Pulverized coal dust collection linkage start-stop method based on neural network prediction

Abstract

The present application relates to the technical field of warehouse dust removal control, and discloses a method for starting and stopping a valley ploughing and dust collecting linkage based on neural network prediction, comprising: controlling the valley ploughing cutting unit to perform a pre-detection action, extracting a high-frequency harmonic component of the cross-axis current and inputting the high-frequency harmonic component combined with static physical parameters into a neural network model to determine a target operating frequency; calculating the overall aerodynamic delay time in combination with the current dynamic aerodynamic impedance coefficient, controlling the dust collecting fan and the valley ploughing cutting unit to start and stop in linkage according to a timing window; obtaining the fan operating sequence within the timing window to perform dynamic decoupling, and updating the aerodynamic impedance coefficient. The method realizes feedforward adjustment of the operating frequency through multi-source parameter prediction, avoids pipe network blockage; the aerodynamic work is decoupled by using the equipment start and stop timing, the pipe network impedance is updated under the condition of no external fluid sensor; the delay time is calculated based on the resistance, and the effective connection between the dust removal wind field and the material cutting action is ensured.

Description

Technical Field

[0001] This invention relates to the field of warehouse dust control technology, specifically a method for the coordinated start-up and shutdown of grain harvesting and dust collection based on neural network prediction. Background Technology

[0002] In the storage of bulk materials, grain unloaders are commonly used for unloading and turning over materials. Their cutting operations generate dust, thus requiring a dust collection system. In existing technology, grain unloaders and dust collectors typically use a basic linked start-stop control system. That is, when a work command is received, the dust collector and the grain unloading cutting components are started sequentially at preset fixed time intervals, and both are stopped sequentially at the end of the work according to a set time difference to control the dust concentration in the working environment.

[0003] The dust collection operating parameters of this control logic are mostly based on design values ​​and operate statically, lacking a feedforward adjustment mechanism that combines the static physical properties of the material with the real-time cutting status. When processing materials with low dust generation, maintaining the rated parameters will generate unnecessary energy consumption; when processing materials with high moisture content, the fixed suction parameters are difficult to adapt to changes in material characteristics, and dust is prone to adherence and accumulation inside the pipeline, affecting the unobstructed flow of the dust collection channel.

[0004] The aerodynamic impedance of the dust collection pipeline changes dynamically during operation. However, in conventional industrial environments, the maintenance cost of directly deploying fluid sensors to measure the pipeline resistance is high, making it impossible for the control system to obtain real-time impedance data to correct the linkage delay time. The constant start-stop delay time used by the system cannot adapt to the dynamic electrical response of the frequency converter at different operating frequencies. This results in material cutting starting before the wind pressure flow field is fully established during the start-up phase, causing initial dust overflow. During the shutdown phase, the timing of the airflow suction action and the stopping of the cutting action is not accurate enough, easily leading to residual dust remaining in the pipeline.

[0005] Therefore, this invention provides a method for the coordinated start-up and shutdown of grain harvesting and dust collection based on neural network prediction, in order to overcome the shortcomings of the prior art. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method for the coordinated start-up and shutdown of grain harvesting and dust collection systems based on neural network prediction. This method solves the problems of existing grain harvesting and dust collection systems using open-loop logic with fixed time constraints for start-up and shutdown control, which does not incorporate material conditions and pipeline resistance for feedforward adjustment. This results in initial dust overflow during the start-up phase, dust adhesion and accumulation on the pipe walls caused by high-moisture materials, and a lack of basis for system linkage delay correction due to the difficulty in directly measuring pipeline resistance.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] The first aspect of this invention provides a method for the coordinated start-up and shutdown of grain harvesting and dust collection based on neural network prediction, applicable to a grain harvesting and dust collection system including a grain harvesting and cutting unit, a dust collection fan unit, and a dust collection pipeline network, comprising the following steps:

[0009] The grain-cutting unit is controlled to perform a pre-detection action, and the high-frequency harmonic components of the cross-axis current are extracted as dynamic mechanical characteristic parameters. The dynamic mechanical characteristic parameters are combined with the obtained static physical property characteristic parameters to generate a system state characteristic vector.

