Dedusting air distribution control system and method based on mathematical model

Abstract

The invention discloses a dust removal and air distribution control system and method based on a mathematical model, relates to the field of power failure prediction methods of power big data, and provides the following scheme that the dust removal and air distribution control system comprises a data acquisition unit, a mathematical model operation unit, a control decision unit, an execution unit and a feedback correction unit; the data acquisition unit is used for acquiring real-time environment parameters and equipment operation parameters in an industrial scene, and transmitting the acquired data to the mathematical model arithmetic unit through wireless transmission or an industrial Ethernet, and the transmission delay is less than or equal to 100ms. According to the invention, traditional control hysteresis is broken through, 'predictive 'accurate dust removal is realized, and the problem of scene adaptation is solved; meanwhile, self-adaptive regulation and control of multiple industrial scenes are achieved, the energy consumption optimization target can be fused, the win-win situation of efficient dust removal and low-consumption operation is achieved, long-term stability and reliability of the system are guaranteed by constructing a closed-loop optimization system, and the technology is high in expandability and adapts to future industrial intelligent upgrading.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial environment control, and in particular to a dust removal and air distribution control system and method based on a mathematical model. BACKGROUND

[0002] In industrial production processes, dust pollution is a key problem affecting the safety of the production environment and the health of workers. The dust removal and air distribution system, as the core equipment for controlling dust dispersion, directly determines the dust removal effect in terms of operational stability and control accuracy. Current mainstream dust removal and air distribution control methods mostly use traditional PID control or manual adjustment, which has the following technical defects:

[0003] 1. Significant control lag: Traditional PID control relies on preset parameters and the deviation feedback of actual dust concentration and air volume. When the dust concentration in the industrial scene suddenly changes (such as production equipment starting and stopping, material conveying volume changing), the control system cannot respond quickly, resulting in an expanded dust dispersion range and reduced dust removal efficiency.

[0004] 2. Poor parameter adaptability: Different industrial scenes (such as coal mine underground and steel plant converter workshop) have large differences in dust particle size distribution and air flow characteristics. Traditional control methods do not consider scene specificity and use fixed control logic, which can easily cause "over-air distribution" (waste of energy) or "under-air distribution" (incomplete dust removal) problems.

[0005] 3. Lack of dynamic prediction capability: Existing systems only passively adjust based on real-time monitoring data and cannot predict dust concentration trends based on production process changes (such as material feeding cycle and equipment operating load), resulting in air distribution strategies always lagging behind actual pollution conditions.

[0006] 4. Lack of model support: Existing technologies do not build specialized teaching models to guide system parameter setting and control logic optimization. When deploying new scenes, manual experience is required for repeated debugging, which not only consumes time and effort but also makes it difficult to ensure the consistency and stability of control effects.

[0007] To solve the above problems, there is an urgent need for a dust removal and air distribution control scheme that combines mathematical models (including teaching models) and has dynamic prediction and self-adaptive adjustment capabilities to improve control accuracy, reduce energy consumption, and adapt to multi-scene application requirements. SUMMARY

[0008] The present application relates to the technical field of industrial environment control, and in particular to a dust removal and air distribution control system and method based on a mathematical model.

[0009] To achieve the above purpose, the present application adopts the following technical solutions:

[0010] The dust removal air distribution control system based on a mathematical model comprises a data acquisition unit, a mathematical model operation unit, a control decision unit, an execution unit and a feedback correction unit.

[0011] The data acquisition unit is used to acquire real-time environmental parameters and equipment operation parameters in an industrial scene, and transmit the acquired data to the mathematical model operation unit through wireless transmission or industrial Ethernet, with a transmission delay of ≤100 ms.

[0012] The mathematical model operation unit serves as the core of the system and is used to perform teaching model operation processing on the data information acquired and transmitted by the data acquisition unit.

[0013] The control decision unit is used to receive the optimal air distribution parameters output by the mathematical model operation unit, generate control instructions in combination with real-time feedback data (provided by the feedback correction module), and the control instructions include fan speed adjustment steps (the smallest adjustment unit is 5 r / min) and air valve opening adjustment steps (the smallest adjustment unit is 1%), so as to avoid equipment impact caused by parameter mutation.

[0014] The execution unit includes a variable frequency fan controller and an electric air valve controller, receives the control instructions issued by the control decision unit, drives the fan and the air valve to perform corresponding actions, and the response time is ≤500 ms.

[0015] The feedback correction unit acquires dust concentration change data after the action of the execution unit in real time, compares the data with the predicted value of the mathematical model operation unit, calculates the change of the deviation value (deviation value = actual dust concentration - predicted dust concentration), and compares the data with the predicted value of the mathematical model operation unit.

[0016] If the absolute value of the deviation is >5%, the model correction mechanism is triggered, the weight parameters of the dust concentration prediction model are adjusted through an online gradient descent algorithm, and the deviation data is fed back to the teaching model. The teaching model records the scene parameters and model configuration corresponding to the deviation, which serves as the basis for subsequent parameter updating and scene classification optimization, and ensures the closed-loop optimization of the entire model system.

[0017] Further, the data acquisition unit comprises:

[0018] The environmental parameter acquisition module is used to acquire dust concentration (unit: mg / m³, acquired by a laser dust sensor, sampling frequency 1 time / s), air temperature (unit: ℃), air humidity (unit: %RH) and air flow rate (unit: m / s).

[0019] The equipment parameter acquisition module is used to acquire data information of fan speed (unit: r / min), air valve opening (unit: %) and dust removal filter bag pressure difference (unit: Pa) in real time.

[0020] Furthermore, the mathematical model operation unit specifically includes a teaching model, a dust concentration prediction model, and an air distribution parameter optimization model;

[0021] The teaching model is used to provide initial parameter configuration guidance, scenario adaptation rules, and abnormal data handling strategies for the dust concentration prediction model and the air distribution parameter optimization model. It is the foundation for ensuring the stable operation of the entire mathematical model system. The specific architecture, functions, training process, and stability assurance measures are as follows:

[0022] Model architecture: It adopts a layered structure, including a data preprocessing layer, a scene classification layer, a parameter configuration layer, and a rule output layer;

[0023] The data preprocessing layer cleans historical data (environmental parameters, equipment parameters, dust removal effect data) by removing outliers. The principles are as follows: data that deviates from the mean by more than 3 standard deviations, normalization (mapping the data to the [0,1] interval and using the min-max standardization formula: x_{norm}=\frac{x-x_{min}}{x_{max}-x_{min}}), and feature selection (selecting feature parameters with a correlation ≥0.6 with dust concentration using the Pearson correlation coefficient, such as temperature, humidity, and fan speed), to ensure the validity of the input data;

[0024] Core functions: When a new system is deployed, the teaching model receives basic parameters of the new scenario (dust type, equipment specifications, target dust removal effect), automatically matches typical scenarios, and outputs the initial configuration parameters of the dust concentration prediction model and the air distribution parameter optimization model, avoiding blind manual debugging; during system operation, the validity of input data is monitored in real time, and the data is preprocessed according to the abnormal data processing rules to provide reliable data support for subsequent model calculations;

[0025] Training process: Training the teaching model is the key to achieving accurate scenario classification and reasonable parameter configuration.

[0026] Furthermore, the scene classification layer is based on the K-means clustering algorithm, using dust particle size distribution (e.g., the proportion of 0-5μm, 5-10μm, and 10-20μm ranges), air flow speed (≤2m / s for low-speed scenes; 2-5m / s for medium-speed scenes; >5m / s for high-speed scenes), and equipment type (single fan system, multi-fan linkage system) as clustering features to divide industrial scenes into six typical scenarios: coal mine underground, metallurgical workshop, and building material factory. Each scenario corresponds to a set of standard feature vectors.

