Machine-made sand screening parameter self-adaptive regulation and control method based on deep learning
By combining a deep learning model with a multi-scale attention LSTM and a deep reinforcement learning module, the problems of single data and lag in the adaptive control of screening parameters of manufactured sand were solved, and efficient adaptive control of the screening process of manufactured sand was achieved.
Patent Information
- Application Number
- CN202511019672.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-23
- Publication Date
- 2025-11-07
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing adaptive control methods for screening parameters of manufactured sand rely on data from a single sensor, cannot integrate multi-dimensional information, have weak model generalization ability, are difficult to adapt to sudden changes in raw material characteristics and fluctuations in operating conditions, have lagging control response, and cannot achieve cross-scenario adaptation.
A deep learning-based dual-module model is adopted, combining a multi-scale attention LSTM module and a deep reinforcement learning module. Through training with historical data and real-time data feedback, the screening parameters are dynamically adjusted to achieve adaptive control.
It improves the control efficiency and effectiveness of the screening process, enabling timely response to changes in raw materials, accurate decision-making and dynamic adjustment, and adaptability to different working conditions.
Smart Images
Figure CN120900943A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of manufactured sand screening, in particular to a manufactured sand screening parameter self-adaptive regulation method based on deep learning. BACKGROUND
[0002] The manufactured sand screening parameter self-adaptive regulation refers to a closed-loop regulation mode in which, in the manufactured sand screening process, the system automatically adjusts key screening parameters (vibration frequency, amplitude, screen inclination angle, feeding speed, screening time, etc.) to dynamically adapt to material changes (particle hardness, moisture content, initial grading difference) and production targets (finished sand grading compliance rate, screening productivity, energy consumption control) by real-time sensing of the screening state (finished sand particle size distribution, screening efficiency, material property fluctuation, etc.), without manual intervention to maintain the screening process in an optimal state.
[0003] In the prior art, the data sensing dimension of the manufactured sand screening parameter self-adaptive regulation is single, and it mainly relies on single sensor data (such as only monitoring the amplitude or identifying the particle size distribution), without fusing multi-dimensional information such as moisture content, powder content, and particle shape, resulting in a lack of comprehensive basis for parameter adjustment; linear or shallow algorithms such as PID and fuzzy control are used, which are difficult to model the strong coupling and nonlinear relationship between multiple factors such as vibration frequency and feeding speed in the screening process, and have insufficient adaptability in the case of sudden changes in material properties; the model generalization ability is weak, and it is mainly based on specific materials (such as granite) or fixed working conditions for training, so when the material is replaced (such as limestone) or the working condition fluctuates (such as a sudden increase in the amount of feed), manual recalibration of the parameters is required, and cross-scene self-adaptation cannot be achieved; at the same time, there is a lag in data processing and parameter adjustment (manual sampling detection takes a long time, and early intelligent systems take a long time to reason), and the response of the regulation and control is difficult to match the speed of material changes. SUMMARY
[0004] In order to solve the problems existing in the prior art, the present application aims to provide a manufactured sand screening parameter self-adaptive regulation method based on deep learning. The manufactured sand screening parameter self-adaptive regulation method based on deep learning realizes accurate decision-making of the manufactured sand screening parameters by combining the deep learning model with historical data training and real-time data feedback, relies on dynamic adjustment mechanism to realize self-adaptive regulation of the screening process, and improves the regulation and control efficiency and effect.
[0005] The manufactured sand screening parameter self-adaptive regulation method based on deep learning described in the present application comprises the following steps:
[0006] S1, obtaining material property historical data of manufactured sand raw materials and running parameter historical data of screening equipment, and performing data processing to obtain a manufactured sand screening fusion data set;
[0007] S2, a double-module deep learning model is constructed, and model training is performed in combination with the mechanism sand screening fusion data set to obtain a screening parameter decision model;
[0008] S3, real-time data of material properties of the mechanism sand raw material and real-time data of operating parameters of the screening equipment are collected in real time, and are input into the screening parameter decision model to obtain optimal screening parameters, and a self-adaptive mechanism sand screening regulation result is obtained through feedback dynamic adjustment.
[0009] Preferably, the step S1 specifically comprises:
[0010] The material property historical data includes mechanism sand particle sphericity, mechanism sand particle angularity, mechanism sand particle size distribution, and mechanism sand moisture content.
[0011] The operating parameter historical data includes amplitude, frequency, energy consumption, and material conveying feeding speed of the screening equipment.
[0012] The material property historical data including mechanism sand particle sphericity, mechanism sand particle angularity, mechanism sand particle size distribution, and mechanism sand moisture content is extracted from the mechanism sand historical database.
[0013] The operating parameter historical data including amplitude, frequency, energy consumption, and material conveying feeding speed of the screening equipment is extracted from the screening equipment operation log.
[0014] Preferably, the step S1 further comprises:
[0015] The data quality comprehensive index Q d is calculated for the extracted material property historical data and operating parameter historical data, respectively. d = 0.4 * R C + 0.3 * L P + 0.3 * C t ;
[0016] The material property historical data and operating parameter historical data with the data quality comprehensive index Q d ≥ 0.85 are obtained, and are matched according to time stamps to obtain a combination of time-series-aligned material property historical data and operating parameter historical data.