[0010] The system state feature vector is input into the predictive neural network model to output the expected dust generation index and the dust pipe wall adhesion tendency index. The expected dust generation index is converted into the initial operating frequency. When the dust pipe wall adhesion tendency index reaches the preset risk threshold, the initial operating frequency is corrected to the target operating frequency. When the target operating frequency is less than the minimum critical frequency threshold, the minimum critical frequency threshold is used as the target operating frequency. When the dust pipe wall adhesion tendency index does not reach the preset risk threshold, the initial operating frequency is used as the target operating frequency.

[0011] The overall aerodynamic delay time is calculated based on the target operating frequency and the current dynamic aerodynamic impedance coefficient of the dust collection network. The dust collection fan unit is started according to the target operating frequency. After the overall aerodynamic delay time, the grain cutting unit is started. The timing phase from the start of the dust collection fan unit to the start of the grain cutting unit is taken as the timing window.

[0012] The electromagnetic torque sequence and actual rotor angular velocity sequence of the dust collection fan unit within the time window are obtained for dynamic decoupling, and the nonlinear additional components are extracted to update the current dynamic aerodynamic impedance coefficient.

[0013] Upon receiving a shutdown command, the system controls the grain-cutting unit to stop. After a preset shutdown delay has elapsed since the grain-cutting unit stopped, the system controls the dust collection fan unit to stop.

[0014] Furthermore, the operation of the predictive neural network model includes: receiving dynamic mechanical feature parameters containing high-frequency harmonic components of cross-axis current through a dynamic feature processing stream; processing the dynamic mechanical feature parameters using a bidirectional long short-term memory network to output a dynamic feature sequence; generating key vectors and value vectors from the dynamic feature sequence; receiving static physical property feature parameters through a static feature processing stream; processing the static physical property feature parameters using a multilayer perceptron to output a static feature vector; generating a query vector from the static feature vector; calculating the dot product of the query vector and the key vector; generating an attention weight matrix using a normalized exponential function; weighting and summing the attention weight matrix and the value vector to generate a high-order cross-feature tensor; inputting the high-order cross-feature tensor into the output layer; and mapping and outputting the expected dust generation index and the dust adhesion tendency index to the pipe wall.

[0015] In a preferred embodiment of the present invention, the training steps of the predictive neural network model include: acquiring unlabeled idle operation data; performing unsupervised pre-training on the multimodal feature fusion and dimensionality reduction coding network within the predictive neural network model based on the unlabeled idle operation data; generating virtual working condition samples by interpolation in the feature space using synthetic minority class oversampling technology; constructing a weighted smoothed average absolute error loss function, multiplying the error where the predicted value is lower than the true label value by a preset penalty amplification coefficient when calculating the gradient backpropagation of the dust pipe wall adhesion tendency index; performing supervised fine-tuning training using an adaptive moment estimation optimization algorithm with momentum decay, and performing integer quantization on the network weight matrix after training convergence.

[0016] Furthermore, the high-frequency harmonic components of the quadrature-axis current are extracted as dynamic mechanical feature parameters. The dynamic mechanical feature parameters are combined with the obtained static physical property feature parameters to generate a system state feature vector. This includes: obtaining a discrete-time domain sequence containing the high-frequency harmonic components of the quadrature-axis current and a transient electromagnetic torque sequence; combining the discrete-time domain sequence and the transient electromagnetic torque sequence into a two-dimensional feature matrix; inputting the two-dimensional feature matrix into a one-dimensional convolutional autoencoder for feature dimensionality reduction to extract a dynamic latent space vector; using the dynamic latent space vector as the dynamic mechanical feature parameter; and concatenating the dynamic mechanical feature parameter with the static physical property feature parameters containing storage cycle data and sampled moisture content data according to a preset dimensional order, and performing linear normalization processing to generate the system state feature vector.

