Coal preparation decision-making method and system based on multi-source data integration

Abstract

The present application provides a coal preparation decision-making method and system based on multi-source data integration, which relates to the field of coal washing and processing technology, and aims to solve technical problems such as insufficient data utilization of existing coal preparation decision-making methods, lack of scientific and adaptive decision-making capabilities, and low coal preparation efficiency. In the present application, a multi-source coal preparation data set including coal quality detection, equipment operating parameters and market dynamic indicators is first obtained; then the coal quality detection data sequence is standardized and pre-processed to extract coal quality feature groups such as ash distribution, and the equipment operating parameter sequence is extracted. The equipment operating status feature group such as the current change of the sorting equipment is extracted; then, based on the preset deep neural network model, the above feature groups and the market dynamic indicator sequence are subjected to multi-dimensional feature fusion to generate a coal preparation process parameter adjustment strategy combination; finally, the coal preparation equipment is controlled to perform operations such as sorting density optimization based on the strategy combination, and the deep neural network model is updated according to the real-time sorting effect feedback.

Description

Technical Field

[0001] The present application relates to the technical field of coal washing and processing, and in particular to a coal preparation decision-making method and system based on multi-source data integration. Background Art

[0002] In the field of intelligent coal washing and sorting, traditional coal preparation decision-making methods primarily rely on manual experience. In terms of data acquisition, the collection of coal preparation production data has traditionally been rather partial. For example, coal quality testing data often only captures a small number of key indicators, lacking a comprehensive and detailed assessment of coal quality. Furthermore, acquisition often relies on periodic, discrete sampling, failing to capture dynamic information about coal quality over time. Key data, such as the changing trends in ash distribution, real-time fluctuations in sulfur content, and the dynamic correlations between volatile matter and other components, are difficult to effectively capture. This leads to serious flaws in the overall assessment of coal quality. Regarding equipment operating parameters, only basic operating status parameters, such as whether the equipment is powered on and how long it has been running, are focused on, neglecting numerous key parameters that significantly impact coal preparation performance, such as current fluctuations in the sorting equipment, vibration frequency, and medium flow rate. This results in a lack of in-depth understanding of the equipment's actual operating status and the synergistic relationships between these parameters. Furthermore, market dynamics data are almost completely ignored in traditional coal preparation decision-making processes. Important information, such as market price fluctuations and the varying demands for coal product quality and specifications over time, is not factored into decision-making, leading to a significant disconnect between coal preparation production and actual market demand.

[0003] When it comes to data processing and analysis, traditional methods are extremely crude and simplistic. Most methods employ simple statistical analysis of the limited data available, such as calculating averages, maximums, and minimums, failing to delve deeper into the underlying characteristics and patterns. In coal quality analysis, it's impossible to extract from the numerous coal quality indicators the characteristics that truly reflect coal quality and provide guidance for coal preparation decisions. Furthermore, it's impossible to effectively extract key features from equipment operating parameters that characterize the equipment's operating status and its impact on the coal preparation process, let alone leverage these features for scientific decision-making.

[0004] When it comes to coal preparation decisions and process parameter adjustments, the traditional decision-making process relies primarily on the operator's personal experience, lacking scientific theoretical support or systematic methodological support. Process parameter adjustments are typically based on pre-set, fixed rules, often formulated under specific conditions, lacking flexibility and adaptability. When coal quality changes, equipment performance fluctuates, or market demand shifts, this rule-based approach to decision-making and parameter adjustments is unable to adapt to actual conditions and make timely, effective adjustments. This results in low coal preparation efficiency and unstable product quality. Summary of the Invention

[0005] In view of the above, in order to at least partially address the deficiencies in the prior art, in a first aspect, embodiments of the present application provide a coal preparation decision-making method based on multi-source data integration, the method comprising:

[0006] Acquire a multi-source coal preparation data set in a coal preparation production process, wherein the multi-source coal preparation data set includes a coal quality detection data sequence, an equipment operation parameter sequence, and a market dynamic indicator sequence;

[0007] performing standardization preprocessing on the coal quality detection data sequence to generate a standardized coal quality data sequence, and extracting a coal quality feature group based on the standardized coal quality data sequence, wherein the coal quality feature group includes an ash distribution feature, a sulfur fluctuation feature, and a volatile matter correlation feature;

[0008] Extracting time series features from the equipment operation parameter sequence to generate an equipment operation status feature group, wherein the equipment operation status feature group includes a current change feature, a vibration frequency correlation feature, and a medium flow rate matching feature of the sorting equipment;

[0009] Based on a preset deep neural network model, multi-dimensional feature fusion processing is performed on the coal quality feature group, the equipment operation status feature group, and the market dynamic indicator sequence, and a coal preparation process parameter adjustment strategy combination is generated;

[0010] According to the coal preparation process parameter adjustment strategy combination, the coal preparation equipment is controlled to perform sorting density optimization operation, medium flow regulation operation and product particle size classification operation, and the deep neural network model is updated based on real-time sorting effect feedback.

[0011] On the second aspect, an embodiment of the present application also provides a coal preparation decision system based on multi-source data integration, 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-mentioned coal preparation decision method based on multi-source data integration.

[0012] In summary, the coal preparation decision-making method and system based on multi-source data integration provided by the embodiments of the present application achieve intelligent and efficient coal preparation processes, improve coal preparation efficiency, and enhance the stability of product quality. Specifically, by integrating a dynamic correlation processing mechanism for multi-source data, different types of data, such as coal quality test data, equipment operating parameters, and market dynamic indicators, are deeply integrated to form a representation method that can comprehensively and accurately reflect the characteristics of the coal preparation process. This is no longer limited to the simple analysis of a single data type by traditional methods, and comprehensively captures the complex relationships between various factors in the coal preparation process. On this basis, a preset deep neural network model uses advanced machine learning algorithms to deeply mine the process feature set, achieving accurate prediction of key coal preparation parameters, eliminating the uncertainty of previous reliance on experience. The generated parameter set not only has the ability to quickly respond to changes in the coal preparation process, but also closely interacts with the production control unit through a closed-loop feedback mechanism, making the generation and execution of instructions adaptive and able to dynamically adjust according to real-time production conditions. This successfully transforms the coal preparation quality optimization process from a traditional experience-based model to a scientific model prediction model, thereby significantly improving the accuracy and timeliness of parameter adjustment, effectively reducing quality fluctuations in the coal preparation process, and improving production efficiency.

[0013] Furthermore, through the dynamic update mechanism of the deep neural network model parameter weights based on real-time sorting effect feedback (which can correspond to quality indicators), a control system with continuous evolution capabilities has been constructed, enabling the model to automatically correct its decision-making logic according to changing production conditions, ensuring stable and efficient performance in complex and changing industrial production environments.

[0014] Other features and advantages of the present application will be described in detail in the subsequent detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] To more clearly illustrate the technical solutions in the embodiments of this application, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of this application. Those skilled in the art can also derive other drawings based on the above drawings without inventive effort.

[0016] In order to more completely understand the present application and its beneficial effects, the following description will be given in conjunction with the accompanying drawings, wherein the same drawing numbers represent the same parts in the following description.

[0017] Figure 1 This is a flow chart of a coal preparation decision-making method based on multi-source data integration provided in an embodiment of the present application.

[0018] Figure 2This is a hardware environment diagram of a coal preparation decision system based on multi-source data integration provided in an embodiment of the present application. DETAILED DESCRIPTION

[0019] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the embodiments described are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present application.

[0020] The present application will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of a coal preparation decision-making method based on multi-source data integration provided by an embodiment of the present application. The coal preparation decision-making method based on multi-source data integration includes the following steps S110-S150, which are described in detail below.

[0021] Step S110: Acquire a multi-source coal preparation data set in the coal preparation production process, wherein the multi-source coal preparation data set includes a coal quality detection data sequence, an equipment operation parameter sequence, and a market dynamic indicator sequence.

[0022] In this embodiment, using the production and manufacturing scenario of a coal preparation plant as an example, the plant's production system continuously collects various types of data. For example, a coal quality test data sequence can be obtained by comprehensively testing different batches of coal samples, assuming that the different batches of coal samples come from different mining areas of the same coal mine. For example, coal samples A1, A2, and A3 correspond to different testing time points t1, t2, and t3, respectively. The coal quality test data sequence includes the test values ​​of ash, sulfur, volatile matter, and other factors for each coal sample at each time point. The equipment operating parameter sequence includes records of the operating status of various equipment within the coal preparation plant. For example, for sorting equipment B1, B2, and B3, the current values, vibration frequency values, medium flow rate values, and other parameters at each time point t1, t2, and t3 constitute this sequence. Market dynamics indicator sequences are a collection of information related to the coal preparation product market. For example, the clean coal price trend indicator comprises a series of data consisting of price change rates over different time periods. Another example is a market demand forecast indicator, which may be based on past sales data and market research, and a set of forecasts for future demand over different time periods. The inventory turnover rate indicator, calculated from data such as inventory quantity and sales volume, reflects inventory turnover. Through these methods, comprehensive data on various aspects of the coal preparation process, including coal quality, equipment operation, and market dynamics, can be collected, providing a data foundation for subsequent analysis and decision-making.

[0023] Step S120: performing standardization preprocessing on the coal quality detection data sequence to generate a standardized coal quality data sequence, and extracting a coal quality feature group based on the standardized coal quality data sequence, wherein the coal quality feature group includes ash distribution characteristics, sulfur fluctuation characteristics, and volatile matter correlation characteristics.

