Coal slurry viscosity on-line monitoring method

Abstract

The invention relates to the field of coal water slurry gasification, and discloses a coal slurry viscosity online monitoring method which comprises the following steps: S1, acquiring voltage, current, coal slurry density, rotating speed and geometric parameters of a stirrer system in real time to form a high-frequency characteristic data stream, and synchronously acquiring the coal slurry viscosity measured by manual analysis as low-frequency label data; s2, carrying out abnormal value filtering and down-sampling processing on the high-frequency characteristic data flow to generate a standardized time sequence characteristic sequence; s3, performing space-time alignment on feature data in the standardized time sequence feature sequence after down-sampling and low-frequency label data according to a preset time window, and constructing a three-dimensional training sample set; and S4, establishing a nonlinear mapping model from the feature sequence to the viscosity by adopting a long short-term memory network, and optimizing model parameters through historical data pre-training and online incremental learning. According to the method, the coal slurry viscosity is obtained through online calculation by collecting real-time data through the data driving model, sampling analysis is not needed, and low-frequency analysis is converted into online analysis close to real time.

Description

Technical Field

[0001] The invention relates to the technical field of water-coal slurry gasification, in particular to an online monitoring method for coal slurry viscosity. Background Art

[0002] In the water-coal slurry gasification process, raw coal is thoroughly ground and then mixed with water and additives to form a water-coal slurry of a certain concentration. This slurry is then pumped into the gasifier as feedstock. While increasing the slurry concentration is desirable to improve the gasifier's effective gas yield and reduce raw material consumption, excessively high slurry concentration increases viscosity and degrades stability, impacting performance such as conveying and atomization. Therefore, in actual operation, it is necessary to balance economic efficiency and operational stability, controlling the slurry viscosity within a reasonable range. Therefore, for the water-coal slurry gasification process, slurry viscosity is a critical control parameter, directly impacting gasification efficiency, slurry transportation and storage, gasifier temperature control, slurry stability, and the economics of the entire gasification unit. Controlling and optimizing slurry viscosity is crucial for ensuring stable operation and improving economic efficiency of the water-coal slurry gasification process.

[0003] During industrial coal-water slurry gasification, the viscosity of the coal slurry is generally controlled between 900 and 1500 cP. Monitoring is typically performed through manual sampling and offline analysis. Sampling and analysis are typically performed once per eight-hour shift. The analysis equipment typically uses a rotational viscometer, which measures the shear force generated at a specified shear rate under fixed conditions, thereby calculating or reading the coal slurry viscosity from the instrument.

[0004] The main problems with traditional sampling and analysis methods are as follows: low analysis frequency, as manual sampling and analysis are required; information lag, as it takes several hours from sampling to analysis, during which time field conditions may have changed, so the analysis results have limited guiding significance for actual production; high cost, as regular sampling and analysis by dedicated personnel is required, resulting in high labor costs for analysts; data reliability risk, as the entire process of sampling, analysis, and data collation is completed manually, there is a risk of error. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the present invention provides an online monitoring method for coal slurry viscosity, which solves the problems of low analysis frequency, information lag and high cost in traditional sampling and analysis methods.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: A method for online monitoring of coal slurry viscosity, comprising the following steps: S1, collect the voltage, current, coal slurry density, rotation speed, and geometric parameters of the stirrer system in real time to form a high-frequency characteristic data stream, and synchronously acquire the coal slurry viscosity measured by artificial analysis as low-frequency label data; S2, perform outlier filtering and downsampling processing on the high-frequency characteristic data stream to generate a standardized time sequence characteristic sequence; S3, perform spatio-temporal alignment on the characteristic data after downsampling according to a preset time window and the label data, and construct a three-dimensional training sample set; S4, use a long short-term memory network to establish a nonlinear mapping model of the characteristic sequence to the viscosity, and optimize the model parameters through historical data pre-training and online incremental learning; S5, based on the trained model, perform dynamic inference on the real-time characteristic sequence, output the coal slurry viscosity prediction value, and perform result calibration; S6, when the result deviation exceeds a preset threshold, retrain and optimize the model, and update and replace the online running model.