[0017] In a preferred embodiment of the present invention, when the dust adhesion tendency index on the pipe wall reaches a preset risk threshold, the initial operating frequency is corrected to a target operating frequency, and when the target operating frequency is less than a minimum critical frequency threshold, the minimum critical frequency threshold is used as the target operating frequency. This includes: determining a preset decrease penalty coefficient by looking up a preset mapping table based on the dust adhesion tendency index on the pipe wall; multiplying the initial target operating frequency by the preset decrease penalty coefficient to generate a candidate target operating frequency; comparing the candidate target operating frequency with the minimum critical frequency threshold; when the candidate target operating frequency is greater than the minimum critical frequency threshold, the candidate target operating frequency is used as the target operating frequency; when the candidate target operating frequency is not greater than the minimum critical frequency threshold, the minimum critical frequency threshold is used as the target operating frequency.

[0018] Furthermore, the overall aerodynamic delay time is calculated based on the target operating frequency and the current dynamic aerodynamic impedance coefficient of the dust collection network, using the following formula:

[0019] ;

[0020] In the formula, The total aerodynamic delay time, For the target operating frequency, To accelerate the slope setting, This represents the current dynamic aerodynamic resistance coefficient. The basic fluid time constant.

[0021] In a preferred embodiment of the present invention, the electromagnetic torque sequence and the actual rotor angular velocity sequence of the dust collection fan unit within the time window are obtained for dynamic decoupling, and the cumulative net aerodynamic work function is constructed by the following formula:

[0022] ;

[0023] In the formula, for The cumulative net aerodynamic work function at time t. The start time of the timing window. The current sampling time within the time window. This refers to the transient electromagnetic torque of the fan motor. For the actual rotor angular velocity sequence variable, The system rotational inertia of the dust collection fan unit. for Actual rotor angular velocity at time t This represents the actual rotor angular velocity at the initial moment.

[0024] Furthermore, the nonlinear additional component is extracted to update the current dynamic aerodynamic impedance coefficient, including: performing differential calculation on the cumulative net aerodynamic work function to obtain the transient net aerodynamic power; dividing the transient net aerodynamic power by the actual rotor angular velocity at the corresponding moment to generate an aerodynamic drag torque sequence; performing quadratic polynomial fitting on the aerodynamic drag torque sequence relative to the actual rotor angular velocity sequence; extracting the quadratic term coefficient in the fitting result as the current drag coefficient; calculating the differential normalized ratio of the current drag coefficient relative to the clean pipeline reference drag coefficient; and using the differential normalized ratio as the nonlinear additional component.

[0025] In a preferred embodiment of the present invention, the extraction of nonlinear additional components to update the current dynamic aerodynamic impedance coefficient further includes: calculating the fusion value of the previous dynamic aerodynamic impedance coefficient and the nonlinear additional components using an exponentially weighted moving average algorithm; when the fusion value does not exceed a preset maximum impedance coefficient, the fusion value is used as the current dynamic aerodynamic impedance coefficient; when the fusion value exceeds the preset maximum impedance coefficient, the preset maximum impedance coefficient is used as the current dynamic aerodynamic impedance coefficient.

[0026] Furthermore, when the pre-detection action fails to extract the high-frequency harmonic component of the quadrature axis current, the historical average value of the dynamic mechanical characteristics of the same type of material in the previous batch recorded in the local non-volatile memory is called as the dynamic mechanical characteristic parameter, and the total aerodynamic delay time is added to the preset safety delay margin as the updated total aerodynamic delay time.

[0027] A second aspect of the present invention provides a grain harvesting and dust collection linkage start-stop system based on neural network prediction, comprising a physical execution module, a frequency conversion drive module, a logic control module, and an edge computing module;

[0028] The physical execution module includes a grain-cutting unit, a dust-collecting fan unit, and a dust-collecting pipe network. The dust-collecting pipe network is connected between the dust-collecting fan unit and the dust suction hood at the work point, and includes a horizontal main pipe, an expanded diameter settling section, and an end dust collection hopper. The expanded diameter settling section and the end dust collection hopper are used to receive high-moisture, high-viscosity dust that settles after the airflow energy is reduced, and the horizontal main pipe is used to form a continuous dust conveying channel.

[0029] The variable frequency drive module is electrically connected to the physical execution module, including a main drive frequency converter for driving the grain-cutting unit and a fan frequency converter for driving the dust collection fan unit; both the main drive frequency converter and the fan frequency converter are equipped with a field-oriented control loop, which is used to analyze the electromagnetic torque, stator current and rotor angular velocity operating parameters during motor operation;

[0030] The logic control module is connected to the frequency converter drive module to send operation control commands to the main drive frequency converter and the fan frequency converter, and to perform timing latching and delayed release on the start command of the grain cutting unit.

