An adaptive control method and system for process parameters in the production of flotation frothers
By using a multi-source data acquisition and edge computing system, combined with deep Q-networks and digital twin virtual production models, the problem of parameter control in flotation frother production relying on manual experience was solved, achieving adaptive control of process parameters and improving production stability and response speed.
Patent Information
- Application Number
- CN202611537404.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-30
- Publication Date
- 2026-11-17
AI Technical Summary
The current process parameters for flotation frother production mainly rely on manual experience and fixed parameter settings, lacking a multi-source data acquisition and edge computing system. This makes it impossible to achieve adaptive control of process parameters, resulting in fluctuations in product quality and instability in production continuity.
A multi-source data acquisition and edge computing system is established, which collects data through temperature sensors, pressure sensors, flow sensors and raw material composition analyzers. Edge computing nodes perform filtering, time alignment and feature extraction. An anomaly filtering module identifies and removes abnormal data points. Combined with deep Q network and digital twin virtual production model, parameter adjustment decisions are made to achieve adaptive control of process parameters.
It improves the real-time sensing capability and data reliability of the flotation frother production process, enhances the accuracy and response speed of process parameter control, and ensures continuous and stable production operation.
Smart Images

Figure CN123113779A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of frothing agent production and processing, specifically relating to an adaptive control method and system for process parameters in flotation frothing agent production. Background Technology
[0002] Flotation frothers are core additives in mineral flotation, and their production involves complex chemical reactions and physical changes. The precision of process parameter control directly affects product performance. Current production processes rely heavily on manual experience and fixed parameter settings for parameter control. This makes it difficult to respond promptly to changes in raw material properties, reaction conditions, or equipment status, easily leading to fluctuations in product quality and affecting production continuity and stability. While some systems collect data, this multi-source data is scattered, asynchronous, and may contain outliers, making it difficult to form reliable production status inputs and limiting the accuracy of subsequent optimization decisions.
[0003] For example, Chinese invention patent CN116637729B discloses an automatic flotation reagent addition control system based on data analysis, including a flotation acquisition module, a server, and a control unit. The server analyzes the slurry flotation parameters to obtain the ore flotation index and generates fine-tuning correction or variable adjustment signals; the control unit executes the fine-tuning or notifies personnel via a smart terminal. This system achieves automatic fine-tuning of reagent addition through parameter monitoring and analysis, optimizing energy consumption and reagent consumption.
[0004] However, the above technical solutions have the following problems.
[0005] First, the current process parameter control of flotation frother production mainly relies on manual experience and fixed parameter settings, lacking a multi-source data acquisition and edge computing system. It is impossible to perform channel-by-channel filtering, time alignment, standardization, feature extraction, and anomaly filtering on the multiple raw sensor signals output by temperature sensors, pressure sensors, flow sensors, raw material composition analyzers, and equipment status monitoring modules at a sampling period of 100ms. As a result, when the raw material properties fluctuate or the equipment status varies, the deep Q network has difficulty obtaining reliable input and cannot achieve adaptive control of process parameters.
[0006] Secondly, existing technologies lack a mechanism to sort the candidate parameter adjustment schemes output by the deep Q-network from high to low Q-values and then select the top N candidate parameter adjustment schemes with the highest Q-values to be fed into the digital twin virtual production model for pre-run verification. It is impossible to calculate a comprehensive score based on product quality indicators, production efficiency indicators, and energy consumption indicators and select the candidate parameter adjustment scheme with the highest comprehensive score as the recommended scheme. The parameter adjustment decision-making risk is high, and it is difficult to balance control accuracy and response speed.
[0007] Therefore, there is an urgent need for an adaptive control method and system for the production process parameters of flotation frothers that can overcome the above-mentioned defects. Summary of the Invention
[0008] The purpose of this invention is to provide an adaptive control method and system for the production process parameters of flotation frothers, which can effectively solve the problems mentioned in the background art.
[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0010] On the one hand, this application provides an adaptive control method for process parameters in the production of flotation frothers, including: establishing a multi-source data acquisition and edge computing system; deploying temperature sensors, pressure sensors, flow sensors, raw material composition analyzers, and equipment status monitoring modules at the flotation frother production site; receiving multiple raw sensor signals at a sampling period of 100ms at the edge computing nodes; performing filtering, time alignment, standardization, and feature extraction sequentially for each channel; and automatically identifying and marking abnormal data points based on physical constraint rules and statistical criteria to remove outliers.
[0011] The cloud platform receives standardized data streams and multidimensional feature vectors to train and update deep Q-networks. The deep Q-networks contain 5 fully connected hidden layers with 256, 128, 64, 32 and 16 neurons in each layer, respectively. The output layer outputs the Q-value for each candidate action and takes the action with the maximum Q-value among all candidate actions as the current optimal decision.
[0012] After sorting the candidate parameter adjustment schemes by Q value from high to low, the top N candidate parameter adjustment schemes with the highest Q values are selected and sent to the digital twin virtual production model for pre-run verification. The digital twin virtual production model adopts a hybrid modeling method that combines mechanistic model and data-driven model. The data-driven model uses a long short-term memory network to calculate a comprehensive score based on product quality indicators, production efficiency indicators and energy consumption indicators. The candidate parameter adjustment scheme with the highest comprehensive score is selected as the recommended scheme.
[0013] On the other hand, an adaptive control system for flotation frother production process parameters disclosed in an embodiment of the present invention includes: a multi-source data acquisition and edge computing system, used to deploy temperature sensors, pressure sensors, flow sensors, raw material composition analyzers and equipment status monitoring modules at the flotation frother production site; the edge computing nodes receive multiple raw sensor signals at a sampling period of 100ms, and perform filtering, time alignment, standardization and feature extraction channel by channel; the anomaly filtering module automatically identifies and marks abnormal data points based on physical constraint rules and statistical criteria, and completes the removal of outliers.
[0014] A deep reinforcement learning-driven process parameter optimization model is used to receive standardized data streams and multidimensional feature vectors on a cloud platform, train and update a deep Q network. The deep Q network contains 5 fully connected hidden layers with 256, 128, 64, 32 and 16 neurons in each layer, respectively. The output layer outputs the Q value for each candidate action, and the action with the maximum Q value among all candidate actions is taken as the current optimal decision.
[0015] The digital twin virtual production model is used to receive the top N candidate parameter adjustment schemes ranked by Q value from high to low for pre-testing and verification. The digital twin virtual production model adopts a hybrid modeling method that combines mechanistic model and data-driven model. The data-driven model uses a long short-term memory network to calculate a comprehensive score based on product quality indicators, production efficiency indicators and energy consumption indicators, and selects the candidate parameter adjustment scheme with the highest comprehensive score as the recommended scheme.
[0016] In summary, this application includes at least one of the following beneficial technical effects:
[0017] 1. This application utilizes a multi-source data acquisition and edge computing system to filter, time-align, standardize, and extract features from multiple raw sensor signals output by temperature sensors, pressure sensors, flow sensors, raw material composition analyzers, and equipment status monitoring modules, with a sampling period of 100ms. An anomaly filtering module then removes outliers based on physical constraints and statistical criteria, enabling the deep Q-network and digital twin virtual production model to obtain quality-controllable inputs and improving the real-time sensing capability and data reliability of the flotation frother production process.
[0018] 2. This application establishes a mapping relationship from production status to process parameter adjustment actions through a deep Q-network. It calculates the reward value by weighted summation of product quality indicators, production efficiency indicators, and energy consumption indicators. Under conditions of changes in raw material properties, fluctuations in operating conditions, and differences in equipment status, it automatically outputs candidate parameter adjustment schemes, breaking through the limitations of the traditional fixed parameter control mode and improving the accuracy and response speed of process parameter control.
[0019] 3. This application uses a digital twin virtual production model to pre-test and verify candidate parameter adjustment schemes. It adopts a hybrid modeling method that combines mechanistic models and long short-term memory network data-driven models, and selects the top N candidate parameter adjustment schemes by ranking them from high to low according to their Q values. The recommended scheme is selected based on the comprehensive score. Quality risks are identified before parameter adjustment is implemented to prevent unqualified batches from entering actual production. At the same time, remote monitoring and full-process traceability are achieved through a distributed architecture that coordinates cloud platforms and edge computing nodes, ensuring continuous and stable production operation. Attached Figure Description
[0020] Figure 1This is a schematic diagram of the overall technical solution for the adaptive control method of process parameters in the production of flotation frothers.