[0024] Among them, step S120 performs standardization preprocessing on the acquired coal quality detection data sequence, which can be achieved through the following steps S121-S129, as described below.

[0025] Step S121: performing data format conversion processing on the original ash detection values, sulfur detection values ​​and volatile matter detection values ​​in the coal quality detection data sequence to generate a unified dimension ash standardization sequence, a sulfur standardization sequence and a volatile matter standardization sequence.

[0026] In this embodiment, since the original coal quality detection data may come from different detection equipment or follow different detection standards, the dimensions of the ash, sulfur and volatile matter detection values ​​may be inconsistent. In order to ensure consistency in subsequent processing, data format conversion is required. Assuming that there is a set of standard dimension systems, all ash detection values ​​are converted to the standard dimensions to generate an ash standardization sequence, such as converting ash detection values ​​expressed in different units through a specific conversion coefficient so that all ash values ​​in the sequence have the same dimension. Similarly, similar operations are performed on sulfur detection values ​​and volatile matter detection values ​​to generate sulfur standardization sequences and volatile matter standardization sequences. In the above manner, different types of coal quality detection data can be unified in dimension, which is convenient for subsequent data processing and analysis.

[0027] Step S122: performing missing value filling processing on the ash content standardized sequence, and generating a complete ash content continuous sequence based on a linear interpolation algorithm of adjacent coal sample detection values.

[0028] In the standardized ash content sequence, missing values ​​may exist due to detection failures or other reasons. For example, in the detection sequence of coal samples D1, D2, D3, and D4, the ash content value corresponding to coal sample D3 is missing. In this case, by traversing the standardized ash content sequence, the timestamp interval containing the missing value is identified, and the timestamp corresponding to coal sample D3 is found. Next, the valid ash content value preceding the timestamp interval containing the missing value, such as the ash content value of coal sample D2, and the valid ash content value following the timestamp interval containing the missing value, such as the ash content value of coal sample D4, are obtained. The timestamp difference between the two valid ash content values ​​of D2 and D4 is calculated. For example, the timestamp of coal sample D2 is t6, and the timestamp of coal sample D4 is t8, and the difference is t8 - t6. Based on this timestamp difference, linear interpolation weight coefficients are generated. For example, the weight coefficients can be (t8 - the timestamp of the current missing value) / (t8 - t6) and (the timestamp of the current missing value - t6) / (t8 - t6). Based on these two linear interpolation weight coefficients, a weighted sum is calculated for the previous and next valid ash content values: ((t8 - current missing value timestamp) / (t8 - t6)) × ash content value of coal sample D2 + ((current missing value timestamp - t6) / (t8 - t6)) × ash content value of coal sample D4. This fill-in ash value for the timestamp corresponding to the missing value is generated. This fill-in ash value is inserted into the corresponding timestamp position of the ash normalized sequence to generate a continuous ash content sequence without missing values. This continuous ash content sequence is then dimensionally verified to ensure that the units of all ash values ​​in the sequence are consistent with the units of the sulfur values ​​in the sulfur normalized sequence and the volatile content values ​​in the volatile content normalized sequence, according to the preset dimensional conversion rules.

[0029] In another alternative embodiment, step S122 may include the following steps S1221 - S1226 .

[0030] Step S1221: traverse each coal sample detection timestamp in the ash content standardization sequence to identify the timestamp interval where the missing value is located.

[0031] In this example, the detection timestamps corresponding to each coal sample in the ash content normalization sequence are sequentially checked. If the ash content values ​​between timestamps T1 and T2 are missing, this interval is identified as the timestamp interval containing the missing values. This approach allows the precise location of missing data in the ash content normalization sequence, preparing for subsequent filling of missing values.

[0032] Step S1222: Obtain the previous valid ash content detection value and the next valid ash content detection value in the timestamp interval where the missing value is located.

[0033] Based on the timestamp interval of the missing values ​​found above, assuming the previous valid ash content detection value is A1, corresponding to timestamp T0, and the next valid ash content detection value is A2, corresponding to timestamp T3, these two valid ash content detection values ​​are found by locating the timestamps. This method can obtain the relevant data for linear interpolation calculations, providing the necessary conditions for filling missing values.

[0034] Step S1223: Calculate the timestamp difference between the previous valid ash content detection value and the next valid ash content detection value, and generate a linear interpolation weight coefficient according to the timestamp difference.

[0035] For example, we can calculate the difference between timestamps T3 and T0, setting it as ΔT. We can then generate linear interpolation weight coefficients based on the ratio of the timestamp difference to the time length of the interval containing the missing value. For example, if the missing value is between T1 and T2, and the time length of this interval is Δt, then the linear interpolation weight coefficients W1 = (T3 - T1) / ΔT, and W2 = (T2 - T0) / ΔT. This approach allows us to obtain weighted coefficients for calculating missing values, ensuring reasonable fill-in values.

[0036] Step S1224: performing weighted summation on the previous valid ash content detection value and the next valid ash content detection value based on the linear interpolation weight coefficient to generate a filled ash content value of the timestamp where the missing value is located.

[0037] For example, based on the linear interpolation weight coefficients generated previously, the fill-in ash value F = W1 * A1 + W2 * A2 is calculated to obtain the fill-in ash value for the timestamp where the missing value is located. In this way, the missing ash values ​​can be reasonably filled with existing data, making the ash sequence more complete.

[0038] Step S1225: inserting the filled ash value into the corresponding timestamp position of the ash standardization sequence to generate a continuous ash sequence without missing values.

[0039] Insert the calculated fill-in ash value into the timestamp position corresponding to the missing value in the ash normalized sequence. For example, insert the fill-in ash value between timestamps T1 and T2 to form a continuous ash sequence without missing values. This makes the ash normalized sequence complete and meets the data integrity requirements of subsequent analysis.

[0040] Step S1226: performing a dimensional consistency check on the ash continuous sequence to ensure that the units of all ash values ​​in the ash continuous sequence are consistent with the units of the sulfur values ​​in the sulfur standardized sequence and the units of the volatile values ​​in the volatile standardized sequence under a preset dimensional conversion rule.

[0041] In this example, based on a pre-set dimensional conversion rule, the units of all ash values ​​in the continuous ash sequence are checked to ensure they are consistent with the units of the sulfur values ​​in the standardized sulfur sequence and the volatile matter values ​​in the standardized volatile matter sequence. This ensures dimensional consistency across different coal quality data sequences and avoids subsequent analysis errors caused by dimensionality issues.

[0042] Step S123: performing outlier correction processing on the sulfur content normalized sequence, truncating and replacing the sulfur content detection values ​​that exceed the sulfur content threshold interval based on a preset sulfur content threshold interval, and generating a corrected sulfur content sequence.

[0043] In this embodiment, a reasonable sulfur threshold range can be preset, for example, with a lower limit of L1 and an upper limit of U1. For each sulfur content detection value in the sulfur normalized sequence, such as the sulfur content detection values ​​corresponding to coal samples E1, E2, and E3, assume that the sulfur content detection value of coal sample E2 is S2. If S2 is less than L1 or greater than U1, it is considered an outlier. For values ​​less than L1, the value is truncated and replaced with L1; for values ​​greater than U1, the value is truncated and replaced with U1, thus generating a corrected sulfur sequence. This removes outliers from the sulfur normalized sequence, making the data more reliable and facilitating accurate subsequent analysis of sulfur-related features.

[0044] Step S124: performing noise filtering processing on the standardized volatile component sequence, and smoothing the standardized volatile component sequence using a sliding window mean algorithm to generate a denoised volatile component sequence.

[0045] For example, a sliding window size of n can be set, assuming n is 5. For a set of volatile content values ​​starting from coal sample F1 in the standardized volatile content sequence, first take the first five values—the volatile content values ​​of coal samples F1, F2, F3, F4, and F5—and calculate their average. This average is used as the denoised volatile content value for coal sample F3. Then, the sliding window is shifted back one position, and the volatile content values ​​of coal samples F2, F3, F4, F5, and F6 are taken. The average is calculated as the denoised volatile content value for coal sample F4. This process is repeated for the entire standardized volatile content sequence to generate a denoised volatile content sequence. This smoothes the standardized volatile content sequence, removes noise interference, and produces a data sequence that better reflects the true trend of volatile content changes.

[0046] Step S125: merging the ash continuous sequence, the corrected sulfur sequence, and the denoised volatile matter sequence into a standardized coal quality data sequence, and storing the standardized coal quality data sequence in a distributed database.

[0047] In this embodiment, the processed ash continuous sequence, corrected sulfur sequence, and denoised volatile sequence can be merged according to the coal sample's timestamp or other preset correspondence. For example, the ash, sulfur, and volatile values ​​of coal sample G1 at time t9 are derived from the values ​​at corresponding time points in the ash continuous sequence, corrected sulfur sequence, and denoised volatile sequence, respectively. These values ​​are combined to form a single record, thereby constituting a standardized coal quality data sequence. This standardized coal quality data sequence is then stored in a distributed database for subsequent access. The distributed database can be a cloud storage space deployed on a cloud server to store coal mine big data and facilitate intelligent analysis of coal mine production. In this way, the integrated and processed coal quality data is securely and reliably stored, providing a unified data source for further extraction of coal quality characteristics and coal selection decisions.

[0048] Step S126: performing frequency domain transformation processing on the ash content continuous sequence to generate ash content spectrum distribution features, and extracting frequency bands in the ash content spectrum distribution features where the main frequency amplitude exceeds a preset ash content amplitude threshold as ash content distribution features.