[0031] The edge computing module is connected to the logic control module via a data communication bus. The data communication bus is also connected to the main drive inverter and the fan inverter, enabling the edge computing module to read the operating parameters of the main drive inverter and the fan inverter, and enabling the logic control module to send control commands to the main drive inverter and the fan inverter. The edge computing module stores a multimodal feature fusion and prediction neural network model and is equipped with a network communication interface, which is used to obtain the static attribute parameters of materials in the external warehouse management system. The edge computing module is used to perform operating condition prediction, frequency calculation, and aerodynamic impedance decoupling operation based on the obtained operating parameters and static attribute parameters of materials, so as to realize the above-mentioned method of grain harvesting and dust collection linkage start-stop based on neural network prediction.

[0032] This invention provides a method for the coordinated start-up and shutdown of grain harvesting and dust collection based on neural network prediction. It has the following beneficial effects:

[0033] 1. This invention controls the grain-cutting unit to perform a pre-detection action, extracts the high-frequency harmonic components of the quadrature-axis current, and combines them with static physical property parameters. These are then input into a predictive neural network model to output the expected dust generation index and the pipe wall adhesion tendency index, which are subsequently converted into the target operating frequency. This design pre-calculates the required operating parameters of the equipment before continuous grain-cutting operations, avoiding the energy waste problem under low-load conditions that occurs when the system uses fixed operating parameters. It also prevents pipe network blockage caused by high moisture content materials.

[0034] 2. This invention utilizes a timing window formed by the priority start-up of the dust collection fan and the delayed start-up of the grain-cutting unit to obtain the electromagnetic torque sequence and actual rotor angular velocity sequence of the fan within this timing window for dynamic decoupling calculation. This process eliminates the influence of mechanical rotational inertia and extracts the pure aerodynamic work component, calculating the real-time impedance state of the pipeline network without adding external fluid sensors, providing a dynamic update basis for calculating the overall aerodynamic delay time of the next coordinated start-up and shutdown.

[0035] 3. This invention calculates the overall aerodynamic delay time based on the target operating frequency and the current dynamic aerodynamic impedance coefficient of the dust collection network, and uses this as the control benchmark for the linkage start-stop sequence. This method combines the electrical response time of the fan with the aerodynamic conduction time of the network, ensuring that the air pressure flow field is established before the dust cutting action during the start-up phase, and that the material cutting action is established before the air field suction stops during the stop phase. This overcomes the defect that conventional constant delay settings cannot adapt to the dynamic response rate of variable frequency equipment. Attached Figure Description

[0036] Figure 1 This is a system framework diagram of the present invention;

[0037] Figure 2 This is a flowchart of the method of the present invention;

[0038] Figure 3 This is a flowchart of the multi-source data fusion and state feature vector construction process of the present invention;

[0039] Figure 4 This is a flowchart of the neural network prediction and nonlinear wind speed control of the present invention;

[0040] Figure 5 This is the overall aerodynamic delay time calculation logic diagram of the present invention;

[0041] Figure 6 This is a diagram illustrating the timing latch interception and clean timing window construction process of the present invention;

[0042] Figure 7 This is a schematic diagram illustrating the kinetic energy decoupling, virtual sensing, and dynamic correction of aerodynamic impedance of the present invention.

[0043] Figure 8 This is a diagram illustrating the heterogeneous node computing power allocation and state switching constraint relationships of the present invention.

[0044] Figure 9 This is a graph showing the verification results of the system's resistance to degradation according to the present invention;

[0045] Figure 10 This is a diagram of the aerodynamic drag state space kernel density distribution model of the present invention.

[0046] Among them, 10 is the physical execution module; 11 is the grain cutting unit; 12 is the dust collection fan unit; 13 is the dust collection pipeline; 20 is the variable frequency drive module; 21 is the main drive frequency converter; 22 is the fan frequency converter; 30 is the logic control module; 40 is the edge computing module; and 50 is the data communication bus. Detailed Implementation

[0047] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Reference Figure 1 This invention provides a grain harvesting and dust collection linkage start-stop system based on neural network prediction. The system includes a physical execution module 10, a variable frequency drive module 20, a logic control module 30, and an edge computing module 40.