[0021] Figure 2 This is a schematic diagram illustrating the core principle of an adaptive control algorithm based on multi-source data fusion and dynamic threshold adjustment.
[0022] Figure 3 This is a flowchart of the process parameter acquisition and feature extraction logic.
[0023] Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow between the field terminal and the cloud control platform.
[0024] Figure 5 This is a schematic diagram comparing the adaptive control principle with existing technologies in terms of parameter control accuracy and response speed. Detailed Implementation
[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the following description is provided in conjunction with the appendix. Figure 1 To be continued Figure 5 The present invention will be further described in detail below with reference to specific embodiments.
[0026] This invention provides an adaptive control method for process parameters in the production of flotation frothers. Its technical architecture comprises the following core modules: a multi-source data acquisition and edge computing system, a deep reinforcement learning-driven process parameter optimization model, an intelligent decision engine module, a digital twin virtual production model, and a distributed architecture where a cloud monitoring platform and edge computing nodes collaborate. These modules work together to achieve autonomous optimization and remote monitoring of the flotation frother production process.
[0027] First, step S1 of the embodiment of this application will be described, which establishes a multi-source data acquisition and edge computing system, as detailed below.
[0028] Step S101: Temperature sensors, pressure sensors, flow sensors, raw material composition analyzers, and equipment status monitoring modules are arranged at the flotation frother production site to form a multi-source data acquisition network.
[0029] The temperature sensor has a measurement range of 0 to 300 degrees Celsius and a measurement accuracy of ±0.5 degrees Celsius.
[0030] The pressure sensor range is set to 0 to 2 MPa.
[0031] The flow sensor uses a high-precision electromagnetic flow meter with a measurement accuracy of ±0.1%.
[0032] Temperature sensors, pressure sensors, flow sensors, raw material composition analyzers, and equipment status monitoring modules are respectively positioned to monitor raw materials, reaction vessels, and key rotating equipment on the production site. The multiple raw sensor signals they output serve as input data for subsequent edge computing nodes.
[0033] In step S102, the raw material composition analyzer uses near-infrared spectroscopy to detect key chemical components in the raw material online. The equipment status monitoring module integrates vibration sensors, acceleration sensors, and acoustic emission sensors to collect operating status information of key rotating equipment.
[0034] In the production of flotation frothers, deviations in the mass fraction of key components in the raw materials can alter the reaction path and the foaming properties of the products. Early wear of rotating parts in the equipment can lead to fluctuations in flow rate and pressure. The raw material composition analyzer outputs the content of key chemical components in the raw materials, while the equipment status monitoring module outputs the operating status information of the rotating equipment. Together, these two components constitute the input data for subsequent edge computing nodes.
[0035] The mass fraction of key components output by the raw material composition analyzer, the vibration characteristic spectrum components output by the equipment condition monitoring module, the feature extraction of the edge computing node in step S103, and the state space vector of the deep Q network in step S201 constitute the data dependency relationship from raw material composition detection to deep Q network state space update.
[0036] When the raw material composition analyzer detects that the mass fraction of key component A deviates from the target value, the mass fraction offset of key component A is standardized by the edge computing node in step S104 and then input as an independent dimension of the state space vector into the deep Q-network. The vibration characteristic spectrum components output by the equipment condition monitoring module are extracted by the edge computing node and then input as the equipment health status dimension into the state space of the deep Q-network. The deep Q-network re-evaluates the Q-values of each candidate action in the action space in the next decision cycle.
[0037] In step S103, the edge computing node has a data processing capability of 10,000 points per second, and processes multi-channel sensor data in parallel. The data sampling period is set to 100ms. The edge computing node receives the multiple raw sensor signals output from steps S101 and S102, and processes them independently and in parallel by channel within the 10,000-point-per-second processing capability. Each channel performs scalar operations separately, without cross-channel matrix coupling operations, and finally outputs multiple channels of data aligned with a 100ms sampling period.
[0038] In the production of flotation frothers, changes in the mass fraction of key components of raw materials, fluctuations in reactor temperature, pressure and flow rate, and equipment health status all enter the subsequent deep reinforcement learning state space update stage under the same time reference.
[0039] In step S104, the edge computing node preprocesses and extracts features from the collected data. Data filtering is performed by taking the arithmetic mean of five consecutive sampling points from the same channel to obtain the current filtered value for that channel. Time synchronization is based on the local clock of the edge computing node, aligning the timestamps of each data channel to a 100ms sampling period. If data is missing from a channel within the same period, it is marked as missing and waits for the next period, not participating in the feature calculation of the current sampling period.
[0040] Standardization uses the mean and standard deviation of historical normal operating data from the same channel as a benchmark. The current value is subtracted from the mean and then divided by the standard deviation to obtain a standardized value with zero mean and unit variance. The calculation method for the standardized value is as follows: Let the mean of historical normal operating data from the same channel be... The standard deviation is The current filter value is Then the standardized value for:
[0041]
[0042] in, These are standardized values, dimensionless, with a mean of 0 and a standard deviation of 1. The current filter value for the same channel has the same dimensions as the process parameters. The average of historical normal operating data for the same channel, with units equal to... same; The standard deviation of historical normal operating data for the same channel, with units of... same. The standardized value is calculated by subtracting the mean from the current filtered value and then dividing by the standard deviation. The formula for the standardized value is derived from the definition of standard scores in statistics. Standardized values are calculated independently for each channel, and there is no cross-coupling between different channels.
[0043] The feature extraction module extracts characteristic parameters reflecting the process status from the raw data, including mean, variance, peak value, peak-to-peak value, and power spectral density. Each feature is calculated independently for each channel, and no cross-coupling operations are performed between different channels. After receiving multi-channel synchronous data, the edge computing node sequentially performs filtering, time alignment, standardization, and feature extraction for each channel, outputting preprocessed standardized data, multi-dimensional feature vectors, and missing value markers. Outlier identification and isolation are performed in step S105; outlier removal is not performed in step S104.
[0044] Step S105: The anomaly filtering module automatically identifies and marks abnormal data points based on physical constraint rules and statistical criteria, completing the outlier removal. The physical constraint rules are: temperature between 0 and 300 degrees Celsius, pressure between 0 and 2 MPa, and flow rate greater than or equal to 0. The statistical criterion is: if the standardized value output in step S104... If the absolute value of the deviation is greater than 3, meaning the absolute deviation of the current point from the average of the historical normal operating condition data of the same channel is greater than 3 times the standard deviation of the historical normal operating condition data of the same channel, then the current point is determined to be a statistical outlier. The criteria for determining the statistical outlier are as follows: .
[0045] in, The standardized value output from step S104 is dimensionless, with a mean of 0 and a standard deviation of 1; 3 is a constant, representing 3 times the standard deviation. The judgment condition is: .
[0046] This indicates that the absolute value of the standardized value is greater than 3, meaning the current point deviates from the mean of historical normal operating condition data for the same channel by more than three times the standard deviation of historical normal operating condition data for the same channel. The statistical criterion originates from statistical process control. Guidelines.
[0047] The anomaly filtering module receives the standardized data output from step S104 and the multidimensional feature vector extracted in step S104. It then determines whether the temperature, pressure, and flow rate meet the physical constraint rules for each channel and whether the current feature value meets the statistical criteria, outputting the anomaly marking results. Points that meet the physical range and whose standardized absolute value does not exceed 3 are judged as normal points and enter the subsequent analysis stage; points that exceed the physical range or whose standardized absolute value is greater than 3 are judged as anomalies, marked, isolated, and their timestamps and channel numbers are recorded.
[0048] For example, an edge computing node in a flotation frother production plant collects data on reactor temperature, pressure, flow rate, and raw material composition A at a sampling period of 100ms. The current filtered value is obtained after filtering five consecutive sampling points in the temperature channel. Celsius, historical average of normal operating conditions Celsius, standard deviation Celsius. Substituting into the standardized value formula yields... . Temperature points are identified as statistical anomalies, their timestamps and channel numbers are recorded, and they are then isolated.