[0027] The core process of the K-means clustering algorithm is shown in the following formula:

[0028] 1) Objective function (minimize the sum of squared errors):

[0029] where is the number of clusters, denotes the k-th cluster, is the i-th data point, is the center (mean) of the k-th cluster, denotes the Euclidean distance;

[0030] 2) Cluster center update formula:

[0031] When the cluster partition is determined, the new center of each cluster is calculated:

[0032] where is the number of samples in the k-th cluster;

[0033] 3) Sample assignment rule:

[0034] Each sample is assigned to the cluster where the nearest cluster center is located:

[0035] The algorithm alternately performs step 2 (update cluster center) and step 3 (reassign samples) until the cluster center no longer changes significantly or the maximum number of iterations is reached, finally making the objective function J reach the minimum value;

[0036] The parameter configuration layer pre-stores the initial weight parameters of the dust concentration prediction model (such as the input layer and hidden layer connection weight initial value of the LSTM model, the number of hidden layer neurons) and the air distribution parameter optimization model constraint conditions (such as the energy consumption weight coefficient λ benchmark value in different scenes: coal mine underground type λ = 0.25, metallurgical workshop type λ = 0.2, building materials plant type λ = 0.18) verified by a large number of experiments for each typical scene, while storing the hyperparameters of model training (such as the learning rate initial value 0.01, iteration number 1000 times of the LSTM model, particle number 50, maximum iteration number 50 of the PSO algorithm);

[0037] The rule output layer outputs scene adaptation rules (when new scene data is input, calculate its cosine similarity with each typical scene standard feature vector, similarity ≥ 0.8, then directly match the corresponding scene parameters; similarity < 0.8, then trigger the parameter fine-tuning mechanism, perform linear interpolation calculation based on adjacent scene parameters) and abnormal data processing rules (when real-time collected data suddenly changes, such as dust concentration increase > 50% within 10 seconds, it is determined as abnormal data, trigger data completion strategy, use sliding window mean method to complete missing or abnormal data, window size is set to 5 seconds).

[0038] Further, the training process is specifically divided into five stages of training data set construction, data preprocessing layer training, scene classification layer training, parameter configuration layer training, and rule output layer training and model overall verification:

[0039] The training data set construction includes:

[0040] Data source: Collect historical operation data of six typical industrial scenes of coal mine, metallurgy, building materials, etc. Collect at least 1000 groups of valid data for each scene (each group of data contains 20 features such as dust particle size distribution, air flow rate, equipment type, model parameter configuration, dust removal effect, etc.), and the data time span covers four seasons (to ensure the diversity of parameters such as temperature, humidity, and dust characteristics), and the total data set size is ≥6000 groups;

[0041] Data annotation: Each group of data is annotated with scene type (such as "coal mine underground type - tunneling working face" and "metallurgical workshop type - converter area") and parameter effectiveness (annotated as "valid" or "invalid" according to whether the dust removal effect meets the standard, and the standard is that the dust removal efficiency is ≥85% and the energy consumption is ≤80% of the rated power of the equipment);

[0042] Data division: Divide the total data set into training set (4200 groups), validation set (1200 groups) and test set (600 groups) according to the ratio of 7:2:1, ensure that the proportion of each scene in each data set is consistent (1 / 6), and avoid the influence of data distribution deviation on the training effect;

[0043] The data preprocessing layer training includes:

[0044] Anomaly value processing rule training: Take the data annotated as "valid" in the training set as a sample, calculate the mean (μ) and standard deviation (σ) of each feature parameter, Determine anomaly value threshold (such as μ=25mg / m³, σ=8mg / m³ for dust concentration feature, then the anomaly value threshold is 25±24mg / m³, that is, data <1mg / m³ or >49mg / m³ is judged as abnormal), and adjust the threshold through the validation set test (if the misjudgment rate of abnormal values in the validation set is >5%, adjust the threshold to 3.5σ);

[0045] Feature selection model training: Calculate the correlation between each feature and dust removal effect (using dust removal efficiency as an indicator) in the training set using Pearson correlation coefficient algorithm, select features with correlation ≥0.6 (such as dust particle size 0-5μm proportion, fan speed, air flow rate), construct a feature subset, and compare the scene classification accuracy (using classification error rate as an indicator) of different feature subsets through the validation set. If the error rate is >8%, reduce the correlation threshold (reduce by 0.05 each time) and reselect, until the error rate is ≤8%, to determine the final feature subset.

[0046] Further, the scene classification layer training includes:

[0047] K value determination: the elbow method is used to determine the optimal K value of the K-means clustering algorithm, the within-cluster sum of squares (WCSS) is calculated when K=3 to K=8, when K increases from 5 to 6, the WCSS decreases by >20%, when K increases from 6 to 7, the WCSS decreases by <5%, and the optimal K=6 (corresponding to 6 typical scenes) is determined;

[0048] Clustering model training: input the preprocessed feature data of the training set into the K-means clustering model, initialize 6 cluster centers (randomly select 6 groups of effective data feature vectors), update the cluster centers through iterative calculation (iteration times are set to 200 times), until the cluster center change amplitude <0.01 (using the Euclidean distance of the feature vector as the index), complete the model training;

[0049] Clustering verification: input the validation set data into the trained clustering model, calculate the classification accuracy (the number of correctly classified samples / total number of samples in the validation set), if the accuracy <85%, re-initialize the cluster center and increase the iteration times (to 300 times) to retrain, until the accuracy ≥85%;

[0050] Standard feature vector determination: take the mean of the sample feature vectors in each cluster result to generate the standard feature vector of each typical scene (such as the standard feature vector of the coal mine underground type scene is [0.6, 0.4, 1.8, 0, 1,...], corresponding to the proportion of dust particle size 0-5 μm, 5-10 μm, air flow rate, multi-fan linkage identifier, single-fan identifier, etc.).

[0051] Further, the parameter configuration layer training includes:

[0052] Initial parameter screening: for each typical scene, select all data labeled as "effective" in the training set, extract the corresponding dust concentration prediction model parameters (such as the number of LSTM hidden layer neurons, initial weight) and the air distribution parameter optimization model constraint conditions (such as the energy consumption weight coefficient λ), and construct a parameter candidate set;

[0053] Optimal parameter determination: use cross-validation method (5-fold cross-validation) to evaluate the performance of each parameter combination in the candidate set, take the "prediction error + energy cost" as the comprehensive index (index value = prediction error x 0.6 + energy cost x 0.4, where energy cost = actual energy consumption / rated energy consumption), and select the parameter combination with the lowest comprehensive index as the initial parameter configuration of the scene (such as the LSTM hidden layer neuron number of the coal mine underground type scene = 64, λ = 0.25);

[0054] Parameter verification: Substitute the determined initial parameter configuration into the verification set data, calculate the model prediction error (≤6%) and dust removal efficiency (≥85%), if not meet the requirements, expand the parameter candidate set (increase 20% related data) and re-screen until the performance indicators are met;

[0055] The rule output layer training and model overall verification include:

[0056] Scene adaptation rule training: Take the scene matching data in the training set and the verification set as samples, calculate the cosine similarity between the new scene feature vector and the standard feature vector, and count the parameter configuration effect corresponding to different similarity intervals (such as similarity ≥ 0.8, parameter configuration effective rate ≥ 90%; 0.6≤similarity<0.8, effective rate 70%~90%), and determine the threshold of scene adaptation rule (similarity ≥ 0.8 direct matching, <0.8 fine-tuning) accordingly;

[0057] Abnormal data processing rule training: Inject simulated abnormal data (such as dust concentration mutation, sensor data missing) in the training set, test the completion error of different data completion strategies (sliding window mean method, linear interpolation method), and select the strategy (sliding window mean method, window size 5 seconds) with completion error <3% as the abnormal data processing rule;

[0058] Overall model verification: Input the test set data into the complete teaching model, evaluate the overall performance of the model, including scene classification accuracy (≥88%), parameter configuration effective rate (≥85%), and abnormal data processing accuracy (≥92%), if any index is not up to standard, then backtrack to the corresponding level and retrain until all indicators meet the requirements.

[0059] Further, the teaching model stability guarantee measures: In order to ensure the long-term stable operation of the teaching model and avoid model failure caused by scene change and data drift, the following measures are taken:

[0060] Regular data update: Set a monthly update cycle, the teaching model automatically collects the running data of each scene in the past month (including environmental parameters, equipment parameters, dust removal effect data, model correction records), updates the typical scene feature vector of the scene classification layer, uses incremental learning algorithm (such as incremental K-means algorithm), only calculates the clustering of new data, does not retrain the whole model, reduces the calculation amount, the change amplitude of each type of scene standard feature vector after updating is ≤10%, ensures the continuity of scene classification;

[0061] Parameter validity verification: The initial parameter configuration output by the teaching model is verified for validity every quarter. 10 groups of data under different operating conditions in each typical scene are selected. The initial parameters output by the teaching model are substituted into the dust concentration prediction model and the air distribution parameter optimization model to calculate the model prediction error and the actual dust removal effect. If the average value of the model prediction error in a certain scene is > 8% or the dust removal efficiency is lower than 80%, the parameter calibration process is triggered. The initial parameters corresponding to the scene are adjusted through comparative experiments (testing the model effect of different parameter combinations under the same working condition). The calibrated parameters must meet the requirements of prediction error ≤ 6% and dust removal efficiency ≥ 85% before they can be put into use.

[0062] Abnormal scene early warning: The teaching model monitors the matching degree (cosine similarity) of the new scene and the typical scene in real time. When the similarity is < 0.6 for 10 consecutive minutes, it is determined to be an abnormal scene. Early warning information (including scene abnormal feature data and current model running parameters) is sent to the system operation and maintenance terminal immediately. At the same time, the emergency parameter configuration (based on the mean value of all typical scene parameters) is automatically enabled to ensure temporary stable operation of the system until the operation and maintenance personnel intervene to adjust the scene classification rule.

[0063] Anti-interference design: A noise filtering module is added in the data preprocessing layer. Wavelet transform algorithm (db4 wavelet basis function is selected, and the decomposition layer is 3 layers) is used to filter common electromagnetic interference (such as high-frequency noise generated by frequency converter) and random error of sensor in industrial scene, so that the signal-to-noise ratio of the preprocessed data is ≥ 30 dB. At the same time, a logic checking mechanism is set in the rule output layer to check the rationality of the output parameter configuration and rule (such as the initial value of fan speed should not exceed 90% of the rated speed of the equipment, and the opening of the air valve should not be less than 10%). If unreasonable configuration is found, parameter correction is automatically triggered. After correction, recheck until it meets the logic requirements.

[0064] Further, the dust concentration prediction model is constructed based on long short-term memory network (LSTM). The dust concentration, temperature, humidity and equipment operating parameters within 1 hour of the historical data preprocessed by the teaching model are used as input features, and the predicted value of dust concentration in the future 5-15 minutes is output (prediction error ≤ 5%). During the model training process, the initial weight parameters and hyperparameters output by the teaching model are used to start the training, and then the gradient descent algorithm (learning rate is dynamically adjusted with iteration times, learning rate is 0.01 for the first 500 iterations, and then it is reduced by 0.001 every 100 iterations after 500 iterations, and the minimum is 0.001) is used to optimize the weight parameters to ensure that the prediction accuracy is dynamically improved with data accumulation.

[0065] The air distribution parameter optimization model takes the dust concentration prediction value, the target dust concentration (preset according to the industrial hygiene standard, such as ≤10 mg / m³), and the energy consumption constraint (fan power ≤50 kW) as inputs, solves the optimal air distribution parameter including the target fan speed and the target air valve opening through the particle swarm optimization algorithm (PSO), and the model objective function is: min (|predicted dust concentration-target dust concentration|) + λ × fan energy consumption (where λ is the energy consumption weight coefficient, the value range is 0.1-0.3, the initial value is output by the teaching model according to the scene type, and can be fine-tuned according to the actual operation effect), in the model solving process, the particle initial position and speed range are set by the teaching model based on the scene type (such as the particle initial speed range of the coal mine underground scene is 5-20 r / min (fan speed), 1%-5% (air valve opening), and the metallurgical workshop scene is 10-30 r / min, 2%-8%), to ensure the solving efficiency and accuracy.

[0066] The dust removal air distribution control method based on a mathematical model is applicable to the dust removal air distribution control system based on a mathematical model in any of the above aspects, and comprises the following steps:

[0067] S1, data initialization and model training:

[0068] When the system starts, the data acquisition module acquires the environmental parameters and equipment operation parameters in the last 30 days, and transmits them to the teaching model;

[0069] The teaching model pre-processes (cleanses, normalizes, and selects features) the historical data, completes scene classification through the K-means clustering algorithm, and generates standard feature vectors and corresponding model initial parameter configurations of various typical scenes;

[0070] The dust concentration prediction model adopts the initial weight parameters and hyperparameters output by the teaching model, takes the pre-processed historical data as the training set, and trains the initial dust concentration prediction model through the gradient descent algorithm;

[0071] The air distribution parameter optimization model initializes the objective function and PSO algorithm parameters (particle number, maximum iteration number, etc.) according to the scene constraint conditions (such as the energy consumption weight coefficient λ reference value) and the target dust concentration output by the teaching model;

[0072] S2, real-time data acquisition and pre-processing:

[0073] The data acquisition module acquires the current environmental parameters and equipment operation parameters in real time, and transmits them to the teaching model;

[0074] The teaching model cleans, normalizes, and filters noise from real-time data through a data preprocessing layer, removes abnormal data, fills in missing values, calculates the cosine similarity between the current scene and the standard feature vectors of various typical scenes, determines the scene matching type, and outputs the recommended model parameter adjustment (such as emergency parameter configuration if there is no matching scene);

[0075] S3, dust concentration prediction operation:

[0076] The dust concentration prediction model takes the preprocessed real-time data from the teaching model as input, combines the parameter adjustment suggestions (such as fine-tuning of partial weight parameters) output by the teaching model, and outputs the predicted dust concentration values for the next 5-15 minutes;

[0077] If the predicted value is greater than the target dust concentration, the air distribution parameter optimization process is triggered; if the predicted value is less than or equal to the target dust concentration, the current air distribution parameters are maintained, and the real-time data and prediction results are fed back to the teaching model for subsequent model updates;

[0078] S4, air distribution parameter optimization and control instruction generation:

[0079] The air distribution parameter optimization model takes the dust concentration prediction value, target dust concentration, and energy consumption constraint (energy consumption weight coefficient λ set by the teaching model) as input, and uses the PSO algorithm parameters configured by the teaching model to solve the optimal target fan speed and target air valve opening;