[0049] As an example, a method for performing frequency domain transformation on a continuous ash sequence can be a Fourier transform method, which converts the continuous ash sequence in the time domain to the frequency domain, obtaining an ash spectrum distribution feature that contains multidimensional results of different frequency components and their corresponding amplitudes. For example, in some scenarios, such as coal preparation, where the ash content exhibits periodic fluctuations due to equipment vibration, raw material batch differences, etc., the Fourier transform method can effectively capture the dominant frequency components. In one possible embodiment, assuming a preset ash amplitude threshold of T1, within the obtained ash spectrum distribution feature, frequency bands whose dominant frequency amplitude exceeds T1 are searched and marked as ash distribution features. For example, after frequency domain transformation, frequencies f1-f10 and their corresponding amplitudes A1-A10 are obtained. If A3, A5, and A7 exceed T1, then the corresponding frequency range constitutes the ash distribution feature.

[0050] Step S127: performing fluctuation period analysis on the corrected sulfur content sequence, calculating a sequence of absolute difference values ​​between adjacent sulfur content detection values, and performing moving average processing on the sequence of absolute difference values ​​to generate a sulfur content fluctuation intensity sequence. The intervals in the sulfur content fluctuation intensity sequence that exceed a preset sulfur content fluctuation threshold are marked as sulfur content fluctuation features.

[0051] In this embodiment, to correct the sulfur content sequence, the absolute differences between adjacent sulfur content detection values ​​can be calculated sequentially. For example, for the sulfur content detection values ​​S_H1, S_H2, and S_H3 of coal samples H1, H2, and H3, |S_H2 - S_H1|, |S_H3 - S_H2|, and so on are calculated to form a sequence of absolute difference values. This sequence of absolute difference values ​​is then subjected to a moving average. Assuming the moving average window size is m, for example, m is 3, the first three absolute difference values ​​are taken and the average is calculated as the first moving average sulfur content fluctuation intensity value. The window is then shifted back one position, and the next three values ​​are taken and the average is calculated. This process is repeated to generate a sulfur content fluctuation intensity sequence. Furthermore, a sulfur content fluctuation threshold T2 can be preset, and intervals in the sulfur content fluctuation intensity sequence exceeding T2 are marked as sulfur content fluctuation characteristics. This fully reflects the stability of sulfur content in coal quality.

[0052] Step S128: performing correlation analysis on the denoised volatile sequence and the ash continuous sequence, calculating the Pearson correlation coefficient between the volatile detection value and the corresponding ash detection value, and marking the interval where the Pearson correlation coefficient is lower than a preset correlation threshold as a volatile correlation feature.

[0053] In this embodiment, for the denoised volatile matter sequence and the continuous ash content sequence, the Pearson correlation coefficient is calculated for each pair of volatile matter and ash content values, based on the same timestamp or coal sample correspondence. For example, the volatile matter value of coal sample I1 at time t10 is V1 and the ash value is A1. The volatile matter value of coal sample I2 at time t11 is V2 and the ash value is A2. Similarly, the Pearson correlation coefficients for multiple data sets are calculated. A correlation threshold T3 is preset, and intervals with a Pearson correlation coefficient below T3 are marked as volatile matter correlation features. This allows the correlation between volatile matter and ash content to be analyzed and volatile matter correlation features to be extracted, providing a reference for coal preparation processes.

[0054] Step S129: Associating the ash distribution characteristics, the sulfur fluctuation characteristics, and the volatile matter correlation characteristics with corresponding coal sample batch numbers to generate the coal quality characteristic group.

[0055] In this embodiment, each coal sample can be assigned a batch number. For example, coal samples J1, J2, and J3 belong to batch K1. The previously obtained ash distribution characteristics, sulfur fluctuation characteristics, and volatile matter correlation characteristics are associated with the corresponding coal sample batch number. For example, the ash distribution characteristics, sulfur fluctuation characteristics, and volatile matter correlation characteristics corresponding to the coal samples in batch K1 are combined to form a coal quality feature group. This allows different coal quality features to be associated with specific coal sample batches, facilitating subsequent coal selection decision analysis based on coal sample batches.

[0056] Step S130: extracting time series features from the equipment operation parameter sequence to generate an equipment operation status feature group, wherein the equipment operation status feature group includes a current change feature of the sorting equipment, a vibration frequency correlation feature, and a medium flow rate matching feature.

[0057] The process of performing time series feature extraction and other processing on the processing device operation parameter sequence in step S130 may include the following steps S131-S134, which will be described in detail below.

[0058] Step S131: performing trend decomposition processing on the current time series data of the sorting equipment in the equipment operation parameter sequence, separating the current long-term trend component, periodic fluctuation component and residual component, and marking the interval where the slope of the current long-term trend component exceeds the preset slope threshold as the current change feature.

[0059] In this embodiment, for the current time series data of the sorting equipment, assume that there are a series of current values ​​I1-I9 at time points t12-t20. A suitable trend decomposition method is used, such as decomposing it into a long-term current trend component, which reflects the overall trend of current changes over a longer period of time, a periodic fluctuation component, which reflects periodic current changes, and a residual component, which is the portion remaining after removing the long-term trend and periodic fluctuations. For example, a slope threshold T4 can be preset, and the slope of the long-term current trend component can be calculated. For two adjacent points on the long-term current trend component, such as (t13, I13_long) and (t14, I14_long), the slope is (I14_long - I13_long) / (t14 - t13). Intervals where the slope exceeds T4 are marked as current change features. This method can extract features that reflect current change trends from the current time series data, helping to understand changes in the operating status of the sorting equipment.

[0060] Step S132: Jointly analyze the vibration frequency time series data and the medium flow rate time series data, calculate the covariance matrix of the vibration frequency mean and the medium flow rate mean in the same time window, and mark the window in which the main diagonal elements in the covariance matrix exceed the preset covariance threshold as a vibration frequency correlation feature.

[0061] For example, let's set a time window size of p, assuming p is 4. For vibration frequency time series data and medium flow rate time series data, within each time window, starting from time point t21, take the vibration frequency values ​​f1 - f4 and medium flow rate values ​​v1 - v4 from t21 to t24, and calculate the vibration frequency mean (f1 + f2 + f3 + f4) / 4 and the medium flow rate mean (v1 + v2 + v3 + v4) / 4, respectively. Then, calculate the covariance matrix of these two means. Preset a covariance threshold T5. In the covariance matrix, examine the main diagonal elements and mark windows exceeding T5 as vibration frequency-related features.

[0062] Step S133: performing mutation point detection on the medium flow rate time series data, calculating the cumulative deviation of the medium flow rate sequence based on the cumulative sum algorithm, and obtaining the medium flow rate matching feature according to the moment when the cumulative deviation exceeds the preset deviation threshold.

[0063] First, the corresponding sequence of timestamps and flow rate values ​​in the media flow rate time series data is extracted. Assuming that starting at time t25, there are flow rate values ​​v5 - v10, a flow rate curve is generated in the order of the timestamps. The global mean of this flow rate curve is calculated, for example, (v5 + v6 + v7 + v8 + v9 + v10) / 6. Based on this global mean, a flow rate baseline value corresponding to each timestamp is generated. The difference between the actual flow rate value and the flow rate baseline value at each timestamp is accumulated. For example, at time t25, the difference is v5 - the baseline value. Starting from t25, this difference is accumulated to generate a cumulative deviation sequence. The trend of the cumulative deviation sequence is monitored in real time, and a deviation threshold T6 is preset. When the cumulative deviation exceeds T6, the current timestamp is recorded as a candidate mutation point. The actual media velocity at the candidate moment is compared with the mean values ​​within the preceding and following time windows. Assuming the size of the preceding and following time windows is q (e.g., q is 2), the mean within the preceding time window (e.g., the mean of the media velocity at the two time points before the candidate moment) and the mean within the following time window are calculated. If the difference between the actual media velocity value and the mean of the preceding and following time windows exceeds the velocity fluctuation tolerance T7, the candidate moment is confirmed as a valid media velocity mutation point. The timestamp of the valid media velocity mutation point and the corresponding media velocity change amplitude are associated with the sorting device number in the device operating parameter sequence to generate a media velocity matching feature.

[0064] Specifically, the above step S133 may include the following steps S1331-S1336.

[0065] Step S1331: extracting a corresponding sequence of timestamps and medium flow rate values ​​in the medium flow rate time series data, and generating a medium flow rate variation curve in the order of timestamps.

[0066] In this embodiment, each timestamp and its corresponding media flow rate value are extracted from the media flow rate time series data, arranged in ascending order of timestamps to form a sequence, and a media flow rate change curve is plotted based on this sequence. This approach allows for a visual representation of how the media flow rate changes over time, facilitating subsequent mutation point detection.

[0067] Step S1332: Calculate the global mean of the medium flow rate change curve, and generate a medium flow rate reference value corresponding to each timestamp based on the global mean.

[0068] In this embodiment, a global mean is calculated by averaging all the medium flow rate values ​​on the medium flow rate change curve. For each timestamp, this global mean is used as the medium flow rate reference value corresponding to that timestamp. This provides a benchmark for determining whether a sudden change in the medium flow rate has occurred.

[0069] Step S1333: Accumulate and calculate the difference between the actual value of the medium flow rate at each time stamp and the medium flow rate reference value to generate a cumulative deviation sequence.