[0049] Current filter value for pressure channel MPa, historical average value under normal operating conditions MPa, standard deviation MPa , The pressure point was determined to be within the normal range, and the analysis proceeded to the next stage. The mass fraction of raw material component A shifted from 15% to 14.5%, corresponding to the channel standardized value. If it is identified as a statistical outlier, it will be isolated and will not be included in the feature calculation of the current sampling period, but will wait for the next period.
[0050] Steps S104 and S105 complete data preprocessing and anomaly filtering. Standardized data retains the physical meaning of process quantities such as temperature, pressure, flow rate, and composition, and outliers are isolated, providing subsequent deep Q-networks, intelligent decision engines, and digital twin models with controllable input quality.
[0051] In step S106, the processed standardized data is uploaded to the cloud monitoring platform via an industrial Ethernet communication interface. The communication protocol adopts the standard Modbus TCP or OPC UA protocol. The industrial Ethernet communication interface receives the standardized data output in step S104, the multi-dimensional feature vector output in step S104, and the anomaly labeling results output in step S105, encapsulates them into industrial Ethernet data frames according to the standard Modbus TCP or OPC UA protocol, and outputs a reliable data stream to the cloud monitoring platform.
[0052] Thus, step S1 completes the unification of data entry points for the flotation frother production process, from on-site sensor deployment to edge computing node processing and cloud monitoring platform reception. Temperature, pressure, flow rate, raw material composition, and equipment status are converted into continuous, synchronous, and traceable data streams, providing a data foundation for step S2 to build a deep reinforcement learning-driven process parameter optimization model.
[0053] Next, step S2 in the embodiments of this application will be described, which mainly involves constructing a deep reinforcement learning-driven process parameter optimization model. The specific implementation process is as follows.
[0054] In step S201, the cloud platform receives historical production data and real-time data streams from edge computing nodes to train and update the deep Q-network. The deep Q-network consists of 5 fully connected hidden layers, with 256, 128, 64, 32, and 16 neurons in each layer, respectively. The hidden layer activation function uses rectified linear units, and the output layer uses a linear activation function, directly outputting the Q-value of each candidate action.
[0055] The number of neurons in the input layer equals the dimension of the state space vector, which ranges from 32 to 64 dimensions and includes raw material ratio parameters, temperature parameters, pressure parameters, flow rate parameters, and historical time window operating condition features. The number of neurons in the output layer equals the total dimension of the action space, which is 2000. The network parameters are initialized using a normal distribution with a standard deviation of 0.1.
[0056] The training data is constructed as follows: a state space vector is extracted from the standardized data stream output in step S1. Each training sample consists of the current state, the action performed in the current state, the reward value obtained after performing the action, and a four-tuple for the next state. Historical production data is used to construct the initial experience pool, and real-time data streams are used to continuously replenish the experience pool. The data cleaning standard follows the normal point data output in step S105; outlier data is not included in the experience pool. Training samples do not require manual labeling, and the reward value is calculated and generated in real time by the reward function in step S205.
[0057] The state space dimensions are dynamically adjusted based on the actual number of process parameters. When an adjustable process parameter increases, the corresponding parameter dimension is added to the state space vector. When an adjustable process parameter decreases, the corresponding parameter dimension is deleted from the state space vector. There is no cross-coupling between the dimensions, and each component in the state space vector takes values independently.
[0058] In step S202, the loss function of the deep Q-network adopts the mean squared Bellman error form. The input of the loss function is the Q-value of the current network output and the target Q-value, and the output is the square of the deviation between the two. The target Q-value is composed of the sum of the immediate reward value and the discount factor multiplied by the maximum Q-value of the next state. The learning rate is set to 0.001, the discount factor is set to 0.95, the parameter optimizer adopts the Adam algorithm, and the training convergence objective is to reduce the loss function value to below 0.01 or complete 10,000 training epochs.
[0059] In step S203, the capacity of the experience replay buffer is set to 100,000 experience records, and historical experience is extracted from the buffer using a uniform random sampling method for network training. The network parameters are updated synchronously every 100 time steps.
[0060] In the production of flotation frothers, fluctuations in raw material composition, changes in reactor temperature and pressure, and differences in equipment condition can all affect the optimal values of process parameters. The optimal value refers to the reward value calculated by the reward function in step S205. The maximum value of the process parameter.
[0061] When the mass fraction of key component A in the raw material shifts from 15% to 14.5%, the state space update module incorporates the shift in the mass fraction of key component A into the current state representation, and the deep Q-network re-evaluates the Q-value of each candidate action. The action with the maximum Q-value among all candidate actions is taken as the current optimal decision.
[0062] The maximum Q-value is calculated by comparing the Q-values of each candidate action and selecting the action with the largest Q-value. If multiple candidate actions have the same maximum Q-value, the action with the smallest action number is selected as the current optimal decision.
[0063] Step S204: The motion space employs a discretization strategy, dividing the continuous parameter adjustment range into multiple discrete levels. The reaction temperature is divided into 20 discrete levels, each corresponding to an adjustment step of 0.5 degrees Celsius, with each level evenly distributed upwards and downwards from the current reaction temperature setting. Pressure is divided into 10 discrete levels, each corresponding to an adjustment step of 0.01 MPa, with each level evenly distributed upwards and downwards from the current pressure setting. Flow rate is divided into 10 discrete levels, each corresponding to an adjustment step of 5% of the rated flow rate, with each level evenly distributed from 75% to 120% of the rated flow rate. The total dimension of the motion space is equal to the product of the number of discrete levels for all adjustable parameters. .
[0064] Step S205: The reward function design comprehensively considers three dimensions: product quality indicators, production efficiency, and energy consumption indicators, and calculates the reward value using a weighted sum method.
[0065]
[0066] in, The reward value is dimensionless and ranges from 0 to 1. This is a product quality award item, dimensionless, with a value range between 0 and 1; This is a production efficiency incentive item, dimensionless, with a value ranging from 0 to 1; This is an energy consumption reward item, dimensionless, with a value ranging from 0 to 1.
[0067] The reward function is derived from the weighted and scalar methods of multi-objective optimization in deep reinforcement learning. The calculation method is to use Multiply by 0.5, Multiply by 0.3, Multiply by 0.2, then add the products of the three terms. , , , All are scalars, calculated independently as scalars and then weighted and summed, without vector or matrix coupling operations.
[0068] The calculation method is as follows: the key product indicators include two indicators: the content of the active component of the foaming agent and the foam half-life. Let the number of key product indicators be... , No. The detection values of key indicators for each product The target value is ,in Take 1 to Integer. The first The relative deviation of key indicators for each product is
[0069]
[0070] Product quality awards equal Subtract all The arithmetic mean of the relative deviations of key product indicators, i.e.
[0071]
[0072] The target value for the active ingredient content of the foaming agent is set at 15%, and the target value for the foam half-life is set at 180 seconds. The smaller the relative deviation of the key indicators of each product, the better. The larger.
[0073] The calculation method is as follows: obtain the actual output and rated capacity per unit time, divide the actual output by the rated capacity to obtain the production efficiency ratio, and then calculate the production efficiency bonus. This equals the production efficiency ratio. The closer the actual output is to the rated capacity, the better. The larger.
[0074] The calculation method is as follows: obtain the unit product energy consumption and the benchmark energy consumption, divide the benchmark energy consumption by the unit product energy consumption to obtain the energy consumption ratio, and then calculate the energy consumption bonus. This equals the energy consumption ratio. The lower the energy consumption per unit of product, the better. The larger.
[0075] 3 reward items , and All values have been normalized, ranging from 0 to 1, with unified dimensions, allowing for direct weighted summation. The specific calculation method for weighted summation is as follows: ... Multiply by 0.5, Multiply by 0.3, Multiply by 0.2, and add the products of the three terms to get the reward value. The reward value is updated in real time based on product quality inspection results and production data, optimizing the model output candidate parameter adjustment scheme set and providing input for subsequent digital twin pre-simulation verification.
[0076] For example, a flotation frother production plant deployed the adaptive process parameter control equipment of this invention. A raw material composition analyzer monitors the raw material ratio entering the reactor in real time, detecting a shift in the mass fraction of key component A from the target value of 15% to 14.5%. The detection result is transmitted to an edge computing node via industrial Ethernet. The edge computing node preprocesses and extracts features from the raw material ratio data, then transmits the standardized data to a cloud platform. The state space update module of the cloud platform receives the data and incorporates the shift in the mass fraction of key component A into the state space vector of a deep Q-network.
[0077] The Deep Q-Network, based on the updated state space vector, propagates forward through five fully connected hidden layers, and the output layer outputs the Q-value for each candidate action. The Deep Q-Network selects the action with the highest Q-value as the current optimal decision; assuming the optimal decision is to increase the reaction temperature from 150 degrees Celsius to 151 degrees Celsius, the parameter adjustment suggestion is sent to the digital twin pre-simulation verification module. This module loads the current production state as the initial simulation condition, takes the temperature adjustment suggestion as input, and runs the digital twin model to simulate the changing trend of the production process over the next 30 minutes.
[0078] Simulation results show that the overall product quality score improved from 95 to 97, production efficiency remained unchanged, and energy consumption increased by 2%. With a positive overall evaluation, the cloud platform sends a temperature adjustment command to the edge control node for execution. The edge control node adjusts the power output of the heating system through a PID controller, gradually raising the reaction temperature to 151 degrees Celsius. The edge computing node continuously monitors the temperature change trajectory; if the deviation between the actual temperature value and the target value converges to within ±0.3% within 10 seconds, the adjustment is considered successful, the state space is updated, and the adjustment experience is recorded in the experience replay buffer.
[0079] Reward Value After temperature adjustment, it increased from 0.72 to 0.81. Among them... It increased from 0.78 to 0.91. The value remains unchanged at 0.85. The reward value decreased from 0.52 to 0.48. The substitution operation is as follows:
[0080]
[0081] Rounded to two decimal places, 0.81, then weighted summed. It is on the rise.
[0082] In summary, step S2 establishes a mapping relationship from production status to process parameter adjustment actions through a deep Q-network. The state space vector consists of the standardized data output from step S1 and multi-dimensional feature vectors. The action space covers discrete levels of adjustable parameters such as reaction temperature, pressure, and flow rate. The reward function uses the weighted sum of product quality, production efficiency, and energy consumption as the evaluation criterion.
[0083] After the network training is completed, the optimized model outputs a set of candidate parameter adjustment schemes, which are used by the intelligent decision engine in step S3 for abnormal working condition diagnosis and parameter drift compensation, and are also used by the digital twin virtual production model in step S4 for pre-run verification.
[0084] After step S2 is completed, step S3 in this embodiment of the application will be described next. This mainly involves developing an intelligent decision engine module to realize automatic diagnosis of abnormal working conditions, automatic compensation for parameter drift, and intelligent identification of equipment faults. Specifically, it will be carried out according to the following sub-steps.
[0085] In step S301, the intelligent decision engine is deployed on edge computing nodes and the cloud platform, forming a layered decision architecture. The edge-side decision engine is responsible for real-time response and handling latency-sensitive control decisions. The cloud-side decision engine is responsible for complex operating condition analysis and the formulation of global optimization strategies.
[0086] The intelligent decision engine receives the standardized data stream and multi-dimensional feature vector output from step S1, as well as the candidate parameter adjustment scheme set output from step S2. It performs automatic diagnosis of abnormal operating conditions, automatic compensation for parameter drift, and intelligent identification of equipment faults, and outputs control commands, compensation adjustment amounts, and fault diagnosis results.
[0087] Step S302: The automatic parameter drift compensation function continuously monitors the deviation between the actual value and the set value or model prediction value of each process parameter. When the deviation exceeds a threshold, the compensation mechanism is triggered. The compensation strategy adopts an incremental adjustment method, using a fixed percentage of the current deviation value as the adjustment amount.
[0088] The parameter deviation is calculated by subtracting the actual value from the set value. ,Right now
[0089]
[0090] in, Indicates the deviation value. This indicates the set value of the process parameters. This represents the actual value of the process parameters. and Same dimensions Its dimensions are the same as both. Deviation value A positive value indicates that the actual value is lower than the set value, representing the deviation value. A negative value indicates that the actual value is higher than the set value.
[0091] The compensation adjustment is the deviation value multiplied by the incremental adjustment coefficient. ,Right now
[0092]
[0093] in, Indicates the amount of compensation adjustment. Indicates the incremental adjustment coefficient. This indicates the deviation value. The coefficient is dimensionless. Dimensions and Same. Incremental adjustment coefficient The setting is 0.3. When the reaction temperature is set to 150 degrees Celsius and the actual temperature is 149 degrees Celsius, the deviation value is... The adjustment amount is 1 degree Celsius. The control output is increased by 0.3 degrees Celsius to adjust the reaction temperature toward the set value.
[0094] Step S303: The parameter drift detection sensitivity is ±0.3%. The deviation percentage is calculated by dividing the absolute value of the deviation by the set value and then multiplying by 100%. Compensation is triggered when the absolute value of the deviation percentage exceeds 0.3%. For a reaction temperature setpoint of 150 degrees Celsius, the corresponding absolute deviation is 0.45 degrees Celsius. That is, temperature fluctuations within the range of 149.55 to 150.45 degrees Celsius are normal; exceeding this range triggers the compensation mechanism. The compensation response time is less than 2 seconds, and the overshoot is less than 5%.
[0095] In the production of flotation frothers, minute drifts in reaction temperature, pressure, and flow rate, if not corrected in time, will gradually accumulate and affect the content of active components in the frother and the foam half-life. The edge-side decision engine, upon detecting a deviation exceeding a threshold, adjusts gradually with an increment coefficient of 0.3 to avoid overshoot or oscillation caused by a large, one-time adjustment.
[0096] Step S304: Intelligent equipment fault identification employs a multi-dimensional feature fusion method, comprehensively processing the time-domain and frequency-domain features of vibration, temperature, and current signals. Time-domain features include mean, standard deviation, peak factor, and kurtosis index. Frequency-domain features are extracted using Fast Fourier Transform (FFT) to determine the main peak frequency, frequency band energy percentage, and spectral entropy. The time-domain and frequency-domain features together constitute a 10-dimensional feature vector. Each feature is calculated independently per channel.
[0097] Step S305: After normalization, the multi-source features are input into the fault classification model. The normalization method is as follows: for each feature, the mean and standard deviation of historical normal operating data are obtained, and the current value is subtracted from the mean and then divided by the standard deviation to obtain a standardized feature value with zero mean and unit variance. The fault classification model adopts a hybrid architecture of support vector machine and deep neural network.
[0098] The Support Vector Machine (SVM) layer uses a radial basis function kernel with a kernel width of 0.5 and a penalty coefficient of 1.0, outputting initial fault category labels. The SVM layer aims to maximize the classification margin and is trained until the classification margin no longer increases or the maximum number of iterations is reached.
[0099] The deep neural network layer consists of three fully connected hidden layers with 64, 32, and 16 neurons respectively. The hidden layer activation function uses rectified linear units (RCLs), and the output layer uses a flexible maximum function to output the probability distribution of each fault type. The 10-dimensional feature vector is concatenated with the initial fault category label to form an 11-dimensional vector, which is then input into the deep neural network layer.
[0100] The training data source is the standardized data stream and multidimensional feature vector output from step S1, combined with vibration signals, temperature signals, and current signals collected by the equipment condition monitoring module. Data cleaning criteria include removing missing data and outlier data marked in step S105. Labeling rules are based on actual maintenance records to label fault types, including normal state, rolling bearing wear, pump cavitation, and valve jamming. The deep neural network layer training uses a learning rate of 0.001, batch size of 64, training epochs of 200, cross-entropy loss function, and Adam optimizer.
[0101] Step S306: The fault identification module supports automatic identification of fault types such as rolling bearing wear, pump cavitation, and valve jamming, with a fault identification accuracy rate of over 95% and an identification response time of less than 500ms. The identification results trigger processing measures through the self-healing control module, including early warning notifications, load adjustments, equipment start-stop switching, and maintenance work order generation. The matching rules for the handling strategy are as follows: when the fault probability is below 80%, only an early warning notification is triggered; between 80% and 95%, an early warning notification and load adjustment are triggered; and above 95%, an early warning notification, load adjustment, and equipment start-stop switching are triggered.
[0102] For example, during the operation of the agitator motor in the reactor of a flotation frother production plant, the equipment status monitoring module collects vibration, temperature, and current signals. Edge computing nodes extract time-domain and frequency-domain features to form a 10-dimensional feature vector, which, after normalization, is input into a fault classification model. The support vector machine layer initially screens it as a suspected bearing fault. The 10-dimensional feature vector is then concatenated with the initial fault category label to form an 11-dimensional vector, which is input into a deep neural network layer. This results in a refined identification of slight wear on the outer ring of the rolling bearing, with a fault probability of 92%. The self-healing control module triggers an early warning notification and load adjustment based on the 92% fault probability. Performance verification: On a test set of 1000 samples, 955 faults were correctly identified, achieving an accuracy of 95.5%, with an average response time of 320ms.
[0103] Steps S304 to S306 form an intelligent fault identification and self-healing control link for equipment. A support vector machine layer performs rapid initial screening, a deep neural network layer performs fine classification, and the self-healing control module matches handling strategies based on fault type and severity to avoid the impact of unplanned downtime on production continuity.
[0104] Therefore, step S3, through hierarchical decision-making architecture, parameter drift compensation, intelligent fault identification, and self-healing control, transforms real-time data streams and parameter adjustment suggestions into control actions, providing a decision-making basis for the digital twin pre-simulation in step S4, and ensuring the continuous and stable operation of the flotation frother production process.
[0105] Based on the above analysis, step S4 in the embodiments of this application will be described in detail. Step S4 mainly involves establishing a digital twin virtual production model for real-time simulation and pre-run optimization, as detailed below.
[0106] In the production of flotation frothers, reaction temperature, pressure, flow rate, and the content of key raw material components all affect the formation rate of active components and the foam half-life. If the adjustment scheme output by the deep Q-network is directly issued, fluctuations in raw material conditions and differences in equipment status may lead to batch quality risks. Therefore, step S4 first pre-simulates candidate parameter adjustment schemes in a digital twin environment, and then selects the recommended scheme based on a comprehensive evaluation of product quality indicators, production efficiency indicators, and energy consumption indicators.
[0107] Step S401: Construct a digital twin model corresponding to the physical production line. The digital twin model employs a hybrid modeling approach combining mechanistic and data-driven models. The mechanistic model describes the modelable physicochemical laws in the flotation frother production process. The data-driven model corrects for nonlinear effects and uncertainties that the mechanistic model fails to describe. The mechanistic and data-driven models work synergistically to improve the overall prediction accuracy of the digital twin model.
[0108] Step S402: The mechanistic model is constructed based on chemical reaction kinetics and heat and mass transfer theories, describing the physical laws governing the chemical reaction pathways, reaction rates, material balance, and energy balance during the production of flotation frothers. The mechanistic model is expressed in the form of a system of differential equations, including reaction kinetic equations, heat transfer equations, and mass transfer equations.
[0109] Reaction kinetic equations describe the relationship between the consumption rate of reactants and the formation rate of products and temperature and pressure. Heat transfer equations describe the temperature distribution and heat transfer process of the reaction system. Mass transfer equations describe the transfer of matter between phases. Parameters of the mechanistic model include the Arrhenius constant, activation energy, heat capacity, and thermal conductivity, which are obtained through experimental calibration or literature data.
[0110] Step S403: The data-driven model uses a Long Short-Term Memory (LSTM) network to supplement and correct the complex nonlinear process. The LSM network contains three gating units: an input gate, a forget gate, and an output gate, which handle long-term dependencies in time-series data. The LSM network contains two hidden layers: the first hidden layer has 64 neurons, and the second hidden layer has 32 neurons. The activation function for the hidden layers is the hyperbolic tangent function, and the activation function for the output layer is a linear function.
[0111] The data-driven model takes the main state variables and input variables of the mechanistic model as input and outputs the deviation correction between the model's predicted and actual values. The training data is constructed as follows: the main state variables and input variables of the mechanistic model are extracted from historical production data; the deviation between the model's predicted and actual values at the same time is calculated; the main state variables and input variables are used as inputs to the Long Short-Term Memory (LSTM) network; and the deviation between the model's predicted and actual values is used as the target output of the LSM network, thus forming input-output sample pairs.
[0112] The data cleaning standard is to remove missing data and outliers marked in step S105. The annotation rule is to use the deviation between the predicted and actual values from the mechanistic model as the target output, requiring no manual annotation. Training hyperparameters are set as follows: learning rate 0.001, batch size 64, training epochs 1000, and the Adam algorithm is used as the optimizer. The mean squared error loss function is used, with the inputs being the deviation correction from the Long Short-Term Memory network output and the actual deviation correction, and the output being the mean of their squared errors. The training convergence target is for the loss function value to drop below 0.01 or for 1000 training epochs to be completed.
[0113] In the production of flotation frothers, reaction kinetics and heat and mass transfer theories provide physical constraints, while long short-term memory networks compensate for nonlinear deviations caused by fluctuations in raw material composition and changes in equipment status, enabling digital twin models to reflect the changing trends of temperature, pressure, and flow rate within the reactor and the impact of on-site disturbances.
[0114] Step S404: The model update cycle is set to 1 hour. The cloud platform periodically collects the latest production data, performs online identification and calibration of the mechanism model parameters, and incrementally trains the data-driven model to maintain the synchronization between the digital twin model and the actual production process. The pre-run verification of the digital twin model takes less than 10 seconds, and the simulation accuracy reaches ±3%.
[0115] Step S405, the execution flow of the digital twin pre-simulation optimization mechanism is as follows: The candidate parameter adjustment schemes output by the optimization model are numbered and marked, and a mapping relationship between scheme identifiers and parameters is established. Each candidate parameter adjustment scheme is loaded sequentially in the digital twin virtual environment, and the current production state is used as the initial simulation condition. The simulation model is then run to simulate the production process within the next 30 minutes.
[0116] Considering that a single digital twin simulation takes less than 10 seconds, and the deep Q-network outputs a large number of candidate parameter adjustment schemes, to improve engineering feasibility, step S405 adopts a candidate parameter adjustment scheme screening strategy: first, the deep Q-network outputs the Q-values of each candidate action, sorts them from high to low Q-values, and only selects the top N candidate parameter adjustment schemes with the highest Q-values to be sent to the digital twin model for pre-simulation verification, where N is between 5 and 10, instead of simulating all candidate parameter adjustment schemes one by one. This candidate parameter adjustment scheme screening strategy controls the total pre-simulation verification time to within 50 to 100 seconds, meeting the timeliness requirements for parameter adjustment during the flotation frother production process.
[0117] There are two feedback loops between the pre-verification mechanism and the deep Q-network in step S2. The first loop is the action space constraint: during the training phase, the deep Q-network uses the pre-verification results output by the digital twin model as the product quality reward term in the reward function. The calibration basis. This is a product quality award item, dimensionless, with a value range between 0 and 1.
[0118] When the digital twin model predicts the product quality indicators within the next 30 minutes based on the candidate parameter adjustment scheme... When the performance falls below the desired threshold, the reward value of the action corresponding to the candidate parameter adjustment scheme in the experience replay buffer is reduced. Product quality metrics The effect threshold is set to 0.8.
[0119] The reduction in reward value depends on product quality indicators. The deviation from the effect threshold is determined; the larger the deviation, the greater the decrease in the reward value. The deviation is defined as 0.8 minus the product quality index. The bias ranges from 0 to 0.8. The reward value decreases linearly with respect to the bias, ranging from 0.1 to 0.5. When the bias is 0, the decrease is 0.1, and when the bias is 0.8, the decrease is 0.5. The deep Q-network reduces the probability of the corresponding action being selected during subsequent training.
[0120] The second closed loop is state space update: when the comprehensive score of all candidate parameter adjustment schemes is lower than 0.8 and the feedback mechanism is triggered, the simulation deviation data output by the digital twin model is fed back to the edge computing node in step S1 as a supplementary input to the feature extraction module, so that the subsequent standardized data contains the deviation information of the pre-verification.
[0121] The aforementioned dual feedback loop couples the digital twin pre-simulation verification with the training and state space update processes of the deep Q-network, forming an adaptive control loop of perception, decision-making, pre-simulation, feedback, and re-decision-making. During simulation, the trajectory of key process parameters and the predicted values of product quality indicators are recorded. After simulation, the simulation results of each candidate parameter adjustment scheme are extracted, and product quality indicators, production efficiency indicators, and energy consumption indicators are extracted as evaluation criteria.
[0122] Step S406, the calculation method for the evaluation indicators is as follows: Product quality indicators adopt a weighted comprehensive scoring method, comprehensively considering the deviation of key product indicators from target values, indicator fluctuation range, and defective product rate. Production efficiency indicators adopt the ratio of actual output to rated capacity. Energy consumption indicators adopt the ratio of benchmark energy consumption to unit product energy consumption.
[0123] Based on the evaluation results, the candidate parameter adjustment schemes are ranked, and the scheme with the highest comprehensive score is selected as the recommended scheme. If none of the candidate parameter adjustment schemes reach the effect threshold, a feedback mechanism is triggered, returning to step S2 to regenerate the candidate parameter adjustment scheme set and adjust the optimization objective or constraints.
[0124] Product quality indicators The calculation method is as follows: Let the key product indicators be... The, the The detection values of key indicators for each product The target value is . No. The relative deviation of the key indicators of each product is the first The absolute value of the difference between the key indicator test value and the target value of each product divided by the first... Target values for key product metrics.
[0125] Deviation Item Subtract all from 1 The arithmetic mean of the relative deviations of key indicators for each product. Indicator fluctuation range. The allowable fluctuation range is the difference between the maximum and minimum values detected within a 30-minute time window. Take 5% of the target value. When the indicator fluctuates... Less than or equal to the allowable fluctuation range At that time, the fluctuation term Set to 1; when the indicator fluctuates within a certain range Greater than the allowable fluctuation range At that time, the fluctuation term Subtract the range of index fluctuations from 1 With the allowable fluctuation range The ratio of the percentage of defective products. The defect rate is the predicted percentage of defective products out of total production. Subtract the defect rate from 1 Product quality indicators for:
[0126]
[0127] in, It is a product quality indicator, dimensionless, and its value ranges from 0 to 1; This is the deviation term, which is dimensionless and ranges from 0 to 1. This is a fluctuation term, dimensionless, with values ranging from 0 to 1; This is the non-conforming rate item, which is dimensionless and ranges from 0 to 1. The calculation method is to use Multiply by 0.5, Multiply by 0.3, Multiply by 0.2 and add the products of the three terms. , , , All are scalars, calculated independently as scalars and then weighted and summed, without vector or matrix coupling operations.
[0128] Production efficiency indicators This is calculated by dividing actual output by rated capacity. (Production efficiency indicator) Dimensionless, with values ranging from 0 to 1. Energy consumption index. The energy consumption index is calculated by dividing the baseline energy consumption by the energy consumption per unit of product. Dimensionless, with values ranging from 0 to 1, the lower the energy consumption per unit of product, the higher the energy consumption index. The larger. Overall score. for:
[0129]
[0130] in, For the overall score, it is dimensionless and its value ranges from 0 to 1; For product quality indicators; As a production efficiency indicator; Energy consumption indicators. The calculation method is to use Multiply by 0.5, Multiply by 0.3, Multiply by 0.2 and add the products of the three terms. , , , All are scalars, calculated independently as scalars and then weighted and summed, without vector or matrix coupling operations.
[0131] Select Overall Score The candidate parameter adjustment scheme with the largest value is selected as the recommended scheme, which is the comprehensive score of all candidate parameter adjustment schemes. Compare the values and take the largest one. The corresponding candidate parameter adjustment schemes. If there are multiple candidate parameter adjustment schemes, the overall score is... If they are the same and all are the maximum values, then the energy consumption index is selected. The option with the larger value is recommended. The effectiveness threshold is set as the overall score. If the overall score of all candidate parameter adjustment schemes is less than 0.8, the feedback mechanism is triggered.
[0132] For example, the reactor temperature setpoint of a flotation frother production plant is 150 degrees Celsius. The mass fraction of key component A in the current raw material has shifted from the target value of 15% to 14.5%. In step S2, the deep Q-network outputs a set of candidate parameter adjustment schemes, including: Candidate parameter adjustment scheme 1, increasing the reaction temperature to 151 degrees Celsius; Candidate parameter adjustment scheme 2, increasing the reaction temperature to 152 degrees Celsius; and Candidate parameter adjustment scheme 3, maintaining the reaction temperature at 150 degrees Celsius and adjusting the flow rate. The digital twin model loads the current production state as the initial simulation condition and simulates the three candidate parameter adjustment schemes sequentially for 30 minutes.
[0133] The simulation results of candidate parameter adjustment scheme 1 are as follows , , Overall score The simulation results for candidate parameter adjustment scheme 2 are as follows: , , Overall score The simulation results for candidate parameter adjustment scheme 3 are as follows: , , Overall score .
[0134] Candidate parameter adjustment scheme 1 achieved the highest overall score of 0.806, which is greater than the effect threshold of 0.8. Therefore, candidate parameter adjustment scheme 1 was selected as the recommended scheme. After implementing candidate parameter adjustment scheme 1, the actual product quality score improved from 95 to 97, production efficiency remained unchanged, and energy consumption increased by 2%, which is consistent with the prediction results of the digital twin simulation.
[0135] In the production of flotation frothers, the above process allows for the early identification of adjustment risks caused by fluctuations in reaction temperature, pressure, flow rate, and raw material composition in a virtual environment, thus preventing unqualified batches from entering actual production.
[0136] Therefore, step S4 establishes a digital twin virtual production model using a hybrid modeling method that combines mechanistic models and data-driven models. The candidate parameter adjustment schemes are then pre-tested and verified in the digital twin virtual environment. The recommended scheme is selected based on the evaluation results of product quality indicators, production efficiency indicators, and energy consumption indicators.
[0137] The pre-test results of the digital twin model provide a basis for decision-making in step S3. Candidate parameter adjustment schemes that fail the pre-test trigger a feedback mechanism to return to step S2 to regenerate the set of candidate parameter adjustment schemes, forming a closed loop from parameter optimization to pre-test and then to decision execution.
[0138] Finally, step S5 in the embodiments of this application is described, which mainly involves building a distributed architecture that coordinates the cloud monitoring platform and the edge computing nodes. The specific implementation process is as follows.
[0139] Step S501: The distributed architecture adopts a layered design, with the edge layer responsible for real-time data acquisition and local rapid response, and the cloud layer undertaking big data storage, complex analysis and global optimization tasks.
[0140] Edge computing nodes are deployed on the production site, equipped with independent data processing and communication units, to achieve data preprocessing, feature extraction, and real-time control decision-making at the edge. In the production of flotation frothers, the edge layer is directly connected to field equipment such as reactors, raw material pumps, and stirring motors, maintaining the normal operation of the local control loop even when there are network fluctuations or cloud latency.
[0141] Step S502: The communication between the edge node and the cloud platform is achieved through asynchronous communication using a message queue middleware. The message queue middleware uses distributed message queue products such as Kafka or RabbitMQ, and adopts a distributed deployment method to support persistent message storage and reliable transmission.
[0142] Edge nodes encapsulate the processed, standardized data into messages and send them to the message queue middleware. Message consumers on the cloud platform retrieve the data from the message queue middleware for subsequent storage and analysis. Asynchronous communication isolates the timing coupling between the edge and cloud sides, ensuring stable uploading of temperature, pressure, flow rate, and raw material composition data during flotation frother production despite network fluctuations.
[0143] Step S503: Configure a distributed database on the cloud platform to store historical data. The distributed database uses database products such as MongoDB or Cassandra that support partitioning, sharding, and replica set mechanisms. The replica set mechanism ensures high data availability. Data storage adopts a partitioning and sharding strategy, partitioning data according to time range and device identifier, supporting efficient time series queries and device-level aggregation analysis.
[0144] The storage period is set to no less than 3 years to meet the data requirements for traceability and long-term trend analysis of the flotation frother production process. The distributed database supports horizontal scaling, allowing for smooth expansion by adding data nodes.
[0145] Steps S501 to S503 construct a distributed data architecture that coordinates the edge layer and the cloud layer. The edge layer is responsible for real-time data acquisition and local control, while the cloud layer is responsible for historical storage and global analysis. A message queue middleware enables asynchronous communication between the two. In the production of flotation frothers, this hierarchical architecture allows data on reactor temperature, pressure, flow rate, and raw material composition to be reliably uploaded to the cloud after preprocessing at the edge. The cloud platform can then uniformly manage and perform cross-batch analysis of historical data from multiple production units.
[0146] Step S504: The remote visual monitoring interface adopts a B / S architecture, supporting real-time display of process parameters, trend analysis, alarm management, and report generation. The real-time display module shows the current value, set value, and trend of key process parameters in the form of a dashboard. The trend analysis module provides functions for querying, comparing, and predicting historical data.
[0147] The alarm management module supports configuring alarm thresholds, querying alarm records, and managing alarm handling processes. The report generation module supports custom report templates and automatically generates daily, weekly, and monthly reports. The monitoring interface supports multi-terminal access, allowing operators to monitor the flotation frother production status in real time via PC or mobile devices.
[0148] Step S505: The production process traceability management module establishes a data recording system covering the entire process from raw material warehousing, production and processing, quality inspection, and product delivery. Each data record includes key fields such as timestamp, equipment number, operator, parameter data, and quality inspection results.
[0149] The timestamps are accurate to the millisecond level, the equipment numbers correspond one-to-one with the actual equipment, the operators record the identity information of the logged-in users, the parameter data records the real-time collected values and statistical values of each process parameter, and the quality inspection results record the test values and judgment conclusions of the key product indicators.
[0150] Data records are anchored and stored using blockchain technology. The blockchain employs a consortium blockchain architecture, where the hash value of each data record is written to the blockchain via smart contracts. Leveraging the decentralized and immutable characteristics of blockchain, the authenticity and integrity of the data records are ensured. The traceability query function supports precise retrieval based on time range, equipment number, product batch, and other criteria, enabling retrospective analysis of the flotation foaming agent production process and root cause analysis of quality issues at any given time.
[0151] Steps S504 and S505 complete remote monitoring and traceability management. The remote visual monitoring interface provides maintenance personnel with real-time images and historical trends of the flotation frother production process. The alarm management module promptly notifies operators when process parameters exceed normal ranges. The traceability management module uses blockchain-anchored storage to ensure data records are tamper-proof. When product quality issues occur, the root cause can be quickly located according to time range, equipment number, or product batch.
[0152] Finally, step S5 integrates the standardized data streams, parameter adjustment schemes, fault diagnosis results, and pre-drill verification results generated in steps S1 to S4 into the cloud monitoring platform through a distributed architecture, asynchronous communication, distributed database, remote monitoring interface, and traceability management module.
[0153] Edge computing nodes work in tandem with the cloud platform to achieve a balance between local real-time control and remote global optimization in the flotation frother production process. The remote monitoring interface and traceability management module provide operators and management with a visual representation of the production process and the ability to retrieve historical data, supporting the continuous and stable operation of the flotation frother production process and the ongoing traceability of product quality.
[0154] To further enhance the understanding of the embodiments of this application, the technical solutions of the present invention will be further described below in conjunction with specific application examples.
[0155] After a flotation frother production plant deployed the adaptive process parameter control equipment of the present invention, the actual operation process is as follows.
[0156] During the raw material feeding stage, a raw material composition analyzer monitors the proportion of raw materials entering the reaction system in real time and transmits the results to an edge computing node via an industrial Ethernet network. The edge computing node preprocesses and extracts features from the raw material proportion data and transmits the standardized data to the cloud platform.
[0157] The cloud platform's state space update module receives data and updates the state space vector of the deep reinforcement learning network. If the mass fraction of key component A in the raw material is detected to have shifted from the target value of 15% to 14.5%, the state space update module incorporates this shift into the current state representation.
[0158] Deep reinforcement learning networks compute the action space based on the updated state space vector. The output layer of a deep Q-network outputs a Q-value for each candidate action, representing an estimate of the long-term cumulative reward obtained after executing the action. The action with the highest Q-value among all candidate actions is taken as the current optimal decision. For example, the optimal decision might be to increase the reaction temperature from 150 degrees Celsius to 151 degrees Celsius.
[0159] The parameter adjustment suggestions are sent to the digital twin pre-simulation verification module. The pre-simulation verification module loads the current production status as the initial conditions for simulation, takes the temperature adjustment suggestions as input, and runs the digital twin model to simulate the changing trend of the production process over the next 30 minutes.
[0160] Simulation results show that the overall product quality score improved from 95 to 97, production efficiency remained unchanged, and energy consumption increased by 2%. According to the evaluation index calculation method described in step S406, the weighted composite score of the product quality index, production efficiency index, and energy consumption index is 0.806, which is greater than the effect threshold of 0.8, indicating a positive overall evaluation result. The cloud platform sends the temperature adjustment command to the edge control node for execution.
[0161] The edge control node receives temperature adjustment commands and uses a PID controller to adjust the power output of the heating system, gradually raising the reaction temperature to 151 degrees Celsius. The edge computing node continuously monitors the temperature change trajectory and provides real-time feedback on the adjustment effect.
[0162] If the deviation between the actual temperature value and the target temperature value converges to within ±0.3% within 10 seconds, the adjustment is considered successful, the state space is updated, and the adjustment experience is recorded in the experience replay buffer. If the temperature deviation continues to exceed the ±0.3% threshold, the automatic parameter drift compensation mechanism is triggered, guiding the temperature back to the target value through an incremental adjustment strategy.
[0163] During production operation, the equipment condition monitoring module continuously collects vibration, temperature, and current signals from key equipment. A periodic component with a frequency three times the power frequency was detected in the vibration signal of a stirring motor in a certain reactor. Spectral analysis showed an increase in the energy proportion of the high-frequency component.
[0164] The fault identification module extracts the time-domain and frequency-domain features of the vibration signal and inputs them into the fault classification model for identification. The identification results show that there is slight wear on the outer ring of the rolling bearing, with a fault probability of 92%. Based on the 92% fault probability, the self-healing control module triggers an early warning notification and load adjustment, pushes the fault information to the maintenance personnel's terminal, and records the fault diagnosis results and trigger time.
[0165] The self-healing control module assesses the scope of the fault's impact. If the fault does not affect current production continuity, it generates a maintenance work order to await planned repairs. If the fault may cause production interruption, it triggers a load reduction operation strategy to avoid unplanned downtime.
[0166] The cloud-based monitoring platform allows maintenance personnel to monitor production status in real time through a remote, visual monitoring interface. The left side of the interface displays real-time values of key process parameters for the current production line, including reaction temperature, pressure, flow rate, and the content of key components. The middle of the interface displays trend curves for each parameter, supporting selection and zooming of the time range. The right side of the interface displays current alarm information and operating status.
[0167] When an alarm event occurs, an alarm details window automatically pops up on the interface, displaying alarm parameters, alarm time, alarm level, and handling suggestions. Maintenance personnel can confirm and handle the alarm event through the interface, filling in handling records and remarks. The reporting module automatically generates daily reports based on templates, summarizing the day's production data, quality data, and equipment operation data for management review.
[0168] The above application examples demonstrate that the technical solutions in steps S1 to S5 form a closed loop in the production of flotation frothers, encompassing data acquisition, parameter optimization, decision execution, pre-testing and verification, and remote monitoring. The collaborative work of edge computing nodes and the cloud platform enables timely detection, optimization, and handling of changes in raw material properties, process parameter drifts, and equipment anomalies, supporting the continuous and stable operation of the flotation frother production process and the ongoing traceability of product quality.
[0169] On the other hand, an adaptive control system for flotation frother production process parameters disclosed in an embodiment of the present invention includes: a multi-source data acquisition and edge computing system, used to deploy temperature sensors, pressure sensors, flow sensors, raw material composition analyzers and equipment status monitoring modules at the flotation frother production site. The edge computing nodes receive multiple raw sensor signals at a sampling period of 100ms, and perform filtering, time alignment, standardization and feature extraction on a channel-by-channel basis. The anomaly filtering module automatically identifies and marks abnormal data points based on physical constraint rules and statistical criteria, and completes the removal of outliers.
[0170] A deep reinforcement learning-driven process parameter optimization model is used to receive standardized data streams and multidimensional feature vectors on a cloud platform, train and update a deep Q-network. The deep Q-network contains 5 fully connected hidden layers with 256, 128, 64, 32 and 16 neurons in each layer, respectively. The output layer outputs the Q-value for each candidate action, and the action with the maximum Q-value among all candidate actions is taken as the current optimal decision.
[0171] The digital twin virtual production model is used to receive the top N candidate parameter adjustment schemes ranked by Q value from high to low for pre-testing and verification. The digital twin virtual production model adopts a hybrid modeling method that combines mechanistic model and data-driven model. The data-driven model uses a long short-term memory network to calculate a comprehensive score based on product quality indicators, production efficiency indicators and energy consumption indicators, and selects the candidate parameter adjustment scheme with the highest comprehensive score as the recommended scheme.
[0172] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.
[0173] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.
Claims
1. A method for adaptive control of process parameters for production of a flotation frother, characterized by: include: A multi-source data acquisition and edge computing system was established. Temperature sensors, pressure sensors, flow sensors, raw material composition analyzers, and equipment status monitoring modules were deployed at the flotation frother production site. Edge computing nodes received multiple raw sensor signals at a sampling period of 100ms. Filtering, time alignment, standardization, and feature extraction were performed sequentially for each channel. The anomaly filtering module automatically identified and marked abnormal data points based on physical constraint rules and statistical criteria, and completed the removal of outliers. The cloud platform receives standardized data streams and multidimensional feature vectors to train and update deep Q-networks. The deep Q-networks contain 5 fully connected hidden layers with 256, 128, 64, 32 and 16 neurons in each layer, respectively. The output layer outputs the Q-value for each candidate action and takes the action with the maximum Q-value among all candidate actions as the current optimal decision. After sorting the candidate parameter adjustment schemes by Q value from high to low, the top N candidate parameter adjustment schemes with the highest Q values are selected and sent to the digital twin virtual production model for pre-run verification. The digital twin virtual production model adopts a hybrid modeling method that combines mechanistic model and data-driven model. The data-driven model uses a long short-term memory network to calculate a comprehensive score based on product quality indicators, production efficiency indicators and energy consumption indicators. The candidate parameter adjustment scheme with the highest comprehensive score is selected as the recommended scheme.
2. The adaptive control method for flotation frother production process parameters according to claim 1, characterized in that: The anomaly filtering module automatically identifies and marks abnormal data points based on physical constraint rules and statistical criteria, and completes the outlier removal process. The physical constraint rules are: temperature between 0 and 300 degrees Celsius, pressure between 0 and 2 MPa, and flow rate greater than or equal to 0. The statistical criteria are: the absolute value of the standardized value is greater than 3. The standardized value is the current filtered value of the same channel minus the mean of the historical normal operating condition data of the same channel, and then divided by the standard deviation of the historical normal operating condition data of the same channel.
3. The adaptive control method for flotation frother production process parameters according to claim 1, characterized in that: When the raw material composition analyzer detects that the mass fraction of key component A deviates from the target value, the mass fraction offset of key component A is normalized by the edge computing nodes and then used as an independent dimension of the state space vector to input the deep Q network. The vibration characteristic spectrum components output by the equipment condition monitoring module are extracted by the edge computing nodes and used as the equipment health status dimension to input the state space of the deep Q network. The raw material composition analyzer uses near-infrared spectroscopy to detect key chemical components in raw materials online. The equipment status monitoring module integrates vibration sensors, acceleration sensors, and acoustic emission sensors to collect operating status information of key rotating equipment.
4. The adaptive control method for flotation frother production process parameters according to claim 1, characterized in that: The digital twin virtual production model adopts a candidate parameter adjustment scheme screening strategy: First, the deep Q network outputs the Q value of each candidate action, sorts them from high to low according to the Q value, and selects only the top N candidate parameter adjustment schemes with the highest Q value to be sent to the digital twin virtual production model for pre-run verification. N is 5 to 10. The candidate parameter adjustment scheme screening strategy controls the total time of pre-run verification to within 50 to 100 seconds.
5. The adaptive control method for flotation frother production process parameters according to claim 1, characterized in that: The reward function calculates the reward value using a weighted sum method: in, As a reward value, For product quality awards, For productivity incentives, The energy consumption reward item; the product quality reward item includes key product indicators such as the content of active foaming agent and foam half-life, with a target value of 15% for the content of active foaming agent and a target value of 180 seconds for the foam half-life; the product quality reward item is equal to 1 minus the arithmetic mean of the relative deviations of all key product indicators.
6. The adaptive control method for flotation frother production process parameters according to claim 1, characterized in that: The automatic parameter drift compensation function continuously monitors the deviation between the actual value and the set value or model prediction value of each process parameter. When the absolute value of the deviation percentage exceeds 0.3%, the compensation mechanism is triggered. The compensation adjustment amount is the deviation value multiplied by the incremental adjustment coefficient, which is set to 0.
3. The compensation response time is less than 2 seconds and the overshoot is less than 5%.
7. The adaptive control method for flotation frother production process parameters according to claim 1, characterized in that: The intelligent identification of equipment faults adopts a multi-dimensional feature fusion method, which comprehensively processes the time-domain and frequency-domain features of vibration signals, temperature signals and current signals. The time-domain features and frequency-domain features together constitute a 10-dimensional feature vector. Multi-source features are normalized and then input into the fault classification model; the fault classification model adopts a hybrid architecture of support vector machine and deep neural network, and the support vector machine layer uses radial basis kernel function to output preliminary fault category labels; The 10-dimensional feature vector is concatenated with the initial fault category label to form an 11-dimensional vector, which is then input into a deep neural network layer. The deep neural network layer contains three fully connected hidden layers with 64, 32, and 16 neurons, respectively. The output layer uses a flexible maximum function to output the probability distribution of each fault type.
8. The adaptive control method for flotation frother production process parameters according to claim 1, characterized in that: There are two feedback loops between the pre-performance verification of the digital twin virtual production model and the deep Q network: The first loop is that during the training phase, the deep Q network uses the pre-performance verification results output by the digital twin virtual production model as the calibration basis for the product quality reward item in the reward function. When the product quality index of the candidate parameter adjustment scheme predicted by the digital twin virtual production model is lower than the effect threshold in the next 30 minutes, the reward value of the action corresponding to the candidate parameter adjustment scheme in the experience replay buffer is reduced. The second closed loop is triggered when the overall score of all candidate parameter adjustment schemes is lower than 0.8, and the simulation deviation data output by the digital twin virtual production model is fed back to the edge computing node as a supplementary input to the feature extraction module.
9. The adaptive control method for flotation frother production process parameters according to claim 1, characterized in that: Communication between edge computing nodes and the cloud platform is achieved asynchronously using message queue middleware. The cloud platform is configured with a distributed database to store historical data. The data storage adopts a partitioning and sharding strategy, and the data is partitioned according to the time range and device identifier. The storage period is set to be no less than 3 years. The production process traceability management module establishes a data recording system covering the entire process of raw material warehousing, production and processing, quality inspection and product delivery. The data records are anchored and stored using blockchain technology. The blockchain adopts a consortium blockchain architecture, and the hash value of the data records is written to the blockchain through smart contracts.
10. An adaptive control system for process parameters in the production of flotation frothers, applied to the method described in any one of claims 1-9, characterized in that: include: The multi-source data acquisition and edge computing system is used to deploy temperature sensors, pressure sensors, flow sensors, raw material composition analyzers and equipment status monitoring modules at the flotation frother production site. The edge computing nodes receive multiple raw sensor signals at a sampling period of 100ms, and perform filtering, time alignment, standardization and feature extraction on each channel in sequence. The anomaly filtering module automatically identifies and marks abnormal data points based on physical constraint rules and statistical criteria, and completes the removal of outliers. A deep reinforcement learning-driven process parameter optimization model is used to receive standardized data streams and multidimensional feature vectors on a cloud platform, train and update a deep Q network. The deep Q network contains 5 fully connected hidden layers with 256, 128, 64, 32 and 16 neurons in each layer, respectively. The output layer outputs the Q value for each candidate action, and the action with the maximum Q value among all candidate actions is taken as the current optimal decision. The digital twin virtual production model is used to receive the top N candidate parameter adjustment schemes ranked by Q value from high to low for pre-testing and verification. The digital twin virtual production model adopts a hybrid modeling method that combines mechanistic model and data-driven model. The data-driven model uses a long short-term memory network to calculate a comprehensive score based on product quality indicators, production efficiency indicators and energy consumption indicators, and selects the candidate parameter adjustment scheme with the highest comprehensive score as the recommended scheme.
Citation Information
Patent Citations
A data analysis-based automatic flotation reagent adding control system
CN116637729B