[0007] Preferably, the downsampling processing in S2 includes: Perform quartile cutoff mean calculation on the original sampling data within every 15 minutes, specifically, remove the 25% of the largest and 25% of the smallest data within every 15 minutes, and then take the arithmetic mean of the remaining 50% data.

[0008] Preferably, the preset time window in S3 is 8 hours, each window contains 32 consecutive characteristic data samples, and each sample contains the temperature T, raw coal mass flow rate m-coal, total fresh water mass flow rate m-water, density p (using an estimated value or an online measured value), rotation speed n, stirrer voltage U, stirrer current I, and liquid level h parameters.

[0009] Preferably, the standardization processing in S2 includes: Perform dynamic Min-Max normalization on the voltage, current, and density parameters, and set the normalization range to [0, 1]; Discretize the stirrer geometric parameters using one-hot encoding.

[0010] Preferably, the long short-term memory network in S4 includes: A three-layer bidirectional LSTM structure, each layer having 128 hidden units; Introduce a gated attention mechanism at the output end of the LSTM, and the calculation method is to apply a learnable weight coefficient to the hidden state of each time step and then perform weighted summation; The output layer uses a fully connected network, and the activation function is LeakyReLU.

[0011] Preferably, the incremental learning in S4 is specifically implemented as: When 20 new label data are accumulated, freeze the first 80% network layer parameters of the model, and only fine-tune the last two layers of the fully connected network; An exponential decay strategy is used to dynamically adjust the learning rate, and the decay coefficient is set to 0.0001.

[0012] Preferably, the dynamic inference in S5 comprises: Cache real-time feature data to a ring buffer, and trigger the prediction process when 32 valid samples are accumulated; Slide window average processing is performed on the prediction results for three consecutive times, and the window weight is distributed according to [0.2, 0.3, 0.5].

[0013] Preferably, the S5 step further comprises model uncertainty monitoring: Calculate the 95% confidence interval of the prediction result, and automatically trigger the sampling request when the prediction variance of three consecutive times exceeds 0.05; Inject the newly obtained laboratory data into the training data set through the digital twin interface, and start online parameter calibration.

[0014] Preferably, the geometric parameters in S1 comprise: The stirrer blade inclination angle α1, the number of blades α2, the container height-diameter ratio α3, and the stirrer diameter d, and the measurement accuracy of each parameter is 0.5°, integer, 0.01 respectively.

[0015] Preferably, the result calibration of S5 comprises: When the absolute error between the predicted viscosity value and the laboratory measured value exceeds 5%, start the model parameter rollback mechanism; Load the optimal model parameters of the last 30 days to cover the current parameters, and re-execute the incremental learning process.

[0016] The present application provides a coal slurry viscosity online monitoring method. It has the following beneficial effects: 1、The present application uses a data-driven model to collect real-time data and calculate the coal slurry viscosity online, without sampling analysis, and converts low-frequency analysis into near-real-time online analysis. Online analysis is automatic analysis of the system, without the need to increase additional sensors or data acquisition devices. Only the existing DCS system has the measurement points and data, and a one-time investment is required to increase data acquisition, analysis, model, server, etc. Manual analysis is only used for periodic verification, and the workload is negligible.

[0017] 2、The automatic analysis of the present application has low error probability, increased data reliability, and can improve the model application range and calculation accuracy through result verification and retraining during operation. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 Fig. 1 is a schematic diagram of a method flow of a coal slurry viscosity online monitoring method according to the present application; Figure 2 Fig. 2 is another schematic diagram of a method flow of a coal slurry viscosity online monitoring method according to the present application. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be clearly and completely described with reference to the accompanying drawings of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0020] The present application provides a coal slurry viscosity online monitoring method, which is executed by an electronic device. The electronic device can be a single physical electronic device, a cluster of physical electronic devices or a distributed system, or a cloud electronic device providing cloud computing services. The present application does not limit the electronic device, and please refer to the accompanying drawings for details. Figures 1-2 The present application provides a coal slurry viscosity online monitoring method, which includes the following steps: S1, collecting the voltage, current, coal slurry density, rotation speed and geometric parameters of the stirrer system in real time to form a high-frequency feature data stream, and synchronously obtaining the coal slurry viscosity measured by artificial analysis as low-frequency label data; S2, performing outlier filtering and downsampling processing on the high-frequency feature data stream to generate a standardized time sequence feature sequence, the downsampling processing including: performing quartile truncation mean calculation on the original sampling data within every 15 minutes, specifically, removing 25% of the largest values and 25% of the smallest values within every 15 minutes, and then taking the arithmetic mean of the remaining 50% data; the standardization processing including: performing dynamic Min-Max normalization on the voltage, current, coal slurry density, rotation speed and geometric parameters, and setting the normalization range to [0, 1]; and discretizing the stirrer geometric parameters by using one-hot encoding; S3, aligning the feature data in the standardized time sequence feature sequence after downsampling with the low-frequency label data in time and space according to a preset time window, and constructing a three-dimensional training sample set, the preset time window being 8 hours, each window containing 32 continuous feature data samples, and each sample containing the temperature T, the raw coal mass flow m-coal, the total fresh water mass flow m-water for slurry preparation, the density p (using an estimated value or an online measured value), the rotation speed n, the stirrer voltage U, the stirrer current I and the liquid level h parameters; S4, a long short-term memory network is used to establish a nonlinear mapping model of the standardized time sequence feature sequence to viscosity, model parameters are pre-trained and optimized through online incremental learning based on historical data, the long short-term memory network comprises: a three-layer bidirectional LSTM structure, each layer has 128 hidden units; a gating attention mechanism is introduced at the output end of the LSTM, the calculation method is: after applying a learnable weight coefficient to the hidden state of each time step, weighted summation is performed; a full connection network is used in the output layer, the activation function is LeakyReLU, and incremental learning is specifically implemented as follows: when 20 new label data are accumulated, the first 80% of the network layer parameters of the model are frozen, and only the last two layers of the full connection network are fine-tuned and trained; an exponential decay strategy is used to dynamically adjust the learning rate, and the decay coefficient is set to 0.0001; S5, based on the trained model, dynamic inference is performed on the standardized time sequence, a coal slurry viscosity prediction value is output, and result calibration is performed, the dynamic inference comprises: real-time feature data is cached to a ring buffer, when 32 valid samples are accumulated, the prediction process is triggered; the prediction results of three consecutive times are subjected to sliding window average processing, the window weight is distributed as [0.2, 0.3, 0.5], and the model uncertainty monitoring is also included, the 95% confidence interval of the prediction result is calculated, when the prediction variance of three consecutive times exceeds 0.05, a sampling request is automatically triggered; Newly obtained laboratory data is injected into the training data set through a digital twin interface, and online parameter calibration is started; S6, when the result deviation exceeds a preset threshold, the latest job data and the coal slurry viscosity value are appended to the historical training data to perform the steps S2, S3 and S4, the model is retrained and optimized, and the model running online is updated and replaced.

[0021] In order to better understand the above technical solutions, the following example experimental data is provided for explanation and description: 1, high-frequency feature data stream 2024 - 01 - 01 08:00:00.000: voltage 220.5V, current 12.3A, coal slurry density 1200.5kg / m³, rotating speed 80.2rpm, stirrer blade inclination angle α1 30.5°, blade number α2 3, vessel height-diameter ratio α3 1.50, stirrer diameter d 0.8m; 2024 - 01 - 01 08:00:00.001: voltage 220.6V, current 12.4A, coal slurry density 1200.6kg / m³, rotating speed 80.3rpm, stirrer blade inclination angle α1 30.5°, blade number α2 3, vessel height-diameter ratio α3 1.50, stirrer diameter d 0.8m; 2024-01-0108:00:00.002: Voltage 220.4V, Current 12.2A, Coal slurry density 1200.4kg / m3, Rotational speed 80.1rpm, Agitator blade tilt angle a1 30.5°, Blade number a2 3, Vessel height-diameter ratio a3 1.50, Agitator diameter d 0.8m; … (Data continuous acquisition) 2024-01-0108:00:10.000: Voltage 221.0V, Current 12.8A, Coal slurry density 1201.0kg / m3, Rotational speed 80.7rpm, Agitator blade tilt angle a1 30.5°, Blade number a2 3, Vessel height-diameter ratio a3 1.50, Agitator diameter d 0.8m; 2. Low-frequency label data Low-frequency label data is measured by manual analysis, with relatively low acquisition frequency, and the timestamp corresponds to the high-frequency feature data stream but with a longer interval: 2024-01-0108:00:00: Coal slurry viscosity 1500.2mPa·s; 2024-01-0108:00:30: Coal slurry viscosity 1502.5mPa·s; 2024-01-0108:01:00: Coal slurry viscosity 1501.8mPa·s; 3. Standardized time series feature sequence After processing the high-frequency feature data stream, the standardized time series feature sequence is obtained, and a 30-second time window is used for statistics: 2024-01-0108:00:00 (time window start time): Average voltage 220.5V, Average current 12.3A, Average coal slurry density 1200.5kg / m3, Average rotational speed 80.2rpm, Average blade tilt angle a1 30.5°, Average blade number a2 3, Average vessel height-diameter ratio a3 1.50, Average agitator diameter d 0.8m; 2024-01-0108:00:30 (time window start time): Average voltage 220.8V, Average current 12.5A, Average coal slurry density 1200.8kg / m3, Average rotational speed 80.4rpm, Average blade tilt angle a1 30.5°, Average blade number a2 3, Average vessel height-diameter ratio a3 1.50, Average agitator diameter d 0.8m; 2024-01-0108:01:00 (time window start time): average voltage 221.2V, average current 12.7A, average coal slurry density 1201.2kg / m3, average rotating speed 80.6rpm, average paddle tilt angle a1 30.5°, average paddle number a2 3, average vessel height-diameter ratio a3 1.50, average agitator diameter d 0.8m.

[0022] The geometric parameters in S1 include: The agitator paddle tilt angle a1, the paddle number a2, the vessel height-diameter ratio a3, and the agitator diameter d, wherein the measurement accuracy of the agitator paddle tilt angle a1 reaches 0.5°, the measurement accuracy of the paddle number a2 reaches an integer, and the measurement accuracy of the vessel height-diameter ratio a3 reaches 0.01.

[0023] The result calibration of S5 includes: When the absolute error between the predicted viscosity value and the laboratory measured value exceeds 5%, a model parameter rollback mechanism is started; Load the latest 30-day history optimal model parameters to overwrite the current parameters and re-execute the incremental learning process.

[0024] Specifically: Data situation description: Install voltage, current sensors, densitometers (if not, use estimated values), tachometers, and liquid level detection devices in the agitator system. The sampling frequency of characteristic values such as agitator voltage, current, coal slurry density, agitator rotating speed, agitator diameter, liquid level, and other related agitator geometric parameters is high (for example, 1 minute / second); the sampling frequency of the tag value coal slurry viscosity is low (8h) due to sampling and testing conditions. All devices are connected to the central database through the industrial bus to establish a time-stamped raw data table. The following figure is an example: "+" indicates that there is data, "-" indicates that there is no data, and "…" indicates the same as the previous time stamp.

[0025] ts n I U a1 a2 ... μ time stamp stirrer speed stirrer current stirrer voltage stirrer blade tilt angle stirrer blade number other parameters coal slurry viscosity 1 + + + + + - 2 + + + + + - ... ... ... ... ... ... ... 480 + + + + + + 481 + + + + + - 482 + + + + + - ... ... ... ... ... ... ... 960 + + + + + + As shown in the table, in the first 479 samplings, the related tag values such as n, U, I, a1, a2, etc. have data, and the tag value μ has no data; In the 480th sampling, in addition to the related tag values such as n, U, I, a1, a2, etc. having data, the tag value μ also has data; In the 481-959th sampling, the related tag values such as n, U, I, a1, a2, etc. have data, and the tag value μ has no data; In the 960th sampling, in addition to the relevant label values such as n, U, I, a1, a2 and the like, the label value μ also has data.

[0026] Data preprocessing and downsampling: The original data stream enters the preprocessing unit, and first performs outlier filtering: set upper and lower threshold values (±3 times the standard deviation) for voltage and current values, and start the sliding window review mechanism when the threshold value is triggered. Adopt the median replacement method to eliminate transient interference signals. In the downsampling stage, the time window aggregation strategy is adopted - 1200 original sampling points within 15 minutes are calculated by quartile truncation mean, which retains the trend characteristics while suppressing pulse noise; On the one hand, due to the sampling environment or other accidental factors, a small number of sampling points and sampling time data may have large errors, and the data needs to be downsampled and aggregated to reduce random noise in the data. On the other hand, by downsampling, the size of the data set can be significantly reduced, reducing storage requirements and computational complexity. In order to facilitate demonstration, in this embodiment, a fixed time interval of 15 minutes is used, and the average calculation of the feature group data within this period is used to realize downsampling, reducing the original feature value data by 15 times.

[0027] In addition to downsampling, in order to improve the convergence data during subsequent model training and reduce the weight difference between features, parameters with large dimensional differences such as density, voltage and current are implemented by grouping standardization: continuous parameters use dynamic Min-Max scaling (range [0, 1]), and geometric parameters perform one-hot encoding. Design an adaptive normalization coefficient update mechanism to recalculate the extreme value parameters every 24 hours to adapt to the working condition changes and adjust the data to a relatively unified range.

[0028] Data alignment: The feature value data with higher frequency is divided into fixed length windows according to the time interval (8 hours), and each window contains feature data within 8 hours. After preprocessing and downsampling in the previous step, the feature data in each window is 32, and each window corresponds to a label value (coal slurry viscosity μ). Build a double-channel data pipeline: the high-frequency feature channel receives the preprocessed 15-minute granularity data, and the low-frequency label channel receives the 8-hour granularity laboratory data. Develop a timestamp backtracking matching algorithm, taking the viscosity sampling time as the reference point, and intercepting 32 consecutive feature samples (32x15 minutes = 8 hours) to form a three-dimensional tensor (batch x time series x feature dimension), and establish a ring buffer to handle the time series misalignment problem.

[0029] Model construction: For this time series application scenario, a long short-term memory network (LSTM) is used to model the data. A bidirectional LSTM network is built, including a 3-layer stack structure. The input layer receives 32-step time series features (p, U, I, n, d, h...), the hidden layer sets 128 memory units and introduces a gated attention mechanism. The output layer connects a time-distributed fully connected network, and finally outputs the viscosity prediction value m. The residual connection is used to solve the gradient vanishing problem, and the DropPath regularization is embedded to prevent overfitting.

[0030] Learning training scheme: The initialization phase uses a historical data set for pre-training: load the three-year operating data (about 100,000 samples), use the AdamW optimizer for 100 rounds of iterative learning. After online deployment, enable incremental learning mode, and start fine-tuning training every 20 new labeled samples, set the learning rate decay coefficient to 0.0001 to balance the sensitivity of the model.

[0031] Real-time inference engine: Deploy a lightweight inference module to convert the trained model into TensorRT format. Design a double-buffer system: real-time data flows into the ring buffer area for preprocessing, and when 32 valid samples are accumulated, the prediction process is triggered. The output result is smoothed by a sliding window, and a 4-20 mA analog signal output is generated in combination with the process industry standard.

[0032] Self-diagnosis and calibration system: Integrate the model uncertainty monitoring module to calculate the confidence interval of the prediction result. When the prediction variance exceeds the threshold for 3 consecutive times, automatically trigger the laboratory sampling request. Develop a digital twin calibration interface to support manual entry of laboratory data to start online parameter fine-tuning, ensuring long-term running accuracy.

[0033] Although embodiments of the present application have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and alterations can be made thereto without departing from the principles and spirit of the present application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for online monitoring of coal slurry viscosity, characterized in that: The following steps are involved: S1. Real-time acquisition of the voltage, current, coal slurry density, rotational speed, and geometric parameters of the agitator system to form a high-frequency feature data stream, and simultaneous acquisition of the coal slurry viscosity measured by manual analysis as low-frequency tag data. The geometric parameters include: agitator blade inclination angle α1, number of blades α2, container height-to-diameter ratio α3, and agitator diameter d. The parameter measurement accuracy of the agitator blade inclination angle α1 reaches 0.5°, the measurement accuracy of the number of blades α2 reaches an integer, and the measurement accuracy of the container height-to-diameter ratio α3 reaches 0.

01. S2. Perform outlier filtering and downsampling on the high-frequency feature data stream to generate a standardized time series feature sequence; S3, aligning the feature data in the downsampled standardized time series feature sequence with the low-frequency label data in time and space according to a preset time window to construct a three-dimensional training sample set; S4. Using a long short-term memory network to establish a nonlinear mapping model from the standardized time series feature sequence to viscosity, and optimizing model parameters through historical data pre-training and online incremental learning; S5. Performing dynamic reasoning on the standardized time series feature sequence based on the trained model, outputting a predicted value of coal slurry viscosity, and performing result calibration; S6. When the result deviation exceeds the preset threshold, the model is retrained and optimized, and the online model is updated and replaced.

2. The method for online monitoring of coal slurry viscosity according to claim 1, characterized in that: The downsampling process in S2 includes: The quartile truncated mean calculation is performed on the original sampling data within every 15 minutes. Specifically, after removing the data with the largest 25% and the smallest 25% values ​​within every 15 minutes, the arithmetic mean of the remaining 50% of the data is taken.

3. The method for online monitoring of coal slurry viscosity according to claim 1, characterized in that: The preset time window in S3 is 8 hours, and each window contains 32 continuous characteristic data samples, each sample including coal slurry temperature T, raw coal mass flow rate m-coal, total mass flow rate of slurrying fresh water m-water, coal slurry density ρ, coal slurry tank agitator speed n, coal slurry tank agitator voltage U, coal slurry tank agitator current I and coal slurry tank liquid level h parameters.

4. The method for online monitoring of coal slurry viscosity according to claim 1, characterized in that: The standardization process in S2 includes: Perform dynamic Min-Max normalization on voltage, current, coal slurry density, speed, and geometric parameters, and set the normalization range to [0,1]; The geometric parameters of the agitator are discretized using one-hot encoding.

5. The method for online monitoring of coal slurry viscosity according to claim 1, characterized in that: The long short-term memory network in S4 includes: Three-layer bidirectional LSTM structure, each layer has 128 hidden units; A gated attention mechanism is introduced at the output of the LSTM. The calculation method is: applying a learnable weight coefficient to the hidden state of each time step and then performing a weighted sum. The output layer uses a fully connected network and the activation function is LeakyReLU.

6. The method for online monitoring of coal slurry viscosity according to claim 1, characterized in that: The incremental learning in S4 is specifically implemented as follows: When 20 new labeled data are accumulated, the parameters of the first 80% of the network layers of the model are frozen, and only the last two layers of the fully connected network are fine-tuned; The exponential decay strategy is used to dynamically adjust the learning rate, and the decay coefficient is set to 0.0001.

7. The method for online monitoring of coal slurry viscosity according to claim 1, characterized in that: The dynamic reasoning in S5 includes: Cache real-time feature data into a ring buffer, and trigger the prediction process when 32 valid samples are accumulated; The sliding window average processing is performed on the three consecutive prediction results, and the window weights are distributed as [0.2, 0.3, 0.5].

8. The method for online monitoring of coal slurry viscosity according to claim 1, characterized in that: The S5 also includes model uncertainty monitoring: Calculate the 95% confidence interval of the prediction result. When the prediction variance exceeds 0.05 for three consecutive times, a sampling request is automatically triggered. Newly acquired laboratory data are injected into the training dataset via the digital twin interface to initiate online parameter calibration.

9. The method for online monitoring of coal slurry viscosity according to claim 1, characterized in that: The S5 result calibration includes: When the absolute error between the predicted viscosity value and the laboratory measured value exceeds 5%, the model parameter rollback mechanism is activated; Load the historical optimal model parameters from the last 30 days to overwrite the current parameters and re-execute the incremental learning process.

Images