[0017] The time-series-aligned material property historical data and operating parameter historical data are standardized and subjected to interval conversion by using a dynamic interval compression algorithm to generate a sliding window segment, missing value checking is performed on each segment, segments with a missing value ratio < 5% are integrated, and a mechanism sand screening fusion data set is obtained.
[0018] Wherein, R C represents the completeness rate; L P represents the accuracy level; and C tindicates the time sequence continuity.
[0019] Preferably, the step S2 specifically comprises:
[0020] The dual-module deep learning model including a multi-scale attention LSTM module and a deep reinforcement learning module is constructed, and time sequence features in the mechanism sand screening fusion dataset are extracted in layers according to the coupling structure of the multi-scale attention LSTM module and the deep reinforcement learning module;
[0021] According to the screening quality, the screening equipment energy consumption and the screening equipment stability, a dynamic target balancing coefficient λ is added in the dual-module deep learning model t =α t *Q q +β t *Q e +γ t *Q s ;
[0022] The dual-module deep learning model dynamically adjusts the priority of each target according to the characteristics of different batches of data in the mechanism sand screening fusion dataset, and obtains a target balancing strategy adapted to different data characteristics;
[0023] Wherein, Q q , Q e and Q s represent the screening quality, the screening equipment energy consumption and the screening equipment stability index respectively; α t , β t and γ t are dynamic weight coefficients updated in real time according to the gradient descent method.
[0024] Preferably, the step S2 further comprises:
[0025] The target balancing strategy is embedded in the dual-module deep learning model, and the mechanism sand screening fusion dataset is divided into a training set and a test set according to the ratio of A1:A2;
[0026] The training set is input into the dual-module deep learning model, the multi-scale attention LSTM module extracts time sequence features in the data, and the data are transmitted to the deep reinforcement learning module;
[0027] The deep reinforcement learning module calculates the loss combined with the target balancing strategy, iteratively optimizes the dual-module deep learning model through parameter updating, and performs Y times of preliminary training;
[0028] After the preliminary training is completed, the multi-scale attention LSTM module mines subtle time sequence features from the training set, the deep reinforcement learning module dynamically matches target balance strategies according to the features of different batches of data in the training set, adjusts the weights of each target error in the loss function, trains the double-module deep learning model in batches according to the training rounds Y and updates the parameters, and dynamically adjusts the learning rate according to the loss change to complete deep training.
[0029] The test set is input into the double-module deep learning model that has completed deep training for testing, and the screening error of the screening parameter decision result and the actual optimal parameter is obtained.
[0030] If the screening error is less than the preset screening error threshold, the double-module deep learning model is used as a screening parameter decision model.
[0031] If the screening error is greater than or equal to the preset screening error threshold, retraining is performed until the screening error is less than the preset screening error threshold.
[0032] Preferably, the step S3 specifically comprises:
[0033] Real-time data of material properties and real-time data of operating parameters of the screening equipment are collected in real time, the real-time data of material properties include real-time particle sphericity of the manufactured sand, real-time particle angularity of the manufactured sand, real-time particle size distribution of the manufactured sand and real-time moisture content of the manufactured sand, and the real-time data of operating parameters include real-time amplitude, real-time frequency, real-time energy consumption and real-time material conveying feeding speed of the screening equipment;
[0034] The collected real-time data of material properties and real-time data of operating parameters are processed according to the edge computing node, time difference is eliminated by a space-time alignment algorithm, and abnormal data is removed by an adaptive filtering mechanism to obtain first processed real-time data;
[0035] The first processed real-time data is input into the screening parameter decision model, a multi-scale attention LSTM module of the screening parameter decision model extracts dynamic time sequence features in the first processed real-time data, a deep reinforcement learning module of the screening parameter decision model combines current data features and calls corresponding target balance strategies, dynamically infers according to the current data features and the target balance strategies, obtains optimal screening parameters, and applies the optimal screening parameters to the screening equipment.
[0036] Preferably, the step S3 further comprises:
[0037] Real-time collection of the actual operation effect data of the screening equipment after applying the optimal screening parameters, deviation analysis of the actual operation effect data and the prediction results of the screening parameter decision model, calculation of the deviation amount of each index actual value and predicted value, and combination of the preset deviation evaluation rule to obtain the overall deviation evaluation result;
[0038] If the overall deviation evaluation result is < the preset deviation range, the original state is maintained;
[0039] If the overall deviation evaluation result is >= the preset deviation range, a dynamic adjustment mechanism is triggered, the deviation data is input into the screening parameter decision model in reverse, the decision logic is updated through the internal parameter adjustment module, and the weight coefficient in the target balancing strategy is corrected according to the deviation characteristics to obtain the self-adaptive mechanism sand screening regulation and control result.
[0040] A computer device of the present application comprises a processor and a memory connected by a signal, the memory stores at least one instruction or at least one program, and the at least one instruction or the at least one program is executed by the processor to perform the above-mentioned mechanism sand screening parameter self-adaptive regulation and control method based on deep learning.
[0041] A computer-readable storage medium of the present application, on which at least one instruction or at least one program is stored, the at least one instruction or the at least one program is loaded by the processor to execute the above-mentioned mechanism sand screening parameter self-adaptive regulation and control method based on deep learning.
[0042] A computer program product of the present application comprises a computer program, and the computer program is executed by the processor to realize the above-mentioned mechanism sand screening parameter self-adaptive regulation and control method based on deep learning.
[0043] The mechanism sand screening parameter self-adaptive regulation and control method based on deep learning of the present application has the following advantages:
[0044] The mechanism sand screening parameter self-adaptive regulation method based on deep learning of the application is closely combined with data processing and model construction, the mechanism sand screening fusion data set obtained by processing the historical data is provided for the training of the double-module deep learning model, the trained screening parameter decision model can better fit the actual screening scene, the construction and training of the double-module deep learning model enable the model to have the ability to extract the dynamic time sequence characteristics of the data and combine the target balancing strategy for reasoning, the obtained screening parameter decision model can output the optimal screening parameter adapted to the real-time data, the real-time data acquisition and model application form a closed loop, the model decision logic and the target balancing strategy are dynamically adjusted through the feedback mechanism, the self-adaptive regulation of the screening parameter is realized, and the change in the actual operation can be responded in time. The mechanism sand screening parameter self-adaptive regulation method based on deep learning realizes the accurate decision of the mechanism sand screening parameter through the deep learning model combined with the historical data training and real-time data feedback, relies on the dynamic adjustment mechanism, realizes the self-adaptive regulation of the screening process, and improves the regulation efficiency and effect. BRIEF DESCRIPTION OF DRAWINGS
[0045] Figure 1 is a flowchart of the mechanism sand screening parameter self-adaptive regulation method based on deep learning described in the application. DETAILED DESCRIPTION
[0046] As shown in Figure 1 , the mechanism sand screening parameter self-adaptive regulation method based on deep learning described in the application comprises the following steps:
[0047] S1, obtaining the material characteristic historical data of the mechanism sand raw material and the operation parameter historical data of the screening equipment, and performing data processing to obtain a mechanism sand screening fusion data set;
[0048] S2, constructing a double-module deep learning model, and combining the mechanism sand screening fusion data set to perform model training to obtain a screening parameter decision model;
[0049] S3, real-time collecting the material characteristic real-time data of the mechanism sand raw material and the operation parameter real-time data of the screening equipment, and inputting the screening parameter decision model to obtain the optimal screening parameter, combining feedback dynamic adjustment to obtain the self-adaptive mechanism sand screening regulation result.
[0050] Further, in the embodiment, step S1 specifically comprises:
[0051] The material characteristic historical data comprises the mechanism sand particle sphericity, the mechanism sand particle angularity, the mechanism sand particle size distribution, and the mechanism sand moisture content;
[0052] The operation parameter historical data comprises the amplitude, frequency, energy consumption, and material conveying feeding speed of the screening equipment;
[0053] extracting material property historical data including the spherical degree of the manufactured sand particles, the angular degree of the manufactured sand particles, the particle size distribution of the manufactured sand and the moisture content of the manufactured sand from the manufactured sand historical database;
[0054] extracting operation parameter historical data including the amplitude, the frequency, the energy consumption and the material conveying feeding speed of the screening equipment from the screening equipment operation log.
[0055] Further, in the embodiment, step S1 further includes:
[0056] calculating data quality comprehensive index Q d = 0.4 * R C + 0.3 * L P + 0.3 * C t for the extracted material property historical data and operation parameter historical data, respectively.
[0057] obtaining material property historical data and operation parameter historical data with data quality comprehensive index Q d ≥ 0.85, and matching the material property historical data and the operation parameter historical data according to timestamps to obtain a combination of time-series-aligned material property historical data and operation parameter historical data;
[0058] performing standardization processing on the time-series-aligned material property historical data and operation parameter historical data, and performing interval conversion by using a dynamic interval compression algorithm to generate a sliding window segment, performing missing value checking on each segment, integrating segments with a missing value ratio < 5% to obtain a manufactured sand screening fusion data set; wherein the interval division principle of the dynamic interval compression algorithm is: according to the time periodicity of manufactured sand production (for example, the operation rhythm of 15 minutes / round of the screening equipment), the size of the sliding window is dynamically adjusted, and the basic window is set to 10 minutes (covering the stable operation period of the equipment); if the data fluctuates sharply (for example, the feeding speed suddenly changes), the size is automatically reduced to 5 minutes (to capture transient changes); if the data is stable, the size is expanded to 20 minutes (to improve efficiency); through time granularity self-adaptation, the local dynamic characteristics of the data (for example, the influence of moisture content mutation on screening effect) are accurately captured.
[0059] The reason for selecting the segment with a missing value ratio < 5% is that when the missing value ratio ≥ 5%, the interpolation repair error ≥ 12%; and when the missing value ratio < 5%, the interpolation error of the adjacent values ≤ 3%, which is the balance point of “data usability” and “calculation efficiency” (too low will filter effective data, and too high will introduce invalid data).
[0060] The standardization processing can select Z-score standardization. Since the material properties (particle size, moisture content) and equipment parameters (amplitude, frequency) of the manufactured sand approximately obey the normal distribution (statistical test shows that the sample skewness and kurtosis meet the normal characteristics), the standardized data meet the distribution assumption of "mean value of 0 and standard deviation of 1" of the deep learning model (such as neural network), which can avoid gradient explosion or disappearance and improve the convergence and generalization ability of the model.
[0061] The time sequence alignment adopts a timestamp synchronization interpolation method. The material properties (for example, 2 seconds / time) and the equipment parameters (for example, 1 second / time) data are supplemented with intermediate time data (for example, the estimated value of 1 second for the material data) through linear interpolation based on the millisecond-level timestamp, so as to ensure that the two types of data are one-to-one corresponding in the time dimension.
[0062] R C represents the completeness rate; L P represents the precision level; C t represents the time sequence continuity.
[0063] The completeness rate refers to the proportion of complete and non-missing data records in the total number of records in the historical data of the material properties (for example, the sphericity of the manufactured sand particles) and the historical data of the operating parameters (for example, the amplitude of the screening equipment). For example, 100 groups of data of the moisture content of the manufactured sand are collected, 95 groups of data are complete, and 5 groups of data are missing. The completeness rate R C = 95 / 100 = 95%, which reflects the degree of keeping the data complete in the collection and storage process. The higher the completeness rate, the stronger the data availability.
[0064] The precision level represents the accuracy of data measurement and recording. Different detection instruments and collection methods have different accuracies for the material properties of the manufactured sand and the operating parameters of the screening equipment.
[0065] The time sequence continuity represents the continuous and coherent state of the data in the time sequence. In the process of screening the manufactured sand, the data is continuously generated over time.
[0066] The step S1 is shown as follows:
[0067] The material properties that need to be obtained include: the sphericity S of the manufactured sand particles (range 0.3-0.8, the closer the value to 1, the closer to the spherical shape), the angularity A of the particles (range 1-5, the larger the value, the sharper the angularity), the particle size distribution D (represented by the average particle size, range 0.5-5 mm), and the moisture content W (range 3%-15%).
[0068] The operating parameters that need to be obtained include: the amplitude V of the screening equipment (range 1-5 mm), the frequency F (range 50-150 Hz), the energy consumption E (range 10-50 kW·h), and the feeding speed S P (range 1-10 t / h).
[0069] Material property history data is extracted from the machine sand history database, extracting material data recorded every hour within 7 days, for example:
[0070] Day 1 10:00: S = 0.5, A = 3, D = 2mm, W = 8%;
[0071] Day 3 14:00: S = 0.6, A = 4, D = 3mm, W = 10%;
[0072] A total of 168 pieces of material data are extracted, and 160 pieces of valid data are actually available due to temporary shutdown of the equipment;
[0073] Equipment operating parameter history data is extracted from the screening equipment operation log, extracting equipment parameters recorded every hour within 7 days, for example:
[0074] Day 1 10:00: V = 3mm, F = 100Hz, E = 20kW·h, S P = 5t / h;
[0075] Day 2 14:00: V = 4mm, F = 120Hz, E = 25kW·h, S P = 6t / h;
[0076] A total of 168 pieces of equipment data are extracted, and 155 pieces of valid data are actually available due to sensor failure;
[0077] Material data completeness rate R C物 = 160 / 168 = 95.2%;
[0078] Equipment data completeness rate R C设 = 155 / 168 = 92.3%;
[0079] Material data is collected by a high-precision laser detector, with precision level L P物 = 0.9 (full score is 1, representing error ≤0.02mm);
[0080] Equipment parameters are collected by ordinary sensors, with precision level L P设 = 0.85 (error allowance ±1Hz);
[0081] Material data is interrupted only once due to shutdown, with time sequence continuity C t物 = 0.85;
[0082] Equipment data is interrupted 3 times due to sensor failure, with time sequence continuity C t设 = 0.7;
[0083] The calculated comprehensive index of material data quality Q d物 = 0.90, and the comprehensive index of equipment data quality Qd设 = 0.85;
[0084] Q d物 ≥ 0.85 in the material data (eliminate 20 with serious interruption and low precision), there are 140 in the equipment data Q d设 ≥ 0.85, 120 meet;
[0085] After screening the material and equipment data, match them according to the time stamp, for example: at 9:00 on the second day, the material data (S = 0.55, A = 3.2, D = 2.5 mm, W = 9%) is aligned with the equipment data (V = 3.5 mm, F = 110 Hz, E = 22 kW·h, S P = 5.5 t / h), forming a set of associated data, and finally obtaining 120 sets of high-quality data pairs aligned in time sequence;
[0086] Standardization processing: taking the particle size distribution D as an example, assuming that the average value μ of D in 7 days of data is 2.5 mm, and the standard deviation σ is 1 mm, and one piece of data D = 3 mm, then after standardization, D' = (3-2.5) / 1 = 0.5; similarly, process the sphericity, angularity, moisture content, amplitude, frequency and other parameters, so that all data have an average value of 0 and a standard deviation of 1;
[0087] Dynamic interval compression and missing value verification: using a sliding window (the window size is set to 4 hours, because the screening equipment is 1 stable period for 4 hours), the standardized data is interval converted, for example, the 4-hour data from 8:00 to 12:00 on the first day generates a window segment;
[0088] Check the missing value in the segment: if the missing value of a certain segment accounts for 3% (<5%), it is retained and integrated; if it accounts for 6% (>5%), it is discarded, and finally 100 effective sliding window segments, i.e. the mechanism sand screening fusion data set, are integrated.
[0089] Further, in the embodiment, step S2 specifically comprises:
[0090] A double-module deep learning model including a multi-scale attention LSTM module and a deep reinforcement learning module is constructed, and the time sequence features in the mechanism sand screening fusion data set are extracted in layers according to the coupling structure of the multi-scale attention LSTM module and the deep reinforcement learning module; the multi-scale attention LSTM module is a multi-scale attention long short-term memory network module;
[0091] According to the screening quality, the energy consumption of the screening equipment and the stability of the screening equipment, a dynamic target balancing coefficient λ t = α t *Q q + β t *Q e + γt *Q s ;
[0092] The double-module deep learning model dynamically adjusts the priorities of each target according to the characteristics of different batches of data in the manufactured sand screening fusion data set, and obtains a target balancing strategy suitable for different data characteristics.
[0093] wherein Q q , Q e and Q s represent the screening quality, the energy consumption of the screening equipment and the stability of the screening equipment, respectively; α t , β t and γ t are dynamic weight coefficients which are updated in real time according to the gradient descent method.
[0094] Further, in the embodiment, step S2 further comprises:
[0095] The target balancing strategy is embedded into the double-module deep learning model, and the training set and the test set are divided from the manufactured sand screening fusion data set according to the ratio of A1:A2.
[0096] A training round Y is set, the training set is input into the double-module deep learning model, the multi-scale attention LSTM module extracts the time sequence characteristics in the data and transmits them to the deep reinforcement learning module.
[0097] The deep reinforcement learning module calculates the loss in combination with the target balancing strategy, iteratively optimizes the double-module deep learning model through parameter updating, and performs Y times of preliminary training.
[0098] After the preliminary training is completed, the multi-scale attention LSTM module mines the subtle time sequence characteristics of the training set, the deep reinforcement learning module dynamically matches the target balancing strategy according to the characteristics of different batches of data in the training set, adjusts the weights of the error of each target in the loss function, trains the double-module deep learning model in batches according to the training round Y and updates the parameters, and dynamically adjusts the learning rate according to the loss change, and completes the deep training.
[0099] The test set is input into the double-module deep learning model which has completed the deep training for testing, and the screening error of the screening parameter decision result and the actual optimal parameters is obtained.
[0100] If the screening error is less than the preset screening error threshold, the double-module deep learning model is taken as the screening parameter decision model.
[0101] If the screening error is greater than or equal to the preset screening error threshold, retraining is performed until the screening error is less than the preset screening error threshold.
[0102] Step S2 is shown as follows:
[0103] A two-module deep learning model is built, which is composed of a multi-scale attention LSTM module and a deep reinforcement learning module. The multi-scale attention LSTM module is responsible for extracting time series features of different time scales from the 100 sliding window segments generated in step S1 (fusion material, equipment data), such as capturing the trend of particle size distribution change within a 4-hour stable period and the fluctuation rule of moisture content in a single day. The deep reinforcement learning module balances the three optimization targets of "screening quality, energy consumption, and equipment stability" by combining a dynamic target balancing strategy, and outputs the optimal screening parameters (amplitude, frequency, etc.) that adapt to the current data characteristics.
[0104] Screening quality Q q The target value is set to 95% pass rate, screening equipment energy consumption Q e The target value is set to 25kW·h, and screening equipment stability Q s The target value is set to device vibration acceleration ≤0.1g.
[0105] Then the dynamic target balancing coefficient dynamic weight coefficient α t =0.4, β t =0.3 and γ t =0.3 (updated in real time according to gradient descent method).
[0106] Then the dynamic target balancing coefficient λ t =0.4*Q q +0.3*Q e +0.3*Q s (During subsequent training, it will be dynamically adjusted according to data characteristics, for example, when the moisture content of the raw material is high, the α t weight will be increased to prioritize the screening quality).
[0107] According to the ratio of A1:A2=7:3, 100 sliding window segments from step S1 are divided into:
[0108] 70 segments of training set (covering typical data of the first 5 days, including moisture content fluctuation and equipment load change scenarios);
[0109] 30 segments of test set (data of the 6th-7th day, used to verify the model generalization ability);
[0110] Set the training round Y=100 (i.e. the model learns the training set for 100 times), and the initial learning rate is set to 0.001 (which will be dynamically adjusted according to the loss change later).
[0111] Preliminary training:
[0112] The multi-scale attention LSTM module extracts the time series features of the training set: for example, it identifies that "when the feed speed S PThe amplitude V needs to be increased from 3 mm to 4 mm to maintain the screening quality when the capacity is increased from 5 t / h to 7 t / h. The rule is transmitted to the deep reinforcement learning module;
[0113] The deep reinforcement learning module combines the dynamic target balancing strategy (prioritize guaranteeing screening quality, \alpha t The weight is temporarily adjusted to 0.5), and the loss of "predicted screening parameters VS actual optimal parameters" is calculated. If a certain group of predicted parameters results in a 5% lower screening quality pass rate than the actual one, the model parameters are optimized in reverse propagation to make subsequent predictions closer to the true optimal solution;
[0114] Deep training:
[0115] The multi-scale attention LSTM module further excavates subtle features: for example, it discovers the implicit rule that "the moisture content of raw materials will be 2% higher than usual at 3 pm on Wednesdays, and the corresponding screening frequency needs to be increased by 5 Hz to stabilize";
[0116] The deep reinforcement learning module dynamically matches the target balancing strategy: for batches of data with high moisture content, the Q q weight is increased;
[0117] For segments with abnormally high equipment energy consumption, the Q e weight is increased, and the weights of various target errors in the loss function are adjusted (for example, the quality error weight is increased from 0.5 to 0.6, and the energy consumption error weight is increased from 0.3 to 0.4). The model is trained in batches (every 10 rounds as one batch) and the parameters are updated;
[0118] If the loss of a certain batch decreases slowly (for example, from 0.2 to 0.15 and then stalls), the learning rate is automatically adjusted from 0.001 to 0.0005 to accelerate convergence;
[0119] Test set input and error calculation:
[0120] The 30 test set segments (data from days 6-7) are input into the trained double-module deep learning model, and the predicted screening parameters (for example, the predicted amplitude V = 3.8 mm and the frequency F = 115 Hz for a certain segment) are output. Compared with the optimal parameters (V = 4 mm and F = 110 Hz) manually adjusted in actual production, the screening error (quantified by the comprehensive index of "quality pass rate difference + energy consumption deviation rate + equipment vibration difference") is calculated;
[0121] The preset screening error threshold is 5% (i.e. the model is usable when the screening error is < 5%), and the comprehensive error after testing is 3% (< 5%). Therefore, the double-module deep learning model meets the requirements and can be used as a screening parameter decision model.
[0122] Further, in this embodiment, step S3 specifically includes:
[0123] The real-time data of the material characteristics of the manufactured sand raw material and the real-time data of the operating parameters of the screening equipment are collected in real time, the real-time data of the material characteristics include real-time particle sphericity of the manufactured sand, real-time particle angularity of the manufactured sand, real-time particle size distribution of the manufactured sand and real-time moisture content of the manufactured sand, and the real-time data of the operating parameters include real-time amplitude, real-time frequency, real-time energy consumption and real-time material conveying feeding speed of the screening equipment;
[0124] The collected real-time data of the material characteristics and the real-time data of the operating parameters are processed by the edge computing node, the time difference in sampling is eliminated by a space-time alignment algorithm, and abnormal data is removed by an adaptive filtering mechanism to obtain first processed real-time data;
[0125] The first processed real-time data is input into the screening parameter decision model, a multi-scale attention LSTM module of the screening parameter decision model extracts dynamic time sequence features in the first processed real-time data, a deep reinforcement learning module of the screening parameter decision model combines current data features and calls corresponding target balancing strategies, dynamically infers according to the current data features and the target balancing strategies, obtains optimal screening parameters, and applies the optimal screening parameters to the screening equipment.
[0126] Further, in the embodiment, step S3 further includes:
[0127] Real-time data of actual operation effect of the screening equipment after applying the optimal screening parameters are collected, deviation analysis is performed on the actual operation effect data and the prediction result of the screening parameter decision model, deviation amounts of each index actual value and predicted value are calculated, and an overall deviation evaluation result is obtained in combination with a preset deviation evaluation rule;
[0128] If the overall deviation evaluation result is less than the preset deviation range, the original state is maintained;
[0129] If the overall deviation evaluation result is greater than or equal to the preset deviation range, a dynamic adjustment mechanism is triggered, the deviation data is input into the screening parameter decision model in reverse, the decision logic is updated by an internal parameter adjustment module, and the weight coefficients in the target balancing strategies are corrected according to the deviation characteristics to obtain adaptive manufactured sand screening regulation and control results;
[0130] Step S3 is shown as follows:
[0131] Collection objects of the real-time data collection:
[0132] Real-time data of the material characteristics (collected once every 2 minutes):
[0133] Instantaneous particle sphericity S 实时 = 0.52, the collection range is 0.3-0.8;
[0134] Real-time particle angularity A 实时 = 3.1, the collection range is 1-5;
[0135] Real-time particle size distribution D 实时 = 2.6 mm (average particle size), collection range 0.5-5 mm;
[0136] Real-time moisture content W 实时 = 8.2%, collection range 3%-15%;
[0137] Real-time equipment operating parameter data (collected every 2 minutes):
[0138] Real-time amplitude V 实时 3.2 mm, collection range 1-5 mm;
[0139] Real-time frequency F 实时 = 105 Hz, collection range 50-150 Hz;
[0140] Real-time energy consumption E 实时 = 21 kW·h, collection range 10-50 kW·h;
[0141] Real-time feed speed S P实时 = 5.2 t / h, collection range 1-10 t / h;
[0142] Due to the time difference in collection of material sensors and equipment sensors (for example, material data is delayed by 2 seconds), linear interpolation is used to align the time stamps (unified to the "8th day 9:05:00" time point);
[0143] Through adaptive filtering (for example, based on the 3σ principle), it is found that "real-time moisture content 8.2%" is within the normal fluctuation range of the same working condition in history (historical average 8%, standard deviation 0.5), and is determined as valid data; if a certain data (for example, amplitude 10 mm, far beyond the historical range), it is determined as an outlier and is excluded;
[0144] After processing, the first real-time data is obtained:
[0145] S 实时 = 0.52, A 实时 = 3.1, D 实时 = 2.6 mm, W 实时 = 8.2%;
[0146] V 实时 = 3.2 mm, F 实时 = 105 Hz, E 实时 = 21 kW·h, S P实时 = 5.2 t / h;
[0147] Extract the dynamic timing features of the first real-time data, such as identifying that "the current water content is 8.2%, which is 0.2% higher than the average value in the same season in historical data, and the feed rate has increased by 0.2t / h", which corresponds to the timing correlation features that need to adjust the amplitude and frequency;
[0148] Combine the target balance strategy (currently prioritize guaranteeing screening quality, α t = 0.45), dynamically infer the optimal parameters: optimal amplitude V 最优 = 3.5mm, optimal frequency F 最优 = 110Hz;
[0149] Parameter application: issue V 最优 = 3.5mm, F 最优 = 110Hz to the screening equipment and adjust the running state;
[0150] Real-time collection of running effect data (collect every 5 minutes after applying the optimal parameters): actual screening quality Q q实际 = 93% (target 95%), actual energy consumption E 实际 = 23kW·h (target 25kW·h), actual stability Q S实际 = 0.08g (target 0.1g, g is the acceleration of gravity);
[0151] Calculate the deviation of each index:
[0152] Quality deviation: 93%-95%=-2% (lower than target);
[0153] Energy consumption deviation: 23kW·h-25kW·h=-2kW·h (better than target);
[0154] Stability deviation: 0.08g-0.1g=-0.02g (better than target);
[0155] The preset deviation range is "-3%≤ comprehensive deviation ≤3%", and the comprehensive deviation is calculated (by weighted summation, for example, quality weight 0.5, energy consumption 0.3, stability 0.2);
[0156] Comprehensive deviation = 0.5*(-2%)+0.3*(-2)+0.2*(-0.02)= -1.604%, since -1.604%>-3% (within the preset range), maintain the current parameters;
[0157] If the deviation exceeds the range (for example, the quality deviation reaches -5%), trigger dynamic adjustment: input the deviation data in reverse to the model, update the decision logic through the "internal parameter adjustment module" (for example, increase the quality target weight) to correct the weight coefficient of the target balance strategy (for example, α t from 0.45 to 0.55), and re-infer the optimal screening parameters.
[0158] The embodiment provides a computer device, comprising a signal-connected processor and a memory, and the memory stores at least one instruction or at least one program, and the at least one instruction or the at least one program is executed by the processor to perform the mechanism sand screening parameter adaptive regulation method based on deep learning described above.
[0159] The method embodiment provided in the embodiment of the application can be executed in a computer terminal, a server or a similar computing device, that is, the computer device can include the computer terminal, the server or the similar computing device, and the internal structure of the computer device can include but is not limited to a processor, a network interface and a memory, wherein the processor, the network interface and the memory in the computer device can be connected through a bus or other manners.
[0160] The embodiment provides a computer readable storage medium, and at least one instruction or at least one program is stored in the computer readable storage medium, and the at least one instruction or the at least one program is loaded by a processor to execute the mechanism sand screening parameter adaptive regulation method based on deep learning described above.
[0161] In the embodiment of the application, the computer readable storage medium can include but is not limited to a U disk, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk and various computer program storage media.
[0162] The embodiment further provides a computer program product, comprising a computer program, and the computer program is executed by a processor to implement the mechanism sand screening parameter adaptive regulation method based on deep learning described above.
[0163] Obviously, those skilled in the art should understand that the steps of the application described above can be realized by using general computing devices, and they can be concentrated on a single computing device or distributed on a network composed of multiple computing devices, and they can be realized by program codes executable by the computing device, so that they can be stored in the storage device and executed by the computing device, and in some cases, the steps shown or described can be executed in different orders, or they can be manufactured into individual integrated circuit modules, or multiple modules or steps can be manufactured into a single integrated circuit module; thus, the application is not limited to any particular combination of hardware and software.
[0164] In the description of the application, it needs to be understood that the orientation words such as "front, back, up, down, left, right", "transverse, vertical, horizontal" and "top, bottom" and the like indicated orientation or position relationship is generally based on the orientation or position relationship shown in the drawings, only for the convenience of describing the application and simplifying the description, without making the opposite statement, these orientation words do not indicate and imply that the indicated device or element must have a particular orientation or be constructed and operated in a particular orientation, therefore it cannot be understood as a limitation on the protection scope of the application.
[0165] For those skilled in the art, according to the above described technical solutions and concepts, other various corresponding changes and transformations can be made, and all these changes and transformations should belong to the protection scope of the claims of the application.
Claims
1. A deep learning-based mechanism sand screening parameter self-adaptive regulation method, characterized in that, The method comprises the following steps: S1, obtaining material characteristic historical data of manufactured sand raw materials and running parameter historical data of a screening device, and performing data processing to obtain a manufactured sand screening fusion data set; S2, constructing a double-module deep learning model, and combining the manufactured sand screening fusion data set to perform model training to obtain a screening parameter decision model; S3, collecting material characteristic real-time data of manufactured sand raw materials and running parameter real-time data of the screening device in real time, and inputting the screening parameter decision model to obtain optimal screening parameters, and combining feedback dynamic adjustment to obtain adaptive manufactured sand screening regulation and control results.
2. The method of claim 1, wherein the method is characterized by, The step S1 specifically comprises: The material characteristic historical data comprises manufactured sand particle sphericity, manufactured sand particle angularity, manufactured sand particle size distribution, and manufactured sand moisture content; The running parameter historical data comprises amplitude, frequency, energy consumption, and material conveying feeding speed of the screening device; The material characteristic historical data comprising manufactured sand particle sphericity, manufactured sand particle angularity, manufactured sand particle size distribution, and manufactured sand moisture content is extracted from a manufactured sand historical database; The running parameter historical data comprising amplitude, frequency, energy consumption, and material conveying feeding speed of the screening device is extracted from a screening device operation log.
3. The method of claim 2, wherein the method is characterized by, The step S1 further comprises: calculating a data quality synthesis index Q for the extracted material property historical data and the operating parameter historical data, respectively d = 0.4 * R C + 0.3 * L P + 0.3 * C t ; obtaining the data quality comprehensive index Q d the material property historical data and the operating parameter historical data with a timestamp, to obtain a time-series aligned combination of the material property historical data and the operating parameter historical data; The material characteristic historical data and the running parameter historical data after time sequence alignment are subjected to standardization processing, and interval conversion is performed by using a dynamic interval compression algorithm to generate a sliding window segment, missing value checking is performed on each segment, segments with a missing value proportion of less than 5% are integrated, and a manufactured sand screening fusion data set is obtained; wherein R C represents the integrity rate; L P represents the accuracy level; C t represents the timing continuity.
4. The method of claim 3, wherein the method is characterized by, The step S2 specifically comprises: A double-module deep learning model comprising a multi-scale attention LSTM module and a deep reinforcement learning module is constructed, time sequence features in the manufactured sand screening fusion data set are extracted in layers according to a coupling structure of the multi-scale attention LSTM module and the deep reinforcement learning module; According to the screening quality, the screening equipment energy consumption and the screening equipment stability, a dynamic target balancing coefficient λ is added in the double-module deep learning model t = alpha t * Q q + beta t * Q e + gamma t * Q s ; The double-module deep learning model dynamically adjusts priorities of each target according to features of different batches of data in the manufactured sand screening fusion data set to obtain a target balancing strategy adapted to different data features; wherein Q q , Q e and Q s represent the screening quality, the energy consumption of the screening device and the stability index of the screening device, respectively; α t , β t and γ t are dynamic weight coefficients which are updated in real time according to the gradient descent method.
5. The method of claim 4, wherein the method further comprises: The step S2 further comprises: The target balancing strategy is embedded in the double-module deep learning model, and the manufactured sand screening fusion data set is divided into a training set and a test set according to a ratio of A1:A2; A training round Y is set, the training set is input into the double-module deep learning model, the multi-scale attention LSTM module extracts time sequence features in the data, and the features are transmitted to the deep reinforcement learning module; The deep reinforcement learning module calculates a loss in combination with the target balancing strategy, iteratively optimizes the double-module deep learning model by parameter updating, and performs Y times of preliminary training; After the preliminary training is completed, the multi-scale attention LSTM module mines subtle time sequence features from the training set, the deep reinforcement learning module dynamically matches target balance strategies according to the features of different batches of data in the training set, adjusts the weights of each target error in the loss function, trains the double-module deep learning model in batches according to the training rounds Y and updates the parameters, and dynamically adjusts the learning rate according to the loss change to complete the deep training; The test set is input into the double-module deep learning model that has completed deep training for testing to obtain a screening error of the screening parameter decision result and the actual optimal parameter; If the screening error is less than a preset screening error threshold, the double-module deep learning model is taken as a screening parameter decision model; If the screening error is greater than or equal to the preset screening error threshold, retraining is performed until the screening error is less than the preset screening error threshold.
6. The method of claim 5, wherein the method further comprises: The step S3 specifically comprises: Real-time data of material characteristics of the manufactured sand raw material and real-time data of operating parameters of the screening equipment are collected in real time, the real-time data of material characteristics include real-time particle sphericity of the manufactured sand, real-time particle angularity of the manufactured sand, real-time particle size distribution of the manufactured sand, and real-time moisture content of the manufactured sand, and the real-time data of operating parameters include real-time amplitude, real-time frequency, real-time energy consumption, and real-time material conveying feeding speed of the screening equipment; The collected real-time data of material characteristics and the real-time data of operating parameters are processed by an edge computing node, abnormal data is removed by using an adaptive filtering mechanism, and first real-time data after processing are obtained by eliminating sampling time difference through a space-time alignment algorithm; The first real-time data are input into the screening parameter decision model, a multi-scale attention LSTM module of the screening parameter decision model extracts dynamic time sequence features in the first real-time data, a deep reinforcement learning module of the screening parameter decision model combines current data features and calls corresponding target balance strategies, performs dynamic inference according to the current data features and the target balance strategies, obtains optimal screening parameters, and applies the optimal screening parameters to the screening equipment.
7. The method of claim 6, wherein the method further comprises: The step S3 further comprises: Actual running effect data after the screening equipment applies the optimal screening parameters are collected in real time, deviation analysis is performed on the actual running effect data and a prediction result of the screening parameter decision model, deviation amounts of each index actual value and predicted value are calculated, and an overall deviation evaluation result is obtained in combination with a preset deviation evaluation rule; If the overall deviation evaluation result is less than a preset deviation range, an original state is maintained; If the overall deviation evaluation result is greater than or equal to the preset deviation range, a dynamic adjustment mechanism is triggered, deviation data are input into the screening parameter decision model in reverse, a decision logic is updated through an internal parameter adjustment module, and a weight coefficient in a target balance strategy is corrected according to deviation features to obtain an adaptive manufactured sand screening regulation result.
8. A computer device comprising a processor and a memory connected by a signal, characterized in that, The memory stores at least one instruction or at least one program, and the at least one instruction or the at least one program is loaded by the processor to execute the method for adaptive regulation and control of manufactured sand screening parameters based on deep learning according to any one of claims 1-7.
9. A computer-readable storage medium having stored thereon, at least one instruction or at least one piece of program, characterized in that, The at least one instruction or the at least one program is loaded by the processor to execute the deep learning-based mechanism sand screening parameter self-adaptive regulation method according to any one of claims 1-7.
10. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the deep learning-based mechanism sand screening parameter self-adaptive regulation method according to any one of claims 1-7.