[0049] The physical execution module 10 includes a grain-cutting unit 11, a dust-collecting fan unit 12, and a dust-collecting pipe network 13. The dust-collecting pipe network 13 connects the dust-collecting fan unit 12 to the dust suction hood at the work point and includes a horizontal main pipe, an expanded-diameter settling section, and an end dust collection hopper. The expanded-diameter settling section and the end dust collection hopper are used to receive high-moisture, high-viscosity dust that settles after the airflow energy decreases, while the horizontal main pipe forms a continuous dust conveying channel.

[0050] The variable frequency drive module 20 is electrically connected to the physical execution module 10, and includes a main drive frequency converter 21 for driving the grain-cutting unit 11, and a fan frequency converter 22 for driving the dust collection fan unit 12. Both the main drive frequency converter 21 and the fan frequency converter 22 are equipped with a field-oriented control loop, which is used to analyze operating parameters such as electromagnetic torque, stator current and rotor angular velocity during motor operation.

[0051] The logic control module 30 is communicatively connected to the frequency converter drive module 20 and is used to issue operation control commands to the main drive frequency converter 21 and the fan frequency converter 22, and to perform timing latching and delayed release on the start command of the grain cutting unit 11.

[0052] The edge computing module 40 is connected to the logic control module 30 via a data communication bus 50. The data communication bus 50 is also connected to the main drive inverter 21 and the fan inverter 22, enabling the edge computing module 40 to read the operating parameters of the main drive inverter 21 and the fan inverter 22, and allowing the logic control module 30 to send control commands to the main drive inverter 21 and the fan inverter 22. The edge computing module 40 stores a pre-trained multimodal feature fusion and prediction neural network model and has a network communication interface. This interface is used to obtain static property parameters of materials from an external warehouse management system. The edge computing module 40 is used to perform operating condition prediction, frequency calculation, and aerodynamic impedance decoupling operations based on the obtained operating parameters and static property parameters of materials.

[0053] Reference Figure 2 The present invention provides a method for starting and stopping grain collection and dust collection based on neural network prediction, which includes the following steps.

[0054] S1, perform pre-job state acquisition and feature vector construction.

[0055] During the pre-sampling phase after the system is powered on, or before the formal continuous grain-cutting operation of the current batch is started, the logic control module 30 first controls the fan frequency converter 22 to operate at a preset minimum dust suction frequency, and controls the main drive frequency converter 21 to drive the grain-cutting unit 11 to contact the material of the current batch in a low-speed, short-time, and light-load manner to perform pre-detection actions.

[0056] During the pre-detection phase, the edge computing module 40 reads the operating data of the main drive inverter 21 via the data communication bus 50, obtains the dynamic mechanical characteristic parameters of the grain-cutting unit 11, and extracts the high-frequency harmonic components of the quadrature-axis current from the field-oriented control loop of the main drive inverter 21. The edge computing module 40 also calls the warehouse management system via the network communication interface to obtain the static physical property characteristics of the current batch of materials; subsequently, it combines the dynamic mechanical characteristic parameters containing the high-frequency harmonic components with the static physical property characteristics of the input model according to a preset dimensional order to form a system state feature vector.

[0057] The pre-detection action is not part of the formal continuous grain-grazing operation. Its duration, operating speed, and cutting depth are all lower than those of the formal cutting operation. It is used to obtain the transient contact characteristics of the current batch of material while minimizing disturbance to the working surface. After the pre-detection action is completed, the logic control module 30 controls the main drive frequency converter 21 to stop output and controls the fan frequency converter 22 to continue running at a preset minimum dust collection frequency for a preset emptying time, or keeps the fan frequency converter 22 in a low-speed standby state to remove residual dust that may have been brought into the dust collection pipe network 13 by the pre-detection action. The preset emptying time is pre-calibrated based on the effective volume of the dust collection pipe network 13 and the volumetric flow rate corresponding to the minimum dust collection frequency.