[0080] The control decision module generates stepwise control instructions based on the difference between the optimal air distribution parameters and the current device operating parameters, combined with the device protection rules (such as the fan speed single adjustment amplitude should not exceed 100 r / min) output by the teaching model (such as the current fan speed is 2000 r / min, and the target speed is 2200 r / min, then adjust in 4 steps, each step adjusts 50 r / min, and the interval is 10 seconds);

[0081] S5, execution and feedback correction:

[0082] The execution module receives the control instructions and drives the fan and air valve to adjust to the target parameters;

[0083] The feedback correction module collects the adjusted real-time dust concentration, compares it with the predicted value to calculate the deviation, and if the deviation is greater than 5%, it corrects the weight parameters of the dust concentration prediction model through an online gradient descent algorithm, and transmits the deviation data and current scene parameters to the teaching model;

[0084] The teaching model analyzes the deviation data, and if it determines that the parameter configuration is problematic (such as unreasonable initial weight parameters), it updates the initial parameters for the corresponding scene;

[0085] If it is determined that the scene changes (e.g., the dust particle size distribution changes significantly), the incremental update of the scene classification layer is triggered, and the scene classification rule is optimized;

[0086] Steps 3-5 are repeated until the deviation is less than or equal to 5%, and the system is stably operated.

[0087] The application realizes accurate scene classification and initial parameter optimization configuration of the teaching model, combines the LSTM dust concentration prediction model and the PSO air distribution parameter optimization model, so that the air distribution strategy is adapted to the dust concentration change in advance, the hysteresis is avoided, the dust removal efficiency is improved, the air distribution parameter optimization model takes "dust removal effect + energy consumption" as a double target function, the teaching model matches the optimal energy consumption weight coefficient for different scenes, the fan energy consumption is reduced under the premise of meeting the dust control requirements, the traditional "over-air distribution" problem is solved, the teaching model divides the typical scenes through K-means clustering, supports incremental update and abnormal scene early warning, combines the online model correction mechanism, the system can dynamically adjust the model parameters according to the dust characteristics and equipment parameters of different industrial scenes, without manual re-commissioning, and is suitable for coal mine, metallurgy, building materials and other multi-field applications; the teaching model guarantees its stable operation through regular data update, parameter validity verification and anti-interference design, and provides reliable parameter guidance for other models; the step-by-step control instruction avoids sudden changes of equipment parameters, reduces mechanical impact of the fan and the air valve, prolongs the service life of the equipment, and reduces maintenance cost; when a new scene is deployed, the teaching model can quickly output the initial parameter configuration of the model, without manual repeated commissioning, shortens the deployment time, and greatly improves the system landing efficiency;

[0088] To sum up, the application not only breaks through the traditional control hysteresis, realizes "predictive" accurate dust removal, solves the scene adaptation problem, but also realizes "self-adaptive" regulation and control in multiple industrial scenes, and can also integrate the energy consumption optimization target to realize "efficient dust removal + low consumption operation" win-win, through building a closed-loop optimization system, the system is guaranteed to be stable and reliable for a long time, and the application technology has strong scalability and is suitable for future industrial intelligent upgrading. BRIEF DESCRIPTION OF DRAWINGS

[0089] Figure 1 The unit structure block diagram of the dust removal air distribution control system based on a mathematical model proposed by the application;

[0090] Figure 2 The step block diagram of the dust removal air distribution control method based on a mathematical model proposed by the application;

[0091] Figure 3 The training data set construction block diagram of the dust removal air distribution control system based on a mathematical model proposed by the application. DETAILED DESCRIPTION

[0092] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be clearly and completely described below, obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work belong to the scope of protection of the present application.

[0093] In the description of the present application, it should be understood that the terms "center", "longitudinal", "transverse", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise" and the like indicate the orientation or positional relationship shown in the drawings, which are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation of the present application.

[0094] In addition, the terms "first", "second" are only for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise explicitly and specifically limited. In addition, the terms "mounting", "connecting", "connecting" should be broadly understood, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium, or it can be the communication between two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0095] Embodiment 1

[0096] Reference Figure 1 And Figure 3 : Dust removal air distribution control system based on mathematical model, comprising: a data acquisition unit, a mathematical model operation unit, a control decision unit, an execution unit and a feedback correction unit;

[0097] The data acquisition unit is used to collect real-time environmental parameters and equipment operation parameters in the industrial scene, and transmit the collected data to the mathematical model operation unit through wireless transmission or industrial Ethernet, and the transmission delay is ≤100ms;

[0098] The mathematical model operation unit as the core of the system is used to perform teaching model operation processing on the data information collected and transmitted by the data acquisition unit;

[0099] The control decision unit is used for receiving the optimal air distribution parameter output by the mathematical model operation unit, combining real-time feedback data (provided by the feedback correction module) to generate a control instruction, the control instruction including a fan speed adjustment step (the smallest adjustment unit 5r / min) and a damper opening adjustment step (the smallest adjustment unit 1%), so as to avoid equipment impact caused by parameter mutation;

[0100] The execution unit includes a variable frequency fan controller and an electric damper controller, receives the control instruction issued by the control decision unit, and drives the fan and the damper to perform corresponding actions, with a response time ≤ 500ms;

[0101] The feedback correction unit collects dust concentration change data after the action of the execution unit in real time, compares the data with the predicted value of the mathematical model operation unit, calculates the change of the deviation value (deviation value = actual dust concentration - predicted dust concentration), and if the absolute value of the deviation is greater than 5%, the model correction mechanism is triggered, the weight parameter of the dust concentration prediction model is adjusted through the online gradient descent algorithm, and the deviation data is fed back to the teaching model, the teaching model records the scene parameters and model configuration corresponding to the deviation as a basis for subsequent parameter updating and scene classification optimization, so as to ensure the closed-loop optimization of the whole model system.

[0102] In the present application, the data acquisition unit includes:

[0103] The environmental parameter acquisition module is used for collecting dust concentration (unit: mg / m³, collected by a laser dust sensor, sampling frequency 1 time / s), air temperature (unit: ℃), air humidity (unit: %RH) and air flow rate (unit: m / s);

[0104] The equipment parameter acquisition module is used for collecting data information of fan speed (unit: r / min), damper opening (unit: %) and dust removal filter bag pressure difference (unit: Pa) in real time.

[0105] In the present application, the mathematical model operation unit specifically includes a teaching model, a dust concentration prediction model and an air distribution parameter optimization model.

[0106] The teaching model is used for providing initial parameter configuration guidance, scene adaptation rules and abnormal data processing strategies for the dust concentration prediction model and the air distribution parameter optimization model, and is the basis for ensuring stable operation of the whole mathematical model system, and the specific architecture, functions, training process and stability guarantee measures are as follows:

[0107] The model architecture adopts a hierarchical structure, including a data preprocessing layer, a scene classification layer, a parameter configuration layer and a rule output layer.

[0108] The data preprocessing layer cleans the historical data (environmental parameters, equipment parameters and dust removal effect data) (eliminates abnormal values, adopts Principle, namely, remove data deviating from the mean by 3 standard deviations, normalization (map data to the interval [0, 1] using the min-max normalization formula: x norm = x-x min / x max-x min) and feature selection (select feature parameters such as temperature, humidity, fan speed, etc. with a Pearson correlation coefficient of >= 0.6 with dust concentration) to ensure the validity of the input data;

[0109] Core function: When deploying a new system, the teaching model automatically matches a typical scene by receiving basic parameters of the new scene (dust type, equipment specification, target dust removal effect), and outputs initial configuration parameters of the dust concentration prediction model and the air distribution parameter optimization model to avoid blind manual debugging; During system operation, the validity of the input data is monitored in real time, and the data is preprocessed according to the abnormal data processing rules to provide reliable data support for subsequent model operation;

[0110] Training process: The training of the teaching model is the key to realizing accurate scene classification and reasonable parameter configuration.

[0111] In the present application, the scene classification layer is based on the K-means clustering algorithm, and the dust particle size distribution (such as the proportion of 0-5 μm, 5-10 μm, 10-20 μm interval), air flow velocity (low speed scene for <= 2 m / s, medium speed scene for 2-5 m / s, high speed scene for > 5 m / s), and equipment type (single fan system, multiple fan linkage system) are used as clustering features to divide industrial scenes into 6 typical scenes, such as coal mine underground type, metallurgical workshop type, and building material plant type, each of which corresponds to a group of standard feature vectors.

[0112] The core process of the K-means clustering algorithm is shown in the following formula:

[0113] 1) Objective function (minimize error sum of squares):

[0114] wherein is the number of clusters, represents the kth cluster, is the ith data point, is the center (mean) of the kth cluster, represents the Euclidean distance;

[0115] 4) Cluster center update formula:

[0116] When the cluster division is determined, the new center of each cluster is calculated:

[0117] wherein is the number of samples in the kth cluster;

[0118] 5) Sample assignment rule:

[0119] Assign each sample to the cluster where the nearest cluster center is located:

[0120] The algorithm alternately performs step 2 (updating cluster centers) and step 3 (reassigning samples) until the cluster centers no longer change significantly or the maximum number of iterations is reached, finally minimizing the objective function J;

[0121] The parameter configuration layer pre-stores the initial weight parameters of the dust concentration prediction model verified by a large number of experiments (such as the input layer and hidden layer connection weight initial value of the LSTM model, the number of hidden layer neurons) and the air distribution parameter optimization model constraint conditions (such as the energy consumption weight coefficient λ reference value in different scenes: coal mine underground type λ=0.25, metallurgical workshop type λ=0.2, building material plant type λ=0.18), and stores the hyperparameters of model training (such as the learning rate initial value 0.01, the number of iterations 1000 times of the LSTM model, the number of particles 50, the maximum number of iterations 50 of the PSO algorithm);

[0122] The rule output layer outputs the scene adaptation rule (when new scene data is input, calculate its cosine similarity with each typical scene standard feature vector, similarity ≥0.8, then directly match the corresponding scene parameters; similarity <0.8, trigger parameter fine-tuning mechanism, linear interpolation calculation based on adjacent scene parameters) and abnormal data processing rule (when real-time collected data suddenly changes, such as dust concentration increase >50% within 10 seconds, it is determined as abnormal data, trigger data completion strategy, use sliding window mean method to complete missing or abnormal data, window size is set to 5 seconds).

[0123] In the present application, the training process is specifically divided into five stages: training data set construction, data preprocessing layer training, scene classification layer training, parameter configuration layer training and rule output layer training, and model overall verification:

[0124] The training data set construction includes:

[0125] Data source: collect historical operation data of six typical industrial scenes of coal mine, metallurgy and building materials, collect at least 1000 groups of effective data for each scene (each group of data contains 20 features such as dust particle size distribution, air flow rate, equipment type, model parameter configuration and dust removal effect), the data time span covers four seasons (to ensure the diversity of temperature, humidity, dust characteristics and other parameters), and the total data set size is ≥6000 groups;

[0126] Data labeling: Scene type labeling (e.g., labeling "coal mine underground type-tunneling working face" and "metallurgical workshop type-converter area") and parameter validity labeling (labeling parameter configuration as "valid" or "invalid" according to whether the dust removal effect meets the standard, and the standard is that the dust removal efficiency is greater than or equal to 85% and the energy consumption is less than or equal to 80% of the rated power of the equipment) are performed on each group of data.

[0127] Data division: The total data set is divided into a training set (4200 groups), a validation set (1200 groups), and a test set (600 groups) in a ratio of 7:2:1, ensuring that the proportion of each type of scene in each data set is consistent (1 / 6), and avoiding the influence of data distribution deviation on the training effect.

[0128] The data preprocessing layer training includes:

[0129] Outlier processing rule training: Taking the data labeled as "valid" in the training set as a sample, the mean (mu) and standard deviation (sigma) of each feature parameter are calculated, and the outlier rejection threshold (for example, the mu of the dust concentration feature is 25 mg / m3, and the sigma is 8 mg / m3, so the outlier threshold is 25+24 mg / m3, that is, data less than 1 mg / m3 or greater than 49 mg / m3 is determined as an outlier), and the threshold is adjusted through the validation set test (if the misjudgment rate of outliers in the validation set is greater than 5%, the threshold is adjusted to 3.5 sigma);

[0130] Feature selection model training: The Pearson correlation coefficient algorithm is used to calculate the correlation between each feature in the training set and the dust removal effect (using dust removal efficiency as an indicator), and features with a correlation greater than or equal to 0.6 (such as dust particle size 0-5 mu m proportion, fan speed, and air flow rate) are selected. A feature subset is constructed, and the scene classification accuracy of different feature subsets is compared through the validation set (using the error rate as an indicator). If the error rate is greater than 8%, the correlation threshold is reduced (by 0.05 each time) and reselected until the error rate is less than or equal to 8%, and the final feature subset is determined.

[0131] In the present application, the scene classification layer training includes:

[0132] K value determination: The elbow method (Elbow Method) is used to determine the optimal K value of the K-means clustering algorithm, and the within-cluster sum of squares (WCSS) is calculated when K=3 to K=8. When K increases from 5 to 6, the WCSS decreases by more than 20%, and when K increases from 6 to 7, the WCSS decreases by less than 5%. Therefore, the optimal K=6 (corresponding to 6 typical scenes) is determined.

[0133] Cluster model training: input the preprocessed feature data of the training set into the K-means clustering model, initialize 6 cluster centers (randomly select 6 groups of feature vectors of valid data), update the cluster centers through iterative calculation (the number of iterations is set to 200 times), until the change amplitude of the cluster center is <0.01 (taking the Euclidean distance of the feature vector as the index), complete the model training;

[0134] Cluster verification: input the verification set data into the trained clustering model, calculate the classification accuracy (the number of correctly classified samples / total number of samples in the verification set), if the accuracy is <85%, re-initialize the cluster center and increase the number of iterations (to 300 times) to retrain until the accuracy is ≥85%;

[0135] Standard feature vector determination: take the mean of the sample feature vectors in each cluster result to generate the standard feature vector of each typical scene (such as the standard feature vector of the coal mine underground type scene is [0.6,0.4,1.8,0,1,...], corresponding to the proportion of dust particle size 0-5um, 5-10um, air flow rate, multi-fan linkage identifier, single-fan identifier, etc.).

[0136] In the present application, the parameter configuration layer training includes:

[0137] Initial parameter screening: for each typical scene, select all data labeled as "valid" in the training set, extract the corresponding dust concentration prediction model parameters (such as the number of LSTM hidden layer neurons, initial weight) and air distribution parameter optimization model constraints (such as energy consumption weight coefficient λ), and construct a parameter candidate set;

[0138] Optimal parameter determination: use cross-validation method (5-fold cross-validation) to evaluate the performance of each parameter combination in the candidate set, take "prediction error + energy cost" as the comprehensive index (index value = prediction error x 0.6 + energy cost x 0.4, where energy cost = actual energy consumption / rated energy consumption), select the parameter combination with the lowest comprehensive index as the initial parameter configuration of the scene (such as the number of LSTM hidden layer neurons of the coal mine underground type scene = 64, λ = 0.25);

[0139] Parameter verification: input the determined initial parameter configuration into the verification set data, calculate the model prediction error (≤6%) and dust removal efficiency (≥85%), if it does not meet the requirements, expand the parameter candidate set (increase 20% related data) and re-screen until the performance indicators are met;

[0140] Rule output layer training and model overall verification include:

[0141] Scene adaptation rule training: Take the scene matching data in the training set and the validation set as samples, calculate the cosine similarity between the new scene feature vector and the standard feature vector, and count the parameter configuration effect corresponding to different similarity intervals (such as when similarity is greater than or equal to 0.8, the parameter configuration effective rate is greater than or equal to 90%; when 0.6 is less than 0.8, the effective rate is 70%-90%), and the threshold of scene adaptation rule is determined (similarity is greater than or equal to 0.8, and less than 0.8, and the threshold is adjusted);

[0142] Abnormal data processing rule training: Inject simulated abnormal data (such as dust concentration mutation, sensor data missing) in the training set, test the completion error of different data completion strategies (sliding window mean method, linear interpolation method), and select the strategy (sliding window mean method, window size 5 seconds) with completion error less than 3% as the abnormal data processing rule;

[0143] Overall model verification: input the test set data into the complete teaching model, evaluate the overall performance of the model, including scene classification accuracy (greater than or equal to 88%), parameter configuration effective rate (greater than or equal to 85%), and abnormal data processing accuracy (greater than or equal to 92%), if any index does not meet the requirements, go back to the corresponding level and retrain until all indexes meet the requirements.

[0144] In the present application, the stability guarantee measures of the teaching model: in order to ensure the long-term stable operation of the teaching model, and avoid the model failure caused by scene change and data drift, the following measures are taken:

[0145] Regular data update: set a monthly update cycle, the teaching model automatically collects the operation data (including environmental parameters, equipment parameters, dust removal effect data, model correction records) of each scene in the past month, updates the typical scene feature vector of the scene classification layer, uses the incremental learning algorithm (such as incremental K-means algorithm), only calculates the clustering of the new data, does not retrain the whole model, reduces the calculation amount, and the change amplitude of the standard feature vector of each type of scene after updating is less than or equal to 10%, which ensures the continuity of scene classification;

[0146] Parameter validity verification: the initial parameter configuration output by the teaching model is verified for validity every quarter, 10 groups of data under different operating conditions in each typical scene are selected, the initial parameters output by the teaching model are substituted into the dust concentration prediction model and the air distribution parameter optimization model, and the model prediction error and the actual dust removal effect are calculated;

[0147] If the average value of the model prediction error in a certain type of scene is greater than 8% or the dust removal efficiency is less than 80%, the parameter calibration process is triggered, the initial parameters corresponding to the scene are adjusted through comparative experiments (test the model effect of different parameter combinations under the same working condition), and the calibrated parameters meet the requirements of prediction error less than or equal to 6% and dust removal efficiency greater than or equal to 85% before they can be put into use;

[0148] Abnormal scene early warning: The teaching model monitors the matching degree (cosine similarity) of the new scene and the typical scene in real time. When the similarity is <0.6 for 10 consecutive minutes, it is determined that the scene is abnormal, and early warning information (including scene abnormal feature data, current model running parameters) is sent to the system operation and maintenance terminal. At the same time, the emergency parameter configuration (based on the mean value of all typical scene parameters) is automatically enabled to ensure temporary stable operation of the system until the operation and maintenance personnel intervene to adjust the scene classification rule.

[0149] Anti-interference design: A noise filtering module is added in the data preprocessing layer to filter common electromagnetic interference (such as high-frequency noise generated by frequency converters) and sensor random errors in industrial scenes using wavelet transform algorithm (db4 wavelet basis function is selected, and 3 layers of decomposition are used). The signal-to-noise ratio of the preprocessed data is ≥30dB. At the same time, a logic verification mechanism is set in the rule output layer to verify the rationality of the output parameter configuration and rule (such as the initial value of the fan speed should not exceed 90% of the rated speed of the equipment, and the opening of the air valve should not be less than 10%). If unreasonable configuration is found, parameter correction is automatically triggered, and after correction, re-verification is performed until the logic requirement is met.

[0150] In the present application, the dust concentration prediction model is constructed based on long short-term memory network (LSTM), and the historical dust concentration, temperature, humidity, and equipment running parameters within 1 hour after preprocessing by the teaching model are used as input features. The predicted value of the dust concentration in the future 5-15 minutes (prediction error ≤5%) is output. During the model training process, the initial weight parameters and hyperparameters output by the teaching model are used to start the training, and then the weight parameters are optimized by gradient descent algorithm (learning rate is dynamically adjusted with iteration times, learning rate is 0.01 for the first 500 iterations, and then it is reduced by 0.001 every 100 iterations, and the minimum is 0.001), to ensure that the prediction accuracy is dynamically improved with data accumulation.

[0151] The air distribution parameter optimization model takes the dust concentration prediction value, the target dust concentration (preset according to the industrial hygiene standard, such as ≤10mg / m³), and the energy consumption constraint (fan power ≤50kW) as input, and solves the optimal air distribution parameters including the target fan speed and the target air valve opening through particle swarm optimization algorithm (PSO). The objective function of the model is: min(|predicted dust concentration-target dust concentration|)+λ×fan energy consumption (where λ is the energy consumption weight coefficient, the value range is 0.1-0.3, the initial value is output by the teaching model according to the scene type, and it can be fine-tuned according to the actual running effect). During the model solving process, the initial position and speed range of the particle are set by the teaching model based on the scene type (such as the initial speed range of the particle in the coal mine underground scene is 5-20r / min (fan speed), 1%-5% (air valve opening), and the initial speed range of the particle in the metallurgical workshop scene is 10-30r / min, 2%-8%). This ensures the solving efficiency and accuracy.

[0152] Embodiment 2

[0153] With reference to Figure 2 : A dust removal air distribution control method based on a mathematical model, applicable to a dust removal air distribution control system of a mathematical model, comprising the following steps:

[0154] S1, data initialization and model training:

[0155] When the system is started, the data acquisition module acquires environmental parameters and equipment operation parameters in the past 30 days, and transmits them to the teaching model;

[0156] The teaching model pre-processes (cleanses, normalizes, and selects features) the historical data, completes scene classification through a K-means clustering algorithm, and generates standard feature vectors and corresponding model initial parameter configurations of each typical scene;

[0157] The dust concentration prediction model uses the initial weight parameters and hyperparameters output by the teaching model, takes the pre-processed historical data as a training set, and trains an initial dust concentration prediction model through a gradient descent algorithm;

[0158] The air distribution parameter optimization model initializes the objective function and PSO algorithm parameters (particle number, maximum iteration number, etc.) according to the scene constraint conditions (such as the energy consumption weight coefficient λ reference value) and the target dust concentration output by the teaching model;

[0159] S2, real-time data acquisition and preprocessing:

[0160] The data acquisition module acquires real-time environmental parameters and equipment operation parameters, and transmits them to the teaching model;

[0161] The teaching model cleanses, normalizes, and filters noise of the real-time data through the data preprocessing layer, eliminates abnormal data and fills in missing values, calculates the cosine similarity between the current scene and the standard feature vectors of each typical scene, determines the scene matching type, and outputs the adaptive model parameter adjustment suggestion (such as the emergency parameter configuration if there is no matching scene);

[0162] S3, dust concentration prediction operation:

[0163] The dust concentration prediction model takes the real-time data pre-processed by the teaching model as input, combines the parameter adjustment suggestion (such as fine-tuning part of the weight parameters) output by the teaching model, and outputs the dust concentration prediction value in the next 5-15 minutes;

[0164] If the prediction value is greater than the target dust concentration, the air distribution parameter optimization process is triggered; if the prediction value is less than or equal to the target dust concentration, the current air distribution parameters are maintained, and the real-time data and the prediction result are fed back to the teaching model for subsequent model update;

[0165] S4, air distribution parameter optimization and control instruction generation:

[0166] The air distribution parameter optimization model takes the dust concentration prediction value, the target dust concentration, and the energy consumption constraint (the energy consumption weight coefficient λ set by the teaching model) as inputs, and solves the optimal target fan speed and target air valve opening using the PSO algorithm parameters configured by the teaching model;

[0167] The control decision module generates a stepwise control instruction (such as current fan speed 2000 r / min, target speed 2200 r / min, then adjust in 4 steps, each step 50 r / min, interval 10 seconds) according to the difference between the optimal air distribution parameters and the current equipment operating parameters, combined with the equipment protection rules (such as the single adjustment amplitude of the fan speed should not exceed 100 r / min) output by the teaching model;

[0168] S5, execution and feedback correction:

[0169] The execution module receives the control instruction and drives the fan and air valve to adjust to the target parameters;

[0170] The feedback correction module collects the real-time dust concentration after adjustment, compares it with the prediction value to calculate the deviation, and if the deviation is > 5%, corrects the weight parameters of the dust concentration prediction model through online gradient descent algorithm, and transmits the deviation data and the current scene parameters to the teaching model;

[0171] The teaching model analyzes the deviation data, and if it is determined to be a parameter configuration problem (such as unreasonable initial weight parameters), the initial parameters of the corresponding scene are updated;

[0172] If it is determined to be a scene change (such as a significant change in dust particle size distribution), trigger the incremental update of the scene classification layer to optimize the scene classification rules;

[0173] Repeat steps 3-5 until the deviation ≤ 5%, the system is stable and runs.

[0174] The above is only the preferred specific implementation of the present application, but the protection scope of the present application is not limited to this, any skilled person in the art can make equivalent replacement or change according to the technical solution and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A dust removal air distribution control system based on a mathematical model, characterized in that, Comprise: Data acquisition unit, mathematical model operation unit, control decision unit, execution unit and feedback correction unit; The data acquisition unit is used for collecting real-time environmental parameters and equipment operation parameters in the industrial scene, and transmitting the collected data to the mathematical model operation unit through wireless transmission or industrial Ethernet, with a transmission delay ≤100ms; The mathematical model operation unit as the core of the system is used for teaching model operation processing on the data information collected and transmitted by the data acquisition unit; The control decision unit is used for receiving the optimal air distribution parameters output by the mathematical model operation unit, generating control instructions in combination with real-time feedback data, and the control instructions include fan speed adjustment step and air valve opening adjustment step to avoid equipment impact caused by parameter mutation; The execution unit includes a variable frequency fan controller and an electric air valve controller, which receives the control instructions issued by the control decision unit, drives the fan and the air valve to perform corresponding actions, and the response time is ≤500ms; The feedback correction unit collects dust concentration change data after the action of the execution unit in real time, compares it with the predicted value of the mathematical model operation unit, calculates the change of the deviation value, and ensures the closed-loop optimization of the whole model system.

2. The model-based dust control system of claim 1, wherein, The data acquisition unit comprises: An environmental parameter acquisition module for collecting dust concentration, air temperature, air humidity and air flow rate; An equipment parameter acquisition module for collecting data information of fan speed, air valve opening and dust removal filter bag pressure difference in real time.

3. The model-based dust control system of claim 1, wherein, The mathematical model operation unit specifically comprises a teaching model, a dust concentration prediction model and an air distribution parameter optimization model; The teaching model is used to provide initial parameter configuration guidance, scene adaptation rules and abnormal data processing strategies for the dust concentration prediction model and the air distribution parameter optimization model, and is the basis for ensuring the stable operation of the whole mathematical model system. The specific architecture, function, training process and stability guarantee measures are as follows: Model architecture: adopts a hierarchical structure, including data preprocessing layer, scene classification layer, parameter configuration layer and rule output layer; The data preprocessing layer cleans, normalizes and selects features for historical data to ensure the effectiveness of the input data; Core function: when the new system is deployed, the teaching model receives the basic parameters of the new scene, automatically matches the typical scene, and outputs the initial configuration parameters of the dust concentration prediction model and the air distribution parameter optimization model to avoid blind manual debugging; during system operation, the effectiveness of the input data is monitored in real time, and the data is preprocessed according to the abnormal data processing rules to provide reliable data support for subsequent model operation; Training process: the training of the teaching model is the key to realizing accurate scene classification and reasonable parameter configuration.

4. The model-based dust control system of claim 3, wherein, The scene classification layer is based on K-means clustering algorithm, taking dust particle size distribution, air flow rate and equipment type as clustering features, and divides the industrial scene into 6 typical scenes of coal mine underground type, metallurgical workshop type and building material plant type, each of which corresponds to a group of standard feature vectors; The parameter configuration layer pre-stores the initial weight parameters of the dust concentration prediction model and the constraint conditions of the air distribution parameter optimization model verified by a large number of experiments for each typical scene, and also stores the hyperparameters of model training; The rule output layer outputs scene adaptation rules and abnormal data processing rules.

5. The model-based dust control system of claim 3, wherein, The training process is specifically divided into five stages of training data set construction, data preprocessing layer training, scene classification layer training, parameter configuration layer training, rule output layer training and model overall verification: The training data set construction includes: Data sources: collect historical operation data of six typical industrial scenes of coal mines, metallurgy, building materials, etc. Collect at least 1000 sets of valid data for each type of scene, and the data time span covers four seasons, with a total data set size of ≥6000 sets; Data labeling: label each set of data with scene type and parameter validity; Data division: divide the total data set into training set, validation set and test set according to the ratio of 7:2:1, ensure that the proportion of each scene in each data set is consistent, and avoid data distribution deviation affecting the training effect; The data preprocessing layer training includes: Abnormal value processing rule training: taking the data labeled as "valid" in the training set as samples, calculating the mean and standard deviation of each feature parameter, determining the abnormal value elimination threshold, and adjusting the threshold through the test of the validation set; Feature selection model training: use Pearson correlation coefficient algorithm to calculate the correlation between each feature in the training set and the dust removal effect, select features with a correlation of ≥0.6, construct a feature subset, and compare the scene classification accuracy of different feature subsets through the validation set. If the error rate is >8%, reduce the correlation threshold and reselect, until the error rate is ≤8%, and determine the final feature subset.

6. The model-based dust control system of claim 5, wherein, The scene classification layer training includes: K value determination: determine the optimal K value of K-means clustering algorithm using the elbow rule. Calculate the within-cluster sum of squares when K=3 to K=8. When K increases from 5 to 6, the WCSS decreases by >20%. When K increases from 6 to 7, the WCSS decreases by <5%. The optimal K is determined to be 6. Cluster model training: input the preprocessed feature data of the training set into the K-means clustering model, initialize 6 cluster centers, update the cluster centers through iterative calculation until the cluster center variation is <0.01, and complete model training. Cluster verification: input the validation set data into the trained clustering model, calculate the classification accuracy, and if the accuracy is <85%, reinitialize the cluster centers and increase the number of iterations to retrain until the accuracy is ≥85%. Standard feature vector determination: take the mean of the sample feature vectors in each cluster result to generate the standard feature vector of each typical scene.

7. The model-based dust control system of claim 5, wherein, The parameter configuration layer training includes: Initial parameter selection: for each typical scene, select all data labeled as "valid" in the training set, extract the corresponding dust concentration prediction model parameters and air distribution parameter optimization model constraints, and construct a parameter candidate set; Optimal parameter determination: use cross-validation method to evaluate the performance of each parameter combination in the candidate set, take the "prediction error + energy cost" as the comprehensive index, and select the parameter combination with the lowest comprehensive index as the initial parameter configuration of the scene; Parameter verification: input the determined initial parameter configuration into the validation set data, calculate the model prediction error and dust removal efficiency, and if it does not meet the requirements, expand the parameter candidate set and reselect until the performance indicators are met; The rule output layer training and model overall verification includes: Scene adaptation rule training: Take the matching data in the training set and the validation set as samples, calculate the cosine similarity between the new scene feature vector and the standard feature vector, and count the parameter configuration effect corresponding to different similarity intervals. According to this, the threshold of the scene adaptation rule is determined; Abnormal data processing rule training: Inject simulated abnormal data into the training set, test the completion error of different data completion strategies, and select the strategy with a completion error of less than 3% as the abnormal data processing rule; Overall model verification: Input the test set data into the complete teaching model to evaluate the overall performance of the model, including scene classification accuracy, parameter configuration efficiency, and abnormal data processing accuracy. If any index does not meet the requirements, go back to the corresponding level and retrain until all indexes meet the requirements.

8. The model-based dust control system of claim 5, wherein, The teaching model stability guarantee measures: To ensure the long-term stable operation of the teaching model and avoid model failure due to scene changes and data drift, the following measures are taken: Regular data update: Set a monthly update cycle, and the teaching model automatically collects the operation data of each scene in the past month. The typical scene feature vectors of the scene classification layer are updated. Incremental learning algorithm is used to only cluster the new data, without retraining the entire model, reducing the amount of calculation. The change amplitude of the standard feature vectors of each type of scene after updating is ≤10%, ensuring the continuity of scene classification; Parameter effectiveness verification: The initial parameter configuration output by the teaching model is verified for effectiveness every quarter. Select 10 groups of data under different operating conditions in each typical scene. The initial parameters output by the teaching model are substituted into the dust concentration prediction model and the air distribution parameter optimization model to calculate the model prediction error and the actual dust removal effect. If the average value of the model prediction error in a certain scene is >8% or the dust removal efficiency is less than 80%, the parameter calibration process is triggered. The initial parameters corresponding to the scene are adjusted through comparative experiments. The calibrated parameters must meet the prediction error ≤6% and the dust removal efficiency ≥85% before they can be put into use; Abnormal scene early warning: The teaching model monitors the matching degree of new scenes and typical scenes in real time. When the similarity is <0.6 for 10 consecutive minutes, it is determined to be an abnormal scene. An early warning message is sent to the system operation terminal immediately, and emergency parameter configuration is automatically enabled to ensure temporary stable operation of the system until the operation personnel intervene to adjust the scene classification rule; Anti-interference design: Add a noise filtering module in the data preprocessing layer to filter common electromagnetic interference and sensor random errors in industrial scenes using wavelet transform algorithm, so that the signal-to-noise ratio of the preprocessed data is ≥30dB. At the same time, a logic verification mechanism is set in the rule output layer to verify the rationality of the output parameter configuration and rule. If unreasonable configuration is found, parameter correction is automatically triggered. After correction, re-verification is performed until the logic requirements are met.

9. The model-based dust control system of claim 3, wherein, The dust concentration prediction model is constructed based on a long short-term memory network, and historical dust concentration, temperature, humidity and equipment operation parameters within 1 hour after preprocessing by the teaching model are taken as input features, and a prediction value of the dust concentration in the future 5-15 minutes is output. In the model training process, the initial weight parameters and hyperparameters output by the teaching model are used to start the training, and then the weight parameters are optimized by the gradient descent algorithm to ensure that the prediction accuracy is dynamically improved with data accumulation. The air distribution parameter optimization model takes the dust concentration prediction value, target dust concentration and energy consumption constraint as input, and solves the optimal air distribution parameters including target fan speed and target air valve opening through the particle swarm optimization algorithm. In the model solving process, the initial position and speed range of particles are set by the teaching model based on the scene type to ensure the solving efficiency and accuracy.

10. A method for control of dust extraction air distribution based on mathematical model, applicable to a system for control of dust extraction air distribution based on mathematical model according to any of the claims 1-9, characterized in that, The method comprises the following steps: S1, data initialization and model training: When the system starts, the data acquisition module acquires environmental parameters and equipment operation parameters within the last 30 days and transmits them to the teaching model. The teaching model preprocesses the historical data, completes scene classification through the K-means clustering algorithm, generates standard feature vectors of various typical scenes and corresponding initial model parameter configurations. The dust concentration prediction model uses the initial weight parameters and hyperparameters output by the teaching model, takes the preprocessed historical data as a training set, and trains an initial dust concentration prediction model through the gradient descent algorithm. The air distribution parameter optimization model initializes the objective function and PSO algorithm parameters according to the scene constraint conditions and target dust concentration output by the teaching model. S2, real-time data acquisition and preprocessing: The data acquisition module acquires real-time environmental parameters and equipment operation parameters and transmits them to the teaching model. The teaching model cleans, normalizes and filters noise of the real-time data through the data preprocessing layer, eliminates abnormal data and fills in missing values, calculates the cosine similarity of the current scene and standard feature vectors of various typical scenes, determines the scene matching type, and outputs the adaptive model parameter adjustment suggestion. S3, dust concentration prediction operation: The dust concentration prediction model takes the real-time data preprocessed by the teaching model as input, combines the parameter adjustment suggestion output by the teaching model, and outputs a prediction value of the dust concentration in the future 5-15 minutes. If the prediction value is greater than the target dust concentration, the air distribution parameter optimization process is triggered. If the prediction value is less than or equal to the target dust concentration, the current air distribution parameters are maintained, and the real-time data and prediction results are fed back to the teaching model for subsequent model updating. S4, air distribution parameter optimization and control instruction generation: The air distribution parameter optimization model takes the dust concentration prediction value, target dust concentration and energy consumption constraint as input, and solves the optimal target fan speed and target air valve opening using the PSO algorithm parameters configured by the teaching model. The control decision module generates a step control instruction according to the difference between the optimal air distribution parameters and the current equipment operation parameters, and combines the equipment protection rules output by the teaching model. S5, execution and feedback correction: The execution module receives the control instruction, drives the fan and air valve to adjust to the target parameters. The feedback correction module collects the adjusted real-time dust concentration, compares it with the predicted value to calculate the deviation, and if the deviation > 5%, corrects the weight parameters of the dust concentration prediction model through the online gradient descent algorithm, and transmits the deviation data and the current scene parameters to the teaching model; The teaching model analyzes the deviation data, and if it is determined that the parameter configuration problem, the initial parameters of the corresponding scene are updated; If it is determined that the scene changes, the incremental update of the scene classification layer is triggered to optimize the scene classification rules; Repeat steps 3-5 until the deviation ≤ 5%, and the system runs stably.

Images