[0070] In this embodiment, for each timestamp, the difference between the actual medium flow rate and the baseline medium flow rate is calculated. These differences are then accumulated to form a cumulative deviation sequence. This allows quantification of the cumulative change in medium flow rate relative to the baseline, making it easier to detect sudden changes in flow rate.

[0071] Step S1334: monitor the changing trend of the cumulative deviation sequence in real time, and when the cumulative deviation exceeds the preset deviation threshold, record the current timestamp as a candidate mutation point moment.

[0072] In this embodiment, the cumulative deviation sequence is continuously monitored and a preset deviation threshold, such as D1, is set. When the cumulative deviation exceeds D1, the timestamp at that moment is recorded as a candidate mutation point. This allows for preliminary screening of possible medium flow rate mutation points, preparing for subsequent confirmation.

[0073] Step S1335: Compare the mean values ​​of the actual medium flow rate at the candidate moment in the previous and next time windows. If the difference between the actual medium flow rate value and the mean value of the previous time window and the difference between the actual medium flow rate value and the mean value of the next time window both exceed the flow rate fluctuation tolerance, then the candidate mutation point moment is confirmed to be a valid medium flow rate mutation point.

[0074] In this embodiment, for candidate mutation point moments, a preceding and following time window can be set, such as a 5-minute window each. The mean of the medium flow velocity within the preceding and following time windows is calculated, for example, M1 and M2, respectively. The difference between the actual medium flow velocity at the candidate moment and M1 and M2 is then calculated. A velocity fluctuation tolerance can then be set, such as E1. If both differences are greater than E1, the candidate moment is identified as a valid medium flow velocity mutation point. This approach allows for accurate determination of the medium flow velocity mutation point, improving the accuracy of mutation point detection.

[0075] Step S1336: Associating the timestamp of the effective medium flow rate mutation point and the corresponding medium flow rate change amplitude with the sorting device number of the device operation parameter sequence to generate a medium flow rate matching feature.

[0076] In this example, the timestamp of the confirmed effective medium flow rate mutation point and the corresponding medium flow rate change amplitude are associated with the corresponding sorting equipment number in the equipment operating parameter sequence to form a medium flow rate matching feature. This method can clearly identify the equipment to which the mutation point belongs, providing detailed information for analyzing the operating status of coal preparation equipment.

[0077] Step S134: aligning the current change feature, the vibration frequency correlation feature, and the medium flow rate matching feature according to timestamps to generate a device operation status feature group.

[0078] In this example, the previously extracted current variation features, vibration frequency correlation features, and medium flow velocity matching features are aligned using timestamps as a benchmark. For example, for a specific timestamp T, the three features corresponding to that timestamp are combined to form a device operating status feature group. This approach integrates the various features of the device's operation process, comprehensively reflecting its operating status and providing a basis for coal preparation decision-making.

[0079] Step S140: Based on a preset deep neural network model, multi-dimensional feature fusion processing is performed on the coal quality feature group, the equipment operation status feature group and the market dynamic indicator sequence to generate a coal preparation process parameter adjustment strategy combination.

[0080] Herein, step S140 may include the following steps S141-S145.

[0081] Step S141: mapping the ash distribution characteristics, sulfur fluctuation characteristics, and volatile matter correlation characteristics in the coal quality characteristic group into a first characteristic vector.

[0082] In this embodiment, based on specific mapping rules, the ash distribution characteristics, sulfur fluctuation characteristics, and volatile matter correlation characteristics are each converted into vector form and then concatenated in a specific order to form a first feature vector. For example, the ash distribution characteristics contain frequency band information, which is converted into vector A, the sulfur fluctuation characteristics are converted into vector B, and the volatile matter correlation characteristics are converted into vector C. The first feature vector is then formed by concatenating A, B, and C in sequence. This approach can transform the coal quality feature set into a vector form suitable for processing by a deep neural network model, facilitating subsequent feature fusion.

[0083] Step S142: mapping the current change feature, the vibration frequency correlation feature, and the medium flow rate matching feature in the equipment operation status feature group into a second feature vector.

[0084] In this embodiment, similar to the generation method of the first eigenvector, according to specific mapping rules, the current change characteristics, vibration frequency correlation characteristics, and medium flow rate matching characteristics are each converted into vector form and then sequentially concatenated into a second eigenvector. For example, the current change characteristics are converted into vector D, the vibration frequency correlation characteristics are converted into vector E, and the medium flow rate matching characteristics are converted into vector F. The second eigenvector is composed of D, E, and F concatenated in sequence. Through this method, the device operating status feature group is converted into a vector form suitable for processing by a deep neural network model, laying the foundation for multi-dimensional feature fusion.

[0085] Step S143: mapping the clean coal price trend index, the market demand forecast index, and the inventory turnover rate index in the market dynamic index sequence into a third eigenvector.

[0086] In this embodiment, the clean coal price trend indicator, market demand forecast indicator, and inventory turnover rate indicator in the market dynamics indicator sequence are also processed according to specific mapping rules. Assuming that the clean coal price trend indicator contains multi-dimensional information such as the price increase or decrease trend and the magnitude of the change, it is converted into a vector G according to the rules; the market demand forecast indicator may include information such as predicted demand for different time periods and is converted into a vector H; the inventory turnover rate indicator contains relevant information such as the number of inventory turnovers and is converted into a vector I. Vectors G, H, and I are then concatenated in sequence to form a third eigenvector. In this way, the market dynamics indicator sequence is converted into a form similar to other related eigenvectors, facilitating integration within the deep neural network model.

[0087] Step S144: splicing the first eigenvector, the second eigenvector and the third eigenvector through the input layer of the deep neural network model to generate a fused feature vector.

[0088] In this embodiment, the input layer of the deep neural network model receives the first eigenvector, the second eigenvector, and the third eigenvector. These three vectors are connected end to end for splicing according to the splicing method preset by the model. For example, the dimensions of the first eigenvector are first arranged in sequence, then the dimensions of the second eigenvector are connected, and finally the dimensions of the third eigenvector are connected, thereby generating a fused feature vector that includes coal quality characteristics, equipment operating status characteristics, and market dynamic indicator characteristics. This fused feature vector integrates information from multiple aspects and provides a comprehensive data foundation for subsequent further processing within the deep neural network model. In this way, feature vectors from different sources can be effectively combined, so that the model can comprehensively consider various factors for analysis.

[0089] Step S145: performing nonlinear transformation on the fused feature vector through the hidden layer of the deep neural network model, outputting the sorting density adjustment amount, the medium flow adjustment coefficient and the particle size classification threshold, and generating a coal preparation process parameter adjustment strategy combination.

[0090] Among them, step S145 may include the following steps S1451-S1458, which are specifically introduced as follows.

[0091] Step S1451: Input the fused feature vector into the first hidden layer of the deep neural network model, perform nonlinear mapping on the fused feature vector through an activation function, and generate a first intermediate feature vector.

[0092] In this embodiment, after the fused feature vector enters the first hidden layer, the neurons in the hidden layer process each dimension of the fused feature vector according to an activation function. The activation function outputs a new value based on the magnitude of the input value according to a specific nonlinear rule. For example, for each dimension value X in the fused feature vector, the activation function may follow a certain rule, such as outputting a large positive value when X is greater than a certain set value, or a small negative value or zero when X is less than another set value. After the activation function processes all dimensions of the fused feature vector, a first intermediate feature vector is generated. This first intermediate feature vector is no longer a simple combination of the original features; instead, after undergoing a nonlinear transformation, it may extract more representative feature information. By utilizing the nonlinear characteristics of the activation function in this manner, a preliminary transformation of the fused feature vector can be performed, preparing for the subsequent extraction of key features.

[0093] Step S1452: Input the first intermediate feature vector into the second hidden layer of the deep neural network model, perform feature similarity matching with the preset process parameter historical adjustment records, and screen out feature dimensions corresponding to historical adjustment records with similarities higher than a preset matching threshold.

[0094] In this embodiment, the second hidden layer compares the first intermediate feature vector with the preset historical adjustment record of the process parameters. For each dimension of the first intermediate feature vector and the corresponding feature dimension in the historical adjustment record of the process parameters, a specific similarity calculation method is used to measure the degree of similarity between them. For example, a feature distance measurement method may be used to calculate the distance between two feature dimensions. The smaller the distance, the higher the similarity. A preset matching threshold is set. For each dimension, if its similarity with the corresponding dimension in the historical adjustment record is higher than the threshold, the feature dimension corresponding to this historical adjustment record is filtered out. These filtered feature dimensions may contain key feature information when the process parameters were successfully adjusted in the past. In the above manner, the part with a higher similarity to the current feature vector is filtered out from historical experience, providing a reference basis for generating a suitable process parameter adjustment strategy.

[0095] Step S1453: Re-allocate the weights of the first intermediate feature vector based on the filtered feature dimensions to generate a second intermediate feature vector.

[0096] In this embodiment, for the selected feature dimensions, the weights of the various dimensions of the first intermediate feature vector are reallocated based on their relevance or importance to the current situation. For example, for a selected feature dimension, if it has been proven to be very critical in previous cases of successful process parameter adjustment, a larger weight is assigned to the corresponding dimension in the first intermediate feature vector; conversely, if a dimension is relatively unimportant, a smaller weight is assigned. Through this weight redistribution, the first intermediate feature vector is adjusted to generate a second intermediate feature vector. The second intermediate feature vector further highlights the important features related to the current coal preparation process parameter adjustment, which helps the model generate adjustment strategies more accurately. At the same time, the weights of the feature vectors are reasonably adjusted so that the model can better focus on key features and improve the accuracy of generating adjustment strategies.

[0097] Step S1454: Input the second intermediate feature vector into the third hidden layer of the deep neural network model, and perform linear transformation through the sorting density adjustment amount output channel, the medium flow adjustment coefficient output channel and the particle size classification threshold output channel respectively to generate the initial sorting density adjustment amount, the initial medium flow adjustment coefficient and the initial particle size classification threshold.

[0098] In this embodiment, after the second intermediate eigenvector enters the third hidden layer, it enters three different output channels. In the sorting density adjustment output channel, the second intermediate eigenvector is processed according to the linear transformation rule preset in the channel. Assume that the linear transformation rule is to multiply each dimension of the second intermediate eigenvector by a set of specific coefficients and then add them together to obtain a value, which is the initial sorting density adjustment. Similarly, in the medium flow adjustment coefficient output channel and the particle size classification threshold output channel, the second intermediate eigenvector is similarly processed according to their respective preset linear transformation rules to obtain the initial medium flow adjustment coefficient and initial particle size classification threshold, respectively. These initial values ​​are the coal preparation process parameter adjustments initially generated by the model, but may require further processing before they can be applied to actual production. In this way, using the linear transformation of different output channels, initial adjustment values ​​related to the coal preparation process parameters can be generated based on the second intermediate eigenvector.

[0099] Step S1455: converting the unit dimension of the initial sorting density adjustment amount to make it consistent with the density measurement unit of the current sorting equipment, and obtaining the target sorting density adjustment amount.

[0100] In this embodiment, since the unit dimension of the initial sorting density adjustment amount may be inconsistent with the density measurement unit of the current sorting equipment, conversion is required. Assume that the unit of the initial sorting density adjustment amount is a universal density unit, while the current sorting equipment uses another specific density unit. By looking up a pre-set unit dimension conversion table or based on a specific conversion formula (the specific formula is not involved here, it is only a conceptual explanation), the value of the initial sorting density adjustment amount is calculated according to the conversion rule to obtain a target sorting density adjustment amount that is consistent with the density measurement unit of the current sorting equipment. For example, if there is a conversion ratio relationship from the initial unit to the target unit, the initial value is multiplied or divided by the ratio value to obtain the target value. In the above manner, it is ensured that the unit of the sorting density adjustment amount matches the actual equipment, so that it can be directly applied to the adjustment operation of the sorting equipment.

[0101] Step S1456: normalizing the initial medium flow adjustment coefficient, limiting its value range to within the controllable range of the medium flow of the sorting device, and obtaining a target medium flow adjustment coefficient.

[0102] In this embodiment, the value range of the initial medium flow adjustment coefficient may be relatively wide and not necessarily within the controllable range of the medium flow of the sorting device. In order for the coefficient to effectively control the medium flow, normalization processing is required. Assuming that the controllable range of the medium flow of the sorting device is from the lower limit value A to the upper limit value B, a specific normalization method is used, such as subtracting the minimum value of the initial medium flow adjustment coefficient and dividing it by the difference between the maximum and minimum values, to obtain a value between 0 and 1. Then, according to the controllable range of the medium flow of the sorting device, the value between 0 and 1 is linearly transformed so that its value range falls between A and B, and the target medium flow adjustment coefficient is obtained. In this way, the initial medium flow adjustment coefficient is adjusted to an appropriate value range to ensure that it can effectively adjust the medium flow within the controllable range of the sorting device.

[0103] Step S1457: performing integer processing on the initial particle size classification threshold value so as to match it with the calibrated value of the mesh aperture of the grading screen, thereby obtaining the target particle size classification threshold value.

[0104] In this embodiment, the initial particle size classification threshold may be a decimal value, while the calibrated value of the grading screen's mesh aperture is typically an integer. In order for the initial particle size classification threshold to match the actual parameters of the grading screen, integer processing is required. For example, the initial particle size classification threshold can be converted to an integer by rounding off, rounding up, or rounding down. Assuming the initial particle size classification threshold is a decimal X, if rounding is used, when the decimal portion of X is greater than or equal to 0.5, the integer portion is added by 1; when the decimal portion is less than 0.5, the integer portion is directly taken to obtain the target particle size classification threshold that matches the calibrated value of the grading screen's mesh aperture. In this way, the particle size classification threshold is ensured to be consistent with the actual parameters of the grading screen, thereby accurately controlling the product particle size classification operation.

[0105] Step S1458: Associating the target separation density adjustment amount, the target medium flow adjustment coefficient and the target particle size classification threshold according to the equipment number to generate the coal preparation process parameter adjustment strategy combination.

[0106] In this embodiment, a coal preparation plant may have multiple sorting equipment, media pumps, grading screens, and other equipment, each with its own corresponding equipment number. The processed target sorting density adjustment, target media flow adjustment coefficient, and target particle size classification threshold are then associated with the corresponding equipment number. For example, for the sorting equipment with equipment number 1, the target sorting density adjustment, target media flow adjustment coefficient, and target particle size classification threshold applicable to it are combined; a similar combination is performed for the equipment with equipment number 2. Finally, the parameter adjustments corresponding to all equipment are combined to form a coal preparation process parameter adjustment strategy combination. By assigning different process parameter adjustments to specific equipment in this manner, a comprehensive strategic guide is provided for the precise control of coal preparation equipment.

[0107] Step S150: According to the coal preparation process parameter adjustment strategy combination, the coal preparation equipment is controlled to perform sorting density optimization operation, medium flow regulation operation and product particle size classification operation, and the deep neural network model is updated based on real-time sorting effect feedback.

[0108] In this embodiment, step S150 controls the coal preparation equipment to perform separation density optimization, medium flow regulation, and product particle size classification according to the coal preparation process parameter adjustment strategy combination, which specifically includes the following steps S151-S15.

[0109] Step S151: Generate an opening adjustment instruction for the density control valve according to the sorting density adjustment amount in the coal preparation process parameter adjustment strategy combination; control the density control valve to increase or decrease the medium density based on the opening adjustment instruction, and monitor the ash content index of the clean coal after sorting in real time; if the difference between the clean coal ash content index and the target ash content value exceeds the preset ash content tolerance threshold, recalculate the sorting density adjustment amount and iteratively perform the density optimization operation until the difference is lower than the ash content tolerance threshold.

[0110] In this embodiment, a sorting density adjustment value for a particular sorting device is obtained from a combination of coal preparation process parameter adjustment strategies. Based on a predefined relationship, this sorting density adjustment value is converted into an opening adjustment instruction for the density control valve. For example, if the sorting density adjustment value is positive, indicating a need to increase the medium density, the required increase in the density control valve opening is calculated based on the device characteristics and the predefined corresponding relationship. Conversely, if the sorting density adjustment value is negative, the required decrease in the valve opening is calculated. The density control valve is then controlled based on this opening adjustment instruction, thereby increasing or decreasing the medium density. While density adjustment is being performed, the ash content of the sorted clean coal is monitored in real time. Assuming a target ash content value Y and a preset ash tolerance threshold Z, the real-time monitored clean coal ash content is compared with the target ash content value Y, and the difference is calculated. If this difference exceeds the preset ash tolerance threshold Z, the current sorting density is inappropriate and the sorting density adjustment value needs to be recalculated. The recalculation may comprehensively consider the current clean coal ash index, previous adjustment history, and other relevant factors (such as coal quality characteristics). A new sorting density adjustment amount is derived again through a deep neural network model or other preset calculation method. The above process is then repeated until the difference between the clean coal ash index and the target ash value falls below the ash tolerance threshold Z. In this way, the sorting density can be dynamically adjusted according to the actual sorting effect, ensuring that the clean coal ash content meets the requirements and improving the quality of the coal preparation product.

[0111] Step S152: Calculate the target speed value of the medium pump according to the medium flow adjustment coefficient in the coal preparation process parameter adjustment strategy combination; adjust the speed of the medium pump to the target speed value through the frequency converter, and collect real-time medium flow rate sensor data; if the deviation between the real-time medium flow rate and the target flow rate exceeds a preset flow rate deviation threshold, dynamically correct the target speed value based on the proportional integral differential algorithm until the deviation is lower than the flow rate deviation threshold.

[0112] In this embodiment, a medium flow adjustment coefficient is obtained from the coal preparation process parameter adjustment strategy combination. The medium flow adjustment coefficient is used to calculate the target speed of the medium pump based on the characteristics of the medium pump and a pre-established relationship model. For example, a functional relationship exists between the medium pump flow rate and speed. This functional relationship and the medium flow adjustment coefficient are used to calculate the corresponding target speed. The medium pump speed is then adjusted to the target speed using a frequency converter. During the adjustment process, a medium flow velocity sensor installed on the pipeline collects medium flow velocity data in real time. Assuming the target flow rate is a set value V and a preset flow velocity deviation threshold W, the real-time collected medium flow velocity is compared with the target flow rate V to calculate the deviation. If the difference exceeds the preset flow velocity deviation threshold W, the current medium pump speed requires further adjustment. At this point, the target speed is dynamically corrected using the proportional-integral-differential (PID) algorithm. The PID algorithm generates a correction value based on the current deviation value, the rate of change of the deviation, and the accumulated deviation, according to its specific calculation rules. This correction value is added to the original target speed to obtain the new target speed. The medium pump speed is then adjusted again using the frequency converter, and the deviation between the medium flow rate and the target flow rate is continuously monitored. The correction process is repeated until the deviation falls below the preset flow rate deviation threshold W. This allows precise control of the medium pump speed, thereby adjusting the medium flow rate to meet the requirements of the coal preparation process.

[0113] Step S153: Generate a screen vibration amplitude adjustment parameter according to the particle size classification threshold in the coal preparation process parameter adjustment strategy combination; adjust the vibration motor current of the grading screen to match the vibration amplitude adjustment parameter, and collect the particle size distribution data of the oversize material and the undersize material; if the proportion of the target particle size range in the oversize material is lower than the preset proportion threshold, increase the particle size classification threshold and re-execute the particle size classification operation until the proportion reaches the proportion threshold.

[0114] In this embodiment, a particle size classification threshold is first obtained from the coal preparation process parameter adjustment strategy combination. Based on the characteristics of the grading screen and the relationship between particle size classification and screen vibration amplitude, the particle size classification threshold is used to generate a screen vibration amplitude adjustment parameter. For example, the particle size classification threshold is substituted into a correspondence relationship obtained from a pre-established mathematical model or experimental data to calculate an appropriate screen vibration amplitude adjustment parameter. This parameter may include adjustment information for vibration amplitude, vibration frequency, and other aspects to ensure that the grading screen can effectively classify product according to the particle size classification threshold.

[0115] Then, based on the generated mesh vibration amplitude adjustment parameters, the current of the grading screen's vibration motor is adjusted. Because the vibration motor current affects the mesh vibration amplitude, adjusting the current can bring the mesh vibration amplitude into compliance with the vibration amplitude adjustment parameters. After adjusting the vibration motor current, the particle size detection device installed on the grading screen collects particle size distribution data for the oversize and undersize materials. This data reflects the effectiveness of the grading screen's grading according to the set particle size threshold, including multi-dimensional information such as the proportion of material within different particle size ranges.

[0116] Finally, the collected particle size distribution data of the oversize material is analyzed to determine the proportion of the target particle size range in the oversize material. Assuming that the preset proportion threshold is P, the actual proportion is compared with the preset proportion threshold P. If the actual proportion is lower than the preset proportion threshold P, it means that the current particle size classification operation has not achieved the expected effect, and the particle size classification threshold may be set unreasonably. At this time, the particle size classification threshold is increased, and the screen vibration amplitude adjustment parameter is re-generated according to the new particle size classification threshold, the vibration motor current is adjusted, the particle size classification operation is performed again, and the particle size distribution data of the oversize and undersize materials are re-collected. The above comparison and adjustment process is repeated until the proportion of the target particle size range in the oversize material reaches the preset proportion threshold P. In this way, the particle size classification threshold can be dynamically adjusted according to the actual particle size classification effect to ensure that the product particle size classification meets the requirements.

[0117] Furthermore, in step 150, updating the deep neural network model based on real-time sorting effect feedback includes the following steps S154-S157.

[0118] Step S154: collecting a set of clean coal quality indicators after executing the separation density optimization operation, the medium flow adjustment operation and the particle size classification operation, wherein the clean coal quality indicator set includes an actual ash value, an actual sulfur value and a particle size qualified rate.

[0119] In this embodiment, after completing sorting density optimization, medium flow adjustment, and particle size classification, the quality indicators of the clean coal are collected using appropriate testing equipment. Ash testing equipment is used to obtain the actual ash content of the clean coal, while sulfur testing equipment is used to obtain the actual sulfur content. The particle size acceptance rate is calculated by analyzing the particle size distribution data of the oversize and undersize materials. The actual ash content, actual sulfur content, and particle size acceptance rate together constitute a set of clean coal quality indicators. This comprehensive collection of clean coal quality information after the coal cleaning operation provides data support for evaluating coal cleaning performance and updating deep neural network models.

[0120] Step S155: performing feature normalization and fusion processing on the difference between the actual ash value and the target ash value, the difference between the actual sulfur value and the target sulfur value, and the difference between the particle size pass rate and the target pass rate to obtain an effect feedback vector.

[0121] In this embodiment, the difference between the actual ash value and the target ash value, the difference between the actual sulfur value and the target sulfur value, and the difference between the particle size pass rate and the target pass rate are first calculated. Assuming that the target ash value is G1 and the actual ash value is G2, the ash difference is G2 - G1; the target sulfur value is S1, the actual sulfur value is S2, and the sulfur difference is S2 - S1; the target pass rate is R1, the actual particle size pass rate is R2, and the particle size pass rate difference is R2 - R1. The three differences are then arranged in a certain order, and then feature fusion processing is performed to obtain an effect feedback vector. For example, the effect feedback vector can be expressed as [G2 - G1, S2 - S1, R2 - R1]. In this way, the difference between the actual quality index and the target index after the coal preparation operation is integrated into a vector, which is convenient for subsequent input into the deep neural network model for analysis.

[0122] Step S156: input the effect feedback vector and the coal preparation process parameter adjustment strategy into the deep neural network model to calculate the strategy effect loss.

[0123] In this embodiment, a deep neural network model receives a performance feedback vector and a combination of coal preparation process parameter adjustment strategies. The model internally calculates the strategy performance loss according to its predefined calculation method. Assume that the performance feedback vector contains the difference between the actual and target indicators, such as the difference between the actual and target ash content, denoted as ΔA, the difference between the actual and target sulfur content, denoted as ΔS, and the difference in particle size pass rate, denoted as ΔR. The coal preparation process parameter adjustment strategy combination includes the sorting density adjustment amount, denoted as D, the medium flow adjustment coefficient, denoted as F, and the particle size classification threshold, denoted as T. These input elements are comprehensively considered to assess the degree of deviation between the actual performance and the expected performance under these parameter combinations. For example, different weights may be assigned to ΔA, ΔS, and ΔR, denoted as w1, w2, and w3, respectively. Different weights may also be assigned to D, F, and T, denoted as v1, v2, and v3, respectively. The strategy performance loss, L, is then calculated based on a predefined loss calculation function. In this way, the gap between the actual separation effect and the expected effect under the current combination of coal preparation process parameter adjustment strategies can be quantified, providing a basis for subsequent model updates.

[0124] Step S157: Fine-tune the weight parameters of the deep neural network model based on the strategy effect loss to generate an updated deep neural network model.

[0125] In this embodiment, after obtaining the policy effect loss L, the optimization algorithm employed by the deep neural network model uses this loss value to fine-tune the model's weight parameters. Assume that the model contains numerous weight parameters, such as the weights connecting different neurons, denoted as w11, w12, w21, and so on. The optimization algorithm adjusts these weight parameters according to specific rules based on the policy effect loss L. For example, the common gradient descent algorithm calculates the gradient of the loss function with respect to each weight parameter. The gradient represents the rate of change of the loss function in the direction of that weight parameter. A positive gradient indicates that increasing the value of that weight parameter will increase the loss function; a negative gradient indicates that increasing the value of that weight parameter will decrease the loss function. The weight parameters are then adjusted based on the direction and magnitude of the gradient and a pre-set learning rate (denoted as η). For weight parameter w11, its adjusted new value w11' = w11 - η × (gradient of the loss function with respect to w11). All weight parameters in the model are adjusted similarly, generating an updated deep neural network model. In this way, the deep neural network model can optimize its own weight parameters according to the difference between the actual sorting effect and the expected effect, thereby improving the model's ability to generate more appropriate process parameter adjustment strategy combinations in subsequent coal preparation decisions.

[0126] Furthermore, the method in this embodiment also includes a process of training the deep neural network model, which includes the following steps S160-S190.

[0127] Step S160: Acquire a sample coal preparation data set, wherein the sample coal preparation data set includes a sample coal quality feature group, a sample equipment operation feature group, a sample market indicator sequence, and corresponding sample process parameter adjustment records.

[0128] In this embodiment, a sample coal preparation data set for training a deep neural network model can be selected based on historical production data stored in the coal preparation plant's data-based coal preparation decision system or data collected in advance for model training. The sample coal quality feature set is derived from testing and feature extraction of different batches of coal samples. For example, over a period of time, multiple coal samples (P1, P2, P3, etc.) were selected, and each sample was tested for indicators such as ash content, sulfur content, and volatile matter. Further features such as ash distribution characteristics, sulfur content fluctuation characteristics, and volatile matter correlation characteristics were extracted to form the sample coal quality feature set. Sample equipment operation feature sets and sample market indicator sequences can be derived using similar methods and are not detailed here.

[0129] Step S170: Input the sample coal quality feature group, the sample equipment operation feature group and the sample market indicator sequence into the initial deep neural network model to obtain the predicted process parameter adjustment amount.

[0130] In this embodiment, a prepared sample coal quality feature set, a sample equipment operation feature set, and a sample market indicator sequence are input into the initial deep neural network model. The ash distribution characteristics, sulfur fluctuation characteristics, and volatile matter correlation characteristics in the sample coal quality feature set can be input in vector form, with each feature dimension representing a specific aspect of the coal quality. For example, the ash distribution characteristics can be a vector containing information from multiple frequency bands, each corresponding to a different ash distribution. The current variation characteristics, vibration frequency correlation characteristics, and medium flow rate matching characteristics in the sample equipment operation feature set are also input in vector form, reflecting different characteristics of the equipment's operating status. The clean coal price trend indicator, market demand forecast indicator, and inventory turnover rate indicator in the sample market indicator sequence are also converted into appropriate vector form and input into the initial deep neural network model. The initial deep neural network model processes the input sample data to obtain corresponding feature representations and outputs predicted process parameter adjustments, including the predicted sorting density adjustment, the predicted medium flow adjustment coefficient, and the predicted particle size classification threshold. Through the above method, the sample data is analyzed and processed using the initial deep neural network model to obtain the predicted value of the adjustment amount of the coal preparation process parameters based on the current model state, providing a basis for subsequent evaluation of model performance and optimization of the model.

[0131] Step S180: Calculate the mean square error loss between the predicted process parameter adjustment amount and the sample process parameter adjustment record, and update the weight parameters of the initial deep neural network model based on the gradient descent algorithm.

[0132] In this embodiment, the mean square error loss between the predicted process parameter adjustment amount and the sample process parameter adjustment record can be first calculated. Assume that the predicted sorting density adjustment amount in the predicted process parameter adjustment amount is D_pred, and the sorting density adjustment amount in the sample process parameter adjustment record is D_real; the predicted medium flow adjustment coefficient is F_pred, and the medium flow adjustment coefficient in the sample is F_real; the predicted particle size classification threshold is T_pred, and the particle size classification threshold in the sample is T_real. The calculation process of the mean square error loss is as follows: for the sorting density adjustment amount, calculate the square of the difference between its predicted value and the actual value, that is, (D_pred - D_real)²; for the medium flow adjustment coefficient, calculate (F_pred - F_real)²; for the particle size classification threshold, calculate (T_pred - T_real)². The three squared differences are then added and averaged to obtain the mean squared error loss, L = [(D_pred - D_real)² + (F_pred - F_real)² + (T_pred - T_real)²] / 3 (this example uses three parameters; in practice, more parameters may be involved, and the calculation method is similar). After obtaining the mean squared error loss, the weight parameters of the initial deep neural network model are updated using the gradient descent algorithm. The core idea of ​​the gradient descent algorithm is to determine the update direction of the weight parameters by calculating the gradient of the loss function (here, the mean squared error loss) with respect to the model weight parameters. Suppose there is a weight parameter w in the model, and its corresponding gradient is ∂L / ∂w. The gradient represents the rate of change of the loss function in the direction of the weight parameter. If the gradient is positive, increasing the value of the weight parameter will increase the loss function; if the gradient is negative, increasing the value of the weight parameter will decrease the loss function. The weight parameters are updated based on the direction of the gradient and the pre-set learning rate η. The update formula is w' = w - η × (∂L / ∂w). All weight parameters in the model are updated in this way, adjusting the weight parameters in a way that minimizes mean squared error. This quantifies the error between the predicted and actual values, and uses a gradient descent algorithm to optimize the model's weight parameters based on this error. This allows the model to more accurately output coal preparation process parameter adjustments that meet actual requirements in subsequent predictions.

[0133] Step S190: When the mean square error loss is lower than a preset convergence threshold, stop training to obtain and save the deep neural network model.

[0134] In this embodiment, after each update of the weight parameters of the initial deep neural network model using the gradient descent algorithm, a check is performed to determine whether the current mean squared error loss (MSE) is below a preset convergence threshold. The preset convergence threshold can be a value set in advance based on experience and expectations for model performance, representing the acceptable range of error between the model's predicted values ​​and the actual sample values. Assume that the preset convergence threshold is ε. When the calculated MSE loss (L) is less than ε, it indicates that the error between the model's predicted process parameter adjustments and the sample process parameter adjustment records is within an acceptable range after training, indicating that the model has achieved a certain level of accuracy and stability. At this point, model training is terminated to avoid overfitting due to excessive training, where the model performs well on sample data but has poor generalization ability in practical applications. After training is terminated, the current deep neural network model is saved, including information such as the model structure and the trained and optimized weight parameters. These saved models can be used in subsequent coal preparation decision-making processes to accurately predict coal preparation process parameter adjustments based on newly input coal quality data, equipment operation data, and market dynamics, providing effective decision support for coal preparation production. Through the above method, we ensure that the model stops training and saves after reaching a certain accuracy, which not only ensures the effectiveness and practicality of the model, but also avoids unnecessary waste of training resources.

[0135] Figure 2 A schematic diagram illustrates exemplary hardware and software components of a multi-source data integration-based coal preparation decision system 100 that can implement the present invention, as provided in some embodiments of the present application. For example, a processor 120 can be used in the multi-source data integration-based coal preparation decision system 100 to perform the functions of the present invention.

[0136] The coal preparation decision system 100 based on multi-source data integration can be a general-purpose server or a special-purpose server, both of which can be used to implement the method of the present application. Although only one server is shown in this application, for convenience, the functions described in this application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0137] For example, the coal preparation decision system 100 based on multi-source data integration may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and various forms of storage media 140, such as a disk, ROM, or RAM, or any combination thereof. For example, the coal preparation decision system 100 based on multi-source data integration 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 may be implemented based on the aforementioned program instructions. The coal preparation decision system 100 based on multi-source data integration also includes an input / output (I / O) interface 150 between the computer and other input / output devices.

[0138] For ease of explanation, only one processor is described in the coal preparation decision system 100 based on multi-source data integration. However, it should be noted that the coal preparation decision system 100 based on multi-source data integration in this application may also include multiple processors, and therefore the steps performed by one processor described in this application may also be performed jointly or individually by multiple processors. For example, if the processor of the coal preparation decision system 100 based on multi-source data integration performs steps A and B, it should be understood that steps A and B may also be performed jointly by two different processors or individually by one processor. For example, the first processor performs step A and the second processor performs step B, or the first processor and the second processor perform steps A and B together.

[0139] In summary, the coal preparation decision-making method and system based on multi-source data integration provided by the embodiments of the present application achieve intelligent and efficient coal preparation processes, improve coal preparation efficiency, and enhance the stability of product quality. Specifically, by integrating a dynamic correlation processing mechanism for multi-source data, different types of data, such as coal quality test data, equipment operating parameters, and market dynamic indicators, are deeply integrated to form a representation method that can comprehensively and accurately reflect the characteristics of the coal preparation process. This is no longer limited to the simple analysis of a single data type by traditional methods, and comprehensively captures the complex relationships between various factors in the coal preparation process. On this basis, a preset deep neural network model uses advanced machine learning algorithms to deeply mine the process feature set, achieving accurate prediction of key coal preparation parameters, eliminating the uncertainty of previous reliance on experience. The generated parameter set not only has the ability to quickly respond to changes in the coal preparation process, but also closely interacts with the production control unit through a closed-loop feedback mechanism, making the generation and execution of instructions adaptive and able to dynamically adjust according to real-time production conditions. This successfully transforms the coal preparation quality optimization process from a traditional experience-based model to a scientific model prediction model, thereby significantly improving the accuracy and timeliness of parameter adjustment, effectively reducing quality fluctuations in the coal preparation process, and improving production efficiency.

[0140] Furthermore, through the dynamic update mechanism of the deep neural network model parameter weights based on real-time sorting effect feedback (which can correspond to quality indicators), a control system with continuous evolution capabilities has been constructed, enabling the model to automatically correct its decision-making logic according to changing production conditions, ensuring stable and efficient performance in complex and changing industrial environments.

[0141] It should be noted that in order to simplify the description disclosed in this application and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of this application, multiple features are sometimes merged into one embodiment, figure or description thereof.

[0142] In the above embodiments, the descriptions of each embodiment have their own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments. The embodiments, implementation methods and related technical features of this application can be combined and replaced with each other in the absence of conflict. The above are only preferred embodiments of this application and are not intended to limit this application in any form. However, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of this application without departing from the content of the technical solution of this application are still within the scope of the technical solution of this application.

Claims

1. A coal preparation decision-making method based on multi-source data integration, characterized in that: The method comprises: Acquire a multi-source coal preparation data set in a coal preparation production process, wherein the multi-source coal preparation data set includes a coal quality detection data sequence, an equipment operation parameter sequence, and a market dynamic indicator sequence; performing standardization preprocessing on the coal quality detection data sequence to generate a standardized coal quality data sequence, and extracting a coal quality feature group based on the standardized coal quality data sequence, wherein the coal quality feature group includes an ash distribution feature, a sulfur fluctuation feature, and a volatile matter correlation feature; Extracting time series features from the equipment operation parameter sequence to generate an equipment operation status feature group, wherein the equipment operation status feature group includes a current change feature, a vibration frequency correlation feature, and a medium flow rate matching feature of the sorting equipment; Based on a preset deep neural network model, multi-dimensional feature fusion processing is performed on the coal quality feature group, the equipment operation status feature group, and the market dynamic indicator sequence, and a coal preparation process parameter adjustment strategy combination is generated; According to the coal preparation process parameter adjustment strategy combination, the coal preparation equipment is controlled to perform sorting density optimization operations, medium flow regulation operations, and product particle size classification operations, and the deep neural network model is updated based on real-time sorting effect feedback; The process of performing standardization preprocessing on the coal quality detection data sequence to generate a standardized coal quality data sequence includes: Performing data format conversion processing on the original ash detection values, sulfur detection values, and volatile matter detection values ​​in the coal quality detection data sequence to generate a unified dimension ash standardization sequence, a sulfur standardization sequence, and a volatile matter standardization sequence; Performing missing value filling processing on the ash content standardized sequence, and generating a complete ash content continuous sequence based on a linear interpolation algorithm of adjacent coal sample detection values; performing outlier correction processing on the sulfur content normalized sequence, truncating and replacing sulfur content detection values ​​that exceed the sulfur content threshold interval based on a preset sulfur content threshold interval, and generating a corrected sulfur content sequence; performing noise filtering on the standardized volatile component sequence, and smoothing the standardized volatile component sequence using a sliding window mean algorithm to generate a denoised volatile component sequence; Merging the ash continuous sequence, the corrected sulfur sequence, and the denoised volatile sequence into a standardized coal quality data sequence, and storing the standardized coal quality data sequence in a distributed database; The step of extracting a coal quality feature group based on the standardized coal quality data sequence includes: Performing frequency domain transformation on the ash content continuous sequence to generate ash content spectrum distribution features, and extracting frequency bands in the ash content spectrum distribution features where the main frequency amplitude exceeds a preset ash content amplitude threshold as ash content distribution features; performing a fluctuation period analysis on the corrected sulfur content sequence, calculating a sequence of absolute difference values ​​of adjacent sulfur content detection values, performing a moving average process on the sequence of absolute difference values ​​to generate a sulfur content fluctuation intensity sequence, and marking intervals in the sulfur content fluctuation intensity sequence that exceed a preset sulfur content fluctuation threshold as sulfur content fluctuation features; performing a correlation analysis on the denoised volatile sequence and the ash continuous sequence, calculating the Pearson correlation coefficient between the volatile detection value and the corresponding ash detection value, and marking the interval where the Pearson correlation coefficient is lower than a preset correlation threshold as a volatile correlation feature; The ash distribution characteristics, the sulfur fluctuation characteristics, and the volatile matter correlation characteristics are associated with corresponding coal sample batch numbers to generate a coal quality characteristic group.

2. The coal preparation decision-making method based on multi-source data integration according to claim 1 is characterized in that: Extracting time series features from the equipment operation parameter sequence to generate an equipment operation status feature group includes: Performing trend decomposition processing on the current time series data of the sorting equipment in the equipment operation parameter sequence to separate the long-term trend component, the periodic fluctuation component and the residual component of the current, and marking the interval where the slope of the long-term trend component of the current exceeds a preset slope threshold as a current change feature; Perform a joint analysis on the vibration frequency time series data and the medium flow rate time series data, calculate the covariance matrix of the vibration frequency mean and the medium flow rate mean in the same time window, and mark the windows in which the main diagonal elements in the covariance matrix exceed the preset covariance threshold as vibration frequency correlation features; Perform mutation point detection on the medium flow rate time series data, calculate the cumulative deviation of the medium flow rate sequence based on the cumulative sum algorithm, and obtain the medium flow rate matching feature according to the moment when the cumulative deviation exceeds the preset deviation threshold; The current change feature, the vibration frequency correlation feature, and the medium flow rate matching feature are aligned according to timestamps to generate a device operation status feature group.

3. The coal preparation decision-making method based on multi-source data integration according to claim 2 is characterized in that: Perform mutation point detection on the medium flow rate time series data, calculate the cumulative deviation of the medium flow rate sequence based on the cumulative sum algorithm, and obtain the medium flow rate matching feature according to the moment when the cumulative deviation exceeds the preset deviation threshold, including: Extracting a corresponding sequence of timestamps and medium flow rate values ​​in the medium flow rate time series data, and generating a medium flow rate change curve in the order of timestamps; Calculating a global mean of the medium flow rate change curve, and generating a medium flow rate reference value corresponding to each timestamp based on the global mean; Accumulate and calculate the difference between the actual value of the medium flow rate at each time stamp and the medium flow rate reference value to generate a cumulative deviation sequence; Monitor the changing trend of the cumulative deviation sequence in real time, and when the cumulative deviation exceeds the preset deviation threshold, record the current timestamp as a candidate mutation point moment; Comparing the actual value of the medium flow rate at the candidate moment with the mean values ​​in the preceding and following time windows, if the difference between the actual value of the medium flow rate and the mean value of the preceding time window and the difference between the actual value of the medium flow rate and the mean value of the following time window both exceed the flow rate fluctuation tolerance, then confirming the candidate mutation point moment as a valid medium flow rate mutation point; The timestamp of the effective medium flow rate mutation point and the corresponding medium flow rate change amplitude are associated with the sorting device number of the device operation parameter sequence to generate a medium flow rate matching feature.

4. The coal preparation decision-making method based on multi-source data integration according to claim 1, characterized in that: Based on a preset deep neural network model, a multi-dimensional feature fusion process is performed on the coal quality feature group, the equipment operation status feature group, and the market dynamic indicator sequence, and a coal preparation process parameter adjustment strategy combination is generated, including: Mapping the ash distribution feature, sulfur fluctuation feature, and volatile matter correlation feature in the coal quality feature group into a first feature vector; Mapping the current change feature, the vibration frequency correlation feature, and the medium flow rate matching feature in the equipment operation state feature group into a second feature vector; Mapping the clean coal price trend index, the market demand forecast index, and the inventory turnover rate index in the market dynamic index sequence into a third eigenvector; splicing the first eigenvector, the second eigenvector, and the third eigenvector through the input layer of the deep neural network model to generate a fused eigenvector; The fused feature vector is nonlinearly transformed through the hidden layer of the deep neural network model, and the sorting density adjustment amount, medium flow adjustment coefficient and particle size classification threshold are output, and a coal preparation process parameter adjustment strategy combination is generated.

5. The coal preparation decision-making method based on multi-source data integration according to claim 4 is characterized in that: The fused feature vector is nonlinearly transformed through the hidden layer of the deep neural network model to output the separation density adjustment amount, the medium flow adjustment coefficient and the particle size classification threshold, and generate a coal preparation process parameter adjustment strategy combination, including: Inputting the fused feature vector into the first hidden layer of the deep neural network model, performing nonlinear mapping on the fused feature vector through an activation function to generate a first intermediate feature vector; Inputting the first intermediate feature vector into the second hidden layer of the deep neural network model, performing feature similarity matching with preset historical adjustment records of process parameters, and screening out feature dimensions corresponding to historical adjustment records with similarities higher than a preset matching threshold; Re-allocating the weights of the first intermediate feature vectors based on the filtered feature dimensions to generate a second intermediate feature vector; Inputting the second intermediate feature vector into the third hidden layer of the deep neural network model, performing linear transformation through the sorting density adjustment amount output channel, the medium flow adjustment coefficient output channel, and the particle size classification threshold output channel, respectively, to generate an initial sorting density adjustment amount, an initial medium flow adjustment coefficient, and an initial particle size classification threshold; Convert the unit dimension of the initial sorting density adjustment amount to be consistent with the density measurement unit of the current sorting equipment to obtain the target sorting density adjustment amount; Normalizing the initial medium flow adjustment coefficient and limiting its value range to within the controllable range of the medium flow of the sorting device to obtain a target medium flow adjustment coefficient; The initial particle size classification threshold is integerized to match the calibrated value of the sieve aperture of the grading screen to obtain the target particle size classification threshold; The target separation density adjustment amount, the target medium flow adjustment coefficient and the target particle size classification threshold are associated according to the equipment number to generate the coal preparation process parameter adjustment strategy combination.

6. The coal preparation decision-making method based on multi-source data integration according to claim 1 is characterized in that: According to the coal preparation process parameter adjustment strategy combination, the coal preparation equipment is controlled to perform separation density optimization operation, medium flow adjustment operation and product particle size classification operation, including: generating an opening adjustment instruction for a density control valve based on the separation density adjustment amount in the coal preparation process parameter adjustment strategy combination; controlling the density control valve to increase or decrease the medium density based on the opening adjustment instruction, and monitoring the ash content index of the clean coal after separation in real time; and recalculating the separation density adjustment amount and iteratively performing a density optimization operation if the difference between the clean coal ash content index and the target ash content exceeds a preset ash content tolerance threshold value, until the difference falls below the ash content tolerance threshold value; Calculating a target speed value of the medium pump based on the medium flow adjustment coefficient in the coal preparation process parameter adjustment strategy combination; adjusting the speed of the medium pump to the target speed value through a frequency converter, and collecting real-time medium flow rate sensor data; if a deviation between the real-time medium flow rate and the target flow rate exceeds a preset flow rate deviation threshold, dynamically correcting the target speed value based on a proportional-integral-differential algorithm until the deviation is lower than the flow rate deviation threshold; and According to the particle size classification threshold in the coal preparation process parameter adjustment strategy combination, a screen vibration amplitude adjustment parameter is generated; the vibration motor current of the grading screen is adjusted to match the vibration amplitude adjustment parameter, and the particle size distribution data of the oversize material and the undersize material are collected; if the proportion of the target particle size range in the oversize material is lower than the preset proportion threshold, the particle size classification threshold is increased and the particle size classification operation is re-executed until the proportion reaches the proportion threshold.

7. The coal preparation decision-making method based on multi-source data integration according to claim 1 is characterized in that: Updating the deep neural network model based on real-time sorting effect feedback includes: collecting a set of clean coal quality indicators after executing the separation density optimization operation, the medium flow adjustment operation, and the particle size classification operation, wherein the set of clean coal quality indicators includes an actual ash value, an actual sulfur value, and a particle size qualified rate; The difference between the actual ash value and the target ash value, the difference between the actual sulfur value and the target sulfur value, and the difference between the particle size qualified rate and the target qualified rate are subjected to feature normalization and fusion processing to obtain an effect feedback vector; Combining the effect feedback vector and the coal preparation process parameter adjustment strategy into the deep neural network model to calculate the strategy effect loss; The weight parameters of the deep neural network model are fine-tuned based on the strategy effect loss to generate an updated deep neural network model.

8. A coal preparation decision system based on multi-source data integration, characterized by: It includes 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 coal preparation decision-making method based on multi-source data integration as described in any one of claims 1-7.

Images