[0058] The pre-detection action is independent of the subsequent formal mechanical cutting command. After the pre-detection action is completed, the logic control module 30 enters a state of waiting for the formal mechanical cutting command after completing the purging process or confirming the end of the pre-detection phase. When the formal mechanical cutting command arrives, the logic control module 30 then executes steps S3 and S4 to pre-start the fan, latch the mechanical cutting command, and release it after a delay. If the fan inverter 22 remains in a low-speed standby state during the waiting period, the logic control module 30 reads the current output frequency in this state in step S3 and uses it as the input parameter for calculating the overall aerodynamic delay time.

[0059] S2, perform dust state prediction and target frequency determination.

[0060] The edge computing module 40 inputs the system state feature vector into the multimodal feature fusion and prediction neural network model, which performs nonlinear mapping to output the expected dust generation index and the dust pipe wall adhesion tendency index.

[0061] The edge computing module 40 converts the expected dust generation index into the target operating frequency of the fan inverter 22 according to the preset control domain mapping table. If the dust adhesion tendency index on the pipe wall reaches a preset high level, a nonlinear suppression strategy is triggered to lower the upper limit of the target operating frequency; at the same time, a physical floor mechanism is used to limit the lowered target operating frequency so that it is not lower than the frequency threshold corresponding to the minimum critical suspension wind speed required for the current material to remain in the horizontal pipe without settling.

[0062] S3 calculates the total aerodynamic delay time required for the fan to start ahead of schedule.

[0063] The logic control module 30 calculates the total aerodynamic delay time required to establish effective fluid negative pressure at the dust collection hood opening of the work point based on the target operating frequency, the acceleration slope set inside the fan inverter 22, and the current dynamic aerodynamic impedance coefficient of the dust collection network 13.

[0064] When the fan inverter 22 is already in the minimum dust collection frequency or low-speed standby state before receiving the formal mechanical cutting command, the logic control module 30 also reads the current output frequency of the fan inverter 22 and uses the current output frequency to calculate the overall aerodynamic delay time.

[0065] S4 performs latching and delayed release of mechanical cutting commands.

[0066] When the logic control module 30 receives the mechanical cutting command to trigger the grain cutting unit 11, it first performs timing latch interception on the mechanical cutting command to keep the main drive frequency converter 21 in an unstarted state; at the same time, the logic control module 30 sends a start-up command or frequency adjustment command of the target operating frequency to the wind turbine frequency converter 22.

[0067] After the overall aerodynamic delay time calculated in step S3, the logic control module 30 releases the latched mechanical cutting command, driving the grain-cutting unit 11 to enter the formal cutting operation. Thus, a clean timing window with a duration equal to the overall aerodynamic delay time is formed between the start of fan operation and the release of the mechanical cutting action.

[0068] When the wind turbine frequency converter 22 is in low-speed standby mode and the current output frequency is lower than the target operating frequency, the frequency adjustment command is an up-frequency command; when the current output frequency is greater than or equal to the target operating frequency, the frequency adjustment command is an operating command to maintain the current frequency or adjust to the target operating frequency.

[0069] S5 updates the pipeline impedance within a clean timing window.

[0070] Within the clean timing window, the grain-cutting unit 11 has not yet started, and there is no new dust input in the dust collection network 13 generated by the current formal cutting action. The edge computing module 40 continuously collects the estimated electromagnetic torque and actual rotor angular velocity of the fan inverter 22, and integrates the product of electromagnetic torque and rotor angular velocity to obtain the total mechanical work of the dust collection fan unit 12 during startup or frequency adjustment; subsequently, according to the rigid body dynamics relationship, the energy increment corresponding to the moment of inertia is deducted, and the net aerodynamic work is obtained through decoupling.

[0071] The edge computing module 40 constructs a cumulative net aerodynamic work function according to the sampling time, and performs differential or derivative calculations on the cumulative net aerodynamic work function to obtain the transient net aerodynamic power; then, combined with the actual rotor angular velocity at the corresponding time, it calculates the aerodynamic drag torque sequence.

[0072] Edge computing module 40 performs quadratic polynomial fitting on the aerodynamic drag torque sequence relative to the rotor angular velocity, extracts the current drag coefficient, and performs differential normalization processing on the current drