Pulse injection adaptive optimization control method and system and computer readable storage medium

Abstract

The invention relates to the technical field of blowing dust removal control of a bag type dust collector, in particular to a pulse blowing self-adaptive optimization control method and system and a computer readable storage medium. The pulse-jet adaptive optimization control method comprises the following steps: collecting running state data of a bag type dust collector in real time, defining a jet parameter combination as the action of an intelligent agent, calculating an energy efficiency reward value according to the pressure difference change before and after jet and the compressed air consumption, and calculating the energy efficiency of the bag type dust collector based on the state data and the reward value. And carrying out online learning and optimization by adopting a reinforcement learning algorithm, and dynamically outputting an optimal injection parameter combination to execute ash removal. According to the method, a control target is upgraded from pressure difference maintenance to maximum ash removal energy efficiency ratio, through online self-learning, on the premise that the ash removal effect is guaranteed, compressed air consumption is remarkably reduced, and the method has the beneficial effects that energy is saved, the adaptability is high, and the service life of the filter bag is prolonged.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bag type dust collector injection and ash cleaning control, and particularly to a pulse injection self-adaptive optimization control method and system and a computer readable storage medium. BACKGROUND

[0002] A bag type dust collector is the most widely used high-efficiency dust removal equipment in the industrial field, and its core component is a filter bag. When dust-containing gas passes through the filter bag, dust is trapped on the outer surface of the filter bag to form a dust layer. As the dust layer thickens, the operating resistance of the dust collector (usually measured by the pressure difference between the inside and outside of the filter bag) will continuously increase. When the resistance reaches a certain level, the filter bag must be cleaned to remove the dust layer and restore the air permeability of the filter bag.

[0003] Pulse injection and ash cleaning is the most mainstream ash cleaning method at present. It sprays high-pressure compressed air into the filter bag through the injection pipe and the Venturi tube, causing the filter bag to expand and vibrate sharply, thereby shaking off the attached dust. The control strategy of the pulse injection system directly determines the operating stability, energy consumption and filter bag life of the dust collector. At present, the pulse injection control scheme mainly has the following two kinds: 1. Timing control scheme: This scheme triggers each pulse valve in turn for injection and ash cleaning according to a pre-set fixed time period. For example, it is set to spray every 30 minutes to all filter bag units in turn. The pressure, pulse width (i.e. solenoid valve on-time) and other parameters used for injection are also pre-set fixed values manually. However, the timing control does not take into account the actual dust load. When the dust concentration is low, the system still sprays on time, causing a lot of compressed air waste. When the dust concentration suddenly increases, it cannot respond in time, resulting in a high pressure difference, affecting the dust removal efficiency, and even damaging the filter bag.

[0004] 2. Fixed pressure difference control scheme: This scheme installs a pressure difference sensor at the inlet and outlet of the dust collector to monitor the pressure difference on both sides of the filter bag in real time. When the pressure difference rises to a pre-set upper limit value (such as 1500 Pa), the controller starts the injection program to clean the filter bag until the pressure difference drops to a pre-set lower limit value (such as 1000 Pa). The injection pressure and pulse width and other parameters in this scheme are also pre-set fixed values manually. Although the fixed pressure difference control achieves on-demand cleaning, its control logic is still rough.

[0005] The parameters such as injection pressure and width in the above two schemes are fixed, and for dusts of different properties (such as viscosity, humidity, and particle size), the optimal dust cleaning parameters are quite different, and the fixed parameters cannot achieve the best energy efficiency ratio under all working conditions. In addition, the existing scheme lacks automatic learning and adaptive optimization capability, cannot dynamically adjust to different dust properties, humidity, and temperature, and the control target is only to maintain the pressure difference within a certain range, without considering the consumption of compressed air as a control target for optimization. SUMMARY

[0006] The technical problem to be solved by the present application is to solve the problems of high operating cost and extensive control strategy in the prior art in the background art, and to provide a pulse injection adaptive optimization control method.

[0007] The technical scheme adopted by the present application to solve the technical problem is: a pulse injection adaptive optimization control method, comprising the following steps: S1, collecting real-time operation state data of the bag-type dust collector system, the operation state data at least including the pressure difference value of the filter bag on both sides; S2, taking the injection parameter combination as the action of the intelligent agent, wherein the injection parameters at least include the injection pressure and the pulse width; S3, calculating the energy efficiency reward value of the action according to the pressure difference change before and after injection and the compressed air consumption of this injection, wherein the energy efficiency reward value is proportional to the pressure difference change and inversely proportional to the compressed air consumption; S4, based on the operation state data and the energy efficiency reward value, using a reinforcement learning algorithm for online learning and optimization, dynamically adjusting and outputting the optimal injection parameter combination; S5, generating a control instruction according to the optimal injection parameter combination to drive the pulse injection actuator to perform the dust cleaning operation; S6, when the strategy of the reinforcement learning algorithm converges, switching to a steady-state control mode, performing the injection action based on the historical optimal parameters, and triggering the exploration mechanism regularly to cope with the working condition changes.

[0008] S1 collects real-time state data to provide a data basis for subsequent intelligent decision-making, S2 defines the injection parameters as actions, and considers the parameters as actions that can be dynamically adjusted, providing conditions for refinement and adaptation; S3 calculates the energy efficiency reward value, combining the pressure difference change and the compressed air consumption to form an evaluation index; S4 reinforcement learning online optimization, through reinforcement learning, constantly interacting with the environment, and automatically searching for the optimal injection parameter combination that can be obtained under different working conditions; S5 executes the optimal instruction, and S6 avoids the energy loss caused by continuous exploration.

[0009] Further, the operation state data in S1 further includes the pressure difference change rate at least one of the differential pressure, the gas tank pressure, the inlet dust concentration, and the ambient humidity.

[0010] The differential pressure change rate, the inlet dust concentration, and the ambient humidity are increased, so that the system perceives the working conditions more finely and comprehensively, and the judgment ability of the system in the face of complex and variable working conditions is enhanced.

[0011] Further, the injection parameters in S2 further include an injection interval , and the injection parameters are discretized or continuously valued within a preset safety range.

[0012] The injection interval is included in the action space that can be optimized, and the intelligent control of the dust removal rhythm is further realized, so that the dust removal strategy is more complete and efficient.

[0013] Further, the calculation of the energy efficiency reward value of the current action in S3 is as follows: S31, before performing the injection action , record the current differential pressure value ; S32, perform the injection action , and record the differential pressure value at a short time after injection , and calculate the dust removal effect E: ; S33, calculate the compressed air consumption C of the current injection: , wherein k is a correlation coefficient of the nozzle and the valve, is the injection pressure at time t, is the pulse width at time t; S34, calculate the energy efficiency reward value of the current action: .

[0014] The ratio of the dust removal effect (i.e. the differential pressure change before and after injection) to the compressed air consumption is taken as the evaluation standard, which provides an optimization direction for the learning algorithm.

[0015] Further, the calculation of the energy efficiency reward value further includes introducing a penalty term and / or a steady-state reward term, the penalty term is triggered when the differential pressure exceeds the upper limit of safety, and the steady-state reward term is triggered when the differential pressure reaches or is lower than the target differential pressure.

[0016] The introduction of the penalty term and / or the steady-state reward term can guide the system to actively avoid risks and ensure stable operation.

[0017] Further, the reinforcement learning algorithm in S4 is Q-Learning or deep Q network algorithm, and the steps of online learning and optimization include: S41. Create a Q-table and initialize all Q values ​​with small random numbers or zero; S42. Perceive the current state Based on the ε-greedy strategy, actions are selected from the current action space. ; S43, Execution Action To obtain a new state Calculate the instant reward value obtained from this action. ; S44. Update the Q value according to the Bellman equation. The specific calculation formula is as follows: , In the formula, Here is the updated Q-value, α is the learning rate, and γ is the discount factor. This is the discount value for the maximum future reward. S45, New State Transfer to current state Repeat the above steps to gradually converge to the optimal strategy.

[0018] By employing an exploration-utilization strategy for decision-making and by iteratively updating the Q-table through the Bellman equation, the system ensures its self-learning and adaptive capabilities.

[0019] Furthermore, S4 employs an evolutionary algorithm for online learning and optimization. The evolutionary algorithm can be either a genetic algorithm or a particle swarm optimization algorithm. The energy efficiency reward value is used as the fitness function to iteratively evolve the optimal combination of jetting parameters.

[0020] In this invention, the algorithm used for online learning and optimization can be broadly categorized as any algorithm that interacts with the environment and iteratively optimizes based on feedback signals (such as the energy efficiency reward value in this invention). Typical examples include reinforcement learning algorithms (such as Q-Learning), but also evolutionary algorithms (such as genetic algorithms) with similar effects.

[0021] A control system for the pulse jet adaptive optimization control method described in the above scheme is also provided, comprising: Sensor array, used to collect real-time operating status data of bag filter system; The controller is communicatively connected to the sensor group and executes the steps of the pulse jet adaptive optimization control method based on dual optimization of energy consumption and performance as described in any one of claims 1 to 7. The actuator group is communicatively connected to the controller and is used to receive the control commands and perform the blowing action.

[0022] Furthermore, the controller includes: The data acquisition and preprocessing module is used to process the data from the sensor group; The dynamic energy efficiency model module is used to calculate the energy efficiency bonus value; An adaptive optimization algorithm module is used to run the reinforcement learning algorithm; A jetting command generation module is used to generate the control commands; The actuator group includes at least a pulse valve, a solenoid valve, and a pressure regulating valve.

[0023] A computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the pulse jet adaptive optimization control method as described in the above scheme.

[0024] The beneficial effects of this invention are: This invention upgrades the control objective from simply maintaining the pressure difference to maximizing the dust removal energy efficiency ratio. Through online optimization, the system can always find the blowing parameters that consume the least amount of compressed air and achieve the desired dust removal effect under the current operating conditions, thus avoiding the huge energy waste caused by fixed parameters in the traditional method. This invention employs an online learning mechanism that can automatically adapt to changes in working conditions such as dust concentration, particle size, viscosity, and humidity. When the production process is adjusted or the raw materials are changed, there is no need to manually reset the parameters. The system can learn on its own and obtain the optimal control strategy, which greatly improves the robustness and intelligence of the system. By optimizing the process, this invention avoids the need for prolonged use of excessively high pressure for violent dust removal, thus reducing mechanical impact and wear on the filter bags. At the same time, precise control also prevents irreversible damage to the filter bags caused by high pressure differentials due to untimely dust removal, thereby extending the filter bag replacement cycle and reducing maintenance costs. Intelligent dust removal control makes the operating pressure difference of the dust collector more stable, maintaining it within the economic resistance range, which is conducive to the stable operation of the front-end main fan, reduces the fluctuation of fan energy consumption, and ensures the process stability of the entire production line. Attached Figure Description

[0025] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0026] Figure 1 This is a structural block diagram of the pulse jet adaptive optimization control system of the present invention.

[0027] Figure 2 This is a flowchart of the pulse jet adaptive optimization control method of the present invention.

[0028] In the diagram: 100, bag filter body; 200, sensor group; 201, differential pressure sensor; 202, air tank pressure sensor; 300, controller; 301, data acquisition and preprocessing module; 302, dynamic energy efficiency model module; 303, adaptive optimization algorithm module; 304, pulse jet command generation module; 400, actuator group; 401, pulse valve solenoid valve; 402, pressure regulating valve. Detailed Implementation

[0029] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0030] Example 1 like Figure 1 As shown, a control system for an adaptive optimization control method of pulse jet cleaning includes a bag filter body 100, a sensor group 200, a controller 300, and an actuator group 400. The bag filter body 100 includes filter bags, pulse valves, and an air tank, etc. The sensor group 200 is used to collect the operating status data of the bag filter system in real time, and includes a differential pressure sensor 201 and an air tank pressure sensor 202. The differential pressure sensor 201 is used to detect the pressure difference across the filter bag. The gas tank pressure sensor 202 is used to detect the blowing pressure P; the controller 300 is connected to the sensor group 200 and is a PLC, industrial computer or embedded controller; the actuator group 400 is connected to the controller 300 and is used to receive control commands and execute the blowing action, which includes a pulse valve solenoid valve 401 and a pressure regulating valve 402.

[0031] Specifically, the controller 300 includes a data acquisition and preprocessing module 301, a dynamic energy efficiency model module 302, an adaptive optimization algorithm module 303, and a spray command generation module 304. The data acquisition and preprocessing module 301 is used to process the data from the sensor group 200 and acquire the pressure difference ΔP and the spray pressure P in real time. The dynamic energy efficiency model module 302 is used to calculate the energy efficiency bonus value and evaluate the energy efficiency of each spray. The adaptive optimization algorithm module 303 is used to run the reinforcement learning algorithm to generate the optimal spray action. The spray command generation module 304 is used to generate control commands to control the actuator group 400 to perform corresponding actions.

[0032] The data acquisition and preprocessing module 301 not only collects data but also provides high-quality, consistent input for subsequent learning algorithms, avoiding noisy data from misleading the learning process. The dynamic energy efficiency model module 302 constructs energy efficiency reward values ​​for dust removal effect (pressure reduction) and energy consumption cost (gas consumption), unifying the originally contradictory performance and energy consumption goals into a single maximizable goal. The adaptive optimization algorithm module 303 models the pulse jet control problem as a sequence decision problem. Through reinforcement learning, the system no longer relies on preset, fixed rules but autonomously discovers the optimal control strategy hidden in complex working conditions through trial and error and experience accumulation with the environment. The jet instruction generation module 304 is responsible for converting intelligent decisions into specific and safe execution instructions to ensure the execution of the algorithm's decisions.

[0033] Example 2 like Figure 2 As shown, an adaptive optimization control method for pulse jet blowing includes the following steps: S1. Real-time acquisition of operating status data of the bag filter system, including at least the pressure difference value on both sides of the filter bag; S2. The action of the agent is based on the combination of jetting parameters, wherein the jetting parameters include at least jetting pressure and pulse width; S3. Calculate the energy efficiency bonus value of this action based on the pressure difference change before and after the injection and the compressed air consumption of this injection. The energy efficiency bonus value is directly proportional to the pressure difference change and inversely proportional to the compressed air consumption. S4. Based on the operating status data and energy efficiency bonus value, a reinforcement learning algorithm is used for online learning and optimization, dynamically adjusting and outputting the optimal combination of injection parameters; S5. Generate control commands based on the optimal combination of pulse jet parameters to drive the pulse jet actuator to perform dust removal operations; S6. After the reinforcement learning algorithm's policy converges, switch to steady-state control mode, execute the blowing action based on the historical optimal parameters, and periodically trigger the exploration mechanism to cope with changes in working conditions.

[0034] After the strategy converges, using the historically optimal parameters can avoid unnecessary performance fluctuations and computational resource consumption caused by continuous exploration, allowing the system to run at a stable and efficient operating point. The periodically triggered exploration mechanism is equivalent to setting up a long-term self-checking and fine-tuning function for the system, which can detect unstable operating conditions caused by slow factors such as filter bag aging and seasonal changes. Once it is detected that the original strategy is no longer optimal, the learning process can be reactivated by changes in the reward signal, ensuring that the system maintains optimal performance throughout its entire life cycle.

[0035] Specifically, the operating status data in S1 also includes the differential pressure change rate. At least one of the following: gas tank pressure, inlet dust concentration, and ambient humidity. The injection parameters in S2 also include the injection interval. Furthermore, the injection parameters are discretely or continuously set within a preset safety range. The energy efficiency bonus value for this action in S3. The calculation is as follows: S31, During the execution of the blowing action Record the current differential pressure value. ; S32, Perform the blowing action Record the pressure difference value shortly after the injection. Calculate the dust removal effect E: ; S33. Calculate the compressed air consumption C for this injection: Where k is the correlation coefficient between the nozzle and the valve. It is the blowing pressure at time t. It is the pulse width at time t; S34. Calculate the energy efficiency bonus value for this action. : .

[0036] Reward function for energy efficiency ratio The document explicitly guides reinforcement learning agents to find strategies that achieve the greatest possible pressure reduction using the least amount of compressed air. When operating conditions change, such as increased dust viscosity, the original blowing parameters may not achieve the same dust removal effect E, or may require more air consumption C. This will directly affect the energy efficiency bonus value. Decrease, Energy Efficiency Bonus Value The changes automatically drive the learning algorithm to find new parameter combinations to restore high rewards, achieving adaptive optimization without any human intervention.

[0037] In S4, the reinforcement learning algorithm is either Q-Learning or a deep Q-network algorithm. The online learning and optimization steps include: S41. Create a Q-table and initialize all Q values ​​with small random numbers or zero, i.e., Q(S, A), where S represents all possible states and A represents all possible actions; S42, Controller 300 senses the current state through data acquisition and preprocessing module 301. such as the current pressure difference and pressure difference change rate Based on the ε-greedy strategy (ε being the exploration rate), a random number rand between 0 and 1 is first generated. If rand < ε, then exploration is performed, that is, an action is randomly selected from the current action space. If rand ≥ ε, then it is utilized, that is, the current state is looked up in the Q table. For the corresponding row, select the action with the largest Q value. ; S43, The jetting instruction generation module 304 executes the action according to the selected action. The actuator group 400 completes one spray cycle, and acquires a new status after completion. Calculate the instant reward value obtained from this action. ; S44. Update the Q value according to the Bellman equation. The specific calculation formula is as follows: , In the formula, Here is the updated Q-value, α is the learning rate, and γ is the discount factor. This is the discount value for the maximum future reward. S45, New State Transfer to current state Repeat the above steps to gradually converge to the optimal strategy.

[0038] By balancing the utilization of known optimal actions with the exploration of unknown possibilities, the system ensures broad search capabilities in the early stages of learning, avoiding getting trapped in local optima, while also ensuring stable utilization of the optimal strategy in the later stages of learning. The update rule based on the Bellman equation essentially approximates the long-term true value of the action corresponding to each state through iteration. The convergence process of the Q(S,A) value is the system's learning process from the control environment. Once the Q-table converges, for any state... Choose to make Q( A) The largest action A is the globally optimal solution under the current working condition; by changing the new state... Continuous feedback to the system forms a closed-loop learning system, enabling the learning process to continue and dynamically track changes in the system state, thereby achieving online self-adaptation.

[0039] As the system continues to operate, the values ​​in the Q-table become increasingly accurate, and the exploration rate ε gradually decreases over time. The system transitions from being primarily exploration-based to primarily utilization-based, meaning it executes more known optimal actions. When operating conditions change (such as changes in dust properties), causing the old optimal strategy to become no longer optimal, the reward value R changes. The Q-learning algorithm automatically relearns and adapts through continuous iteration, finding new optimal combinations of injection parameters, thereby achieving dynamic adaptive adjustment of the control strategy.

[0040] The implementation process of this method will be explained in detail below with reference to a set of preferred initial parameters.

[0041] The state S of the system is defined as { , }, that is, the current pressure difference and its rate of change, pressure difference The monitoring range is set to 0–1800 Pa, and the differential pressure change rate is... It is obtained by calculating the difference between the current pressure differential and the pressure differential of the previous cycle (e.g., after a 10-second interval); Action A of the agent is the combination of jetting parameters {P, The spray interval will not be optimized for the time being. The injection pressure P is between 0.3 MPa and 0.7 MPa, discrete in 0.1 MPa increments, with five selectable values: {0.3, 0.4, 0.5, 0.6, 0.7} MPa; pulse width... Between 80ms and 200ms, with a step size of 40ms, the options are discretized into four values: {80, 120, 160, 200}ms. There are a total of 5 × 4 = 20 possible combinations of actions. The compressed air consumption calculation coefficient k is set to 0.15 (L / (MPa·ms)) based on the on-site valve and nozzle models; to prevent excessive pressure differential, a safe upper limit for pressure differential is set. =1500 Pa, when > At that time, apply a large negative penalty of -50 to the reward value; Let the learning rate α = 0.1, the discount factor γ = 0.9, and the initial exploration rate be... =0.3, and set it to decay over time. For example, after every 100 spray decisions, the exploration rate is multiplied by a decay coefficient of 0.95, with a minimum of 0.05; the Q table is initialized as a matrix of state number × 20, and all Q values ​​are initially 0; Assume the data acquisition module obtains the current system status as follows: = (1250 Pa, +15 Pa / s), indicating that the current pressure difference is 1250 Pa and is on an upward trend; Controller 300 generates a random number 0.4. Since 0.4 > the current exploration rate ε (e.g., 0.25), it enters the utilization mode and queries the status in the Q table. The corresponding line detects the action. = (P=0.5MPa, The Q value is highest at 120ms, therefore this action is selected for this jetting action; The injection command generation module 304 controls the pressure regulating valve 402 to adjust the pressure to 0.5 MPa and triggers the pulse valve to open for 120 ms. After the injection is completed, wait for the system to stabilize (e.g., wait 3 seconds) and then measure the pressure difference again to obtain the result. =980 Pa; Calculate the dust removal effect E=1250-980=270Pa, compressed air consumption C=0.15×0.5×120=9L, basic energy efficiency bonus value. =270 / 9=30; due to < If there is no penalty, then the total reward for this action is... =30; The system has entered a new state. = (980 Pa, -30 Pa / s), assuming that in the Q-table, the maximum Q value for all actions in the new state is (980 Pa, -30 Pa / s), =35, then update the Q value according to the Bellman equation: Q(S_t, A_t)=0+0.1×[30+0.9×35-0]=0.1×(30+31.5)=6.15, Then, under the condition of (1250Pa, +15Pa / s), the Q value of the action (0.5MPa, 120ms) will be updated from 0 to 6.15.

[0042] Example 3 The difference from Example 2 is that the calculation of the energy efficiency bonus value introduces a penalty term and / or a steady-state bonus term. The penalty term is triggered when the pressure difference exceeds the safety limit, and the steady-state bonus term is triggered when the pressure difference reaches or falls below the target pressure difference.

[0043] The penalty is triggered when the pressure difference exceeds the safety limit: After cleaning, if the pressure difference across the filter bag still exceeds the system's preset safe operating limit (e.g., 1500 Pa), a fixed penalty value (e.g., 100 points) with a large absolute value is subtracted from the total reward value. This forces the learning agent to learn to avoid jet cleaning strategies that may cause excessive system resistance, thus prioritizing the safe operation of the equipment and preventing damage to the filter bag or impact on the main process due to inadequate cleaning.

[0044] The steady-state reward is triggered when the pressure difference reaches or falls below the target pressure difference: After cleaning is completed, if the pressure difference across the filter bag reaches or falls below an ideal target pressure difference value (e.g., 1000 Pa), once triggered, a fixed positive reward score (e.g., 10 points) is added to the total reward value. This positively guides and encourages the agent, prompting the system to actively seek out the blowing parameters that enable the system to maintain ideal operating conditions with low resistance and low energy consumption, thereby ensuring effectiveness while further achieving energy saving, consumption reduction, and extending filter bag life.

[0045] Example 4 The difference from Example 2 is that S4 uses an evolutionary algorithm for online learning and optimization. Specifically, the evolutionary algorithm uses a genetic algorithm, which can treat a set of injection parameters {P, Tw, Ti} as an individual and use the energy efficiency ratio R=E / C as the fitness function. By performing selection, crossover, mutation and other operations on multiple sets of parameters, the individual with the highest fitness, i.e. the optimal parameter combination, is evolved iteratively.

[0046] Specifically, the process is as follows: First, within the safe range allowed by the blowing pressure, pulse width, and blowing interval, multiple different combinations of blowing parameters are randomly generated, for example, 20 combinations. This initial set of parameters is called the initial population, and each set of parameters is called an individual. Each set of parameters (i.e., each individual) in the population is applied sequentially to the actual blowing and cleaning operation. The energy efficiency bonus value defined in Example 2 is used as the evaluation standard to calculate the merits of each set of parameters. This value is called fitness in this algorithm. The higher the fitness, the better the overall performance of the set of parameters in terms of cleaning effect and energy consumption.

[0047] Secondly, based on the calculated fitness, a survival-of-the-fittest selection is carried out. Parameter sets with high fitness are more likely to be retained and participate in the generation of new parameter sets as parents. Parent parameter sets are randomly paired and crossover operations are performed to exchange and combine some values ​​of the two sets of parameters, thereby generating a brand new offspring parameter set. Furthermore, for the parameter group generated by newly, a small, random adjustment is made to one of the parameters (such as the injection pressure) with a small probability (e.g., 10%). Finally, the above steps are repeated to form a continuous evolutionary cycle. After multiple generations (e.g., 100 generations), the overall fitness of the entire population will continuously improve. Ultimately, the set of spray parameters with the highest fitness from all generations is selected as the optimal solution found by the genetic algorithm.

[0048] Example 5 The difference from Example 2 is that S4 uses an evolutionary algorithm for online learning and optimization. Specifically, the evolutionary algorithm is a particle swarm optimization algorithm, which can also be used to find the optimal combination of parameters. Each particle represents a potential solution, and the solution space is searched by following the individual optimum and the global optimum.

[0049] Specifically, within the safe range of the jetting parameters, a large number of candidate jetting parameter combinations are randomly generated, such as 30 groups. Each parameter combination is called a particle. All these particles constitute a particle swarm. Each particle has its own position (i.e., the current parameter value) and flight speed (i.e., the trend and step size of parameter changes). Secondly, the parameter combination represented by each particle is applied to the jet cleaning, and its energy efficiency bonus value is calculated as the fitness of that particle to evaluate the quality of the position; each particle will remember the best historical position it found during the entire flight (i.e., the best set of parameters it has ever found); at the same time, the entire particle swarm will share information and record the best performing position among all particles, i.e., the global best position. Furthermore, in each iteration, each particle adjusts its flight speed (i.e., the next search direction) based on two types of information: moving closer to its own historical best position; or moving closer to the global best position of the entire particle swarm. At the same time, the particle also retains some of its previous flight inertia. Then, the particle moves to a new position based on the updated speed, thus generating a new set of jet parameters. During the update process, if the new parameters calculated by a particle exceed the safe operating range, the system will force it to be corrected to the nearest boundary value to ensure the feasibility of the parameters. Finally, repeating the above steps, the particle swarm gradually gathers towards the optimal parameter region under the guidance of the historical best and the global best. When the preset number of iterations is reached or the performance of the solution no longer improves significantly, the algorithm ends and outputs the currently found global best position, that is, the optimal combination of jetting parameters.

[0050] Example 6 A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the pulse jet adaptive optimization control method of any one of Embodiments 2 to 5.

[0051] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A pulse-jet adaptive optimization control method, characterized by, Comprising the following steps: S1, collecting the running state data of the bag-type dust collector system in real time, the running state data at least including the differential pressure value of the filter bag on both sides; S2, taking the injection parameter combination as the action of the intelligent agent, wherein the injection parameter at least includes the injection pressure and the pulse width; S3, calculating the energy efficiency reward value of the action according to the differential pressure change before and after injection and the compressed air consumption of this injection, wherein the energy efficiency reward value is proportional to the differential pressure change and inversely proportional to the compressed air consumption; S4, based on the running state data and the energy efficiency reward value, using the reinforcement learning algorithm for online learning and optimization, dynamically adjusting and outputting the optimal injection parameter combination; S5, generating the control instruction according to the optimal injection parameter combination to drive the pulse injection actuator to perform the dust removal operation; S6, when the strategy of the reinforcement learning algorithm converges, switching to the steady-state control mode, performing the injection action based on the historical optimal parameters, and triggering the exploration mechanism regularly to cope with the working condition changes.

2. The pulse-jet adaptive seeking control method of claim 1, wherein: The operating status data in S1 also includes the differential pressure change rate. At least one of the following: gas tank pressure, inlet dust concentration, and ambient humidity.

3. The pulse-jet adaptive seeking control method of claim 1, wherein: The injection parameters in S2 also include injection interval , and the injection parameters are discretized or continuously valued within a preset safety range.

4. The pulse-jet adaptive seeking control method of claim 1, wherein: S3 the energy efficiency reward value of this action the calculation of which is specifically: S31, in performing the blowing action Before, record the current differential pressure value ; S32, execute the blowing action At a short time after the blowing, record the differential pressure value at this time Calculate the dust cleaning effect E: ; S33, calculate the compressed air consumption C of this time of injection: wherein k is a correlation coefficient of the nozzle and the valve, is the injection pressure at time t, is the pulse width at time t; S34, calculate the energy efficiency reward value of this action : .

5. The pulse-jet adaptive seeking control method of claim 4, wherein: The calculation of the energy efficiency reward value further includes introducing a penalty term and / or a steady-state reward term, the penalty term being triggered when the differential pressure exceeds the upper limit of safety, and the steady-state reward term being triggered when the differential pressure reaches or is lower than the target differential pressure.

6. The pulse-jet adaptive seeking control method of claim 1, wherein: The reinforcement learning algorithm in S4 is Q-Learning or deep Q network algorithm, and the steps of online learning and optimization include: S41, creating a Q table, initializing all Q values with small random numbers or zeros; S42, sensing a current state selecting an action from the current action space based on an e-greedy policy ; S43, performing an action , obtaining a new state , calculating an immediate reward value obtained by the action ; S44, updating the Q value according to the Bellman equation, and the specific calculation formula is: ， In the formula, Qnewis the updated Q value, a is the learning rate, and g is the discount factor, is the discount value of the maximum future reward value; S45, the new state transition to the current state repeat the above steps to gradually converge to the optimal policy.

7. The pulse-jet adaptive seeking control method of claim 1, wherein: In S4, the evolutionary algorithm is used for online learning and optimization, the evolutionary algorithm is genetic algorithm or particle swarm optimization algorithm, the energy efficiency reward value is used as the fitness function, and the optimal injection parameter combination is evolved iteratively.

8. A control system implementing the pulse-jet adaptive optimization control method according to any one of claims 1 to 7, characterized in that, Comprise: A sensor group (200) for collecting the running state data of the bag-type dust collector system in real time; A controller (300) in communication connection with the sensor group (200) and executing the steps of the pulse injection adaptive optimization control method based on the energy consumption and performance double optimization according to any one of claims 1-7; An actuator group (400) in communication connection with the controller (300) and used for receiving the control instruction and performing the injection action.

9. The pulse-jet adaptive seeking control system of claim 8, wherein, The controller (300) comprises: A data acquisition and preprocessing module (301) for processing the data of the sensor group (200); A dynamic energy efficiency model module (302) for calculating the energy efficiency reward value; An adaptive optimization algorithm module (303) for running the reinforcement learning algorithm; An injection instruction generation module (304) for generating the control instruction; The actuator group (400) at least includes a pulse valve electromagnetic valve (401) and a pressure regulating valve (402).

10. A computer readable storage medium having a computer program stored thereon, when the computer program is executed by a processor, realizing the pulse injection adaptive optimization control method according to any one of claims 1-7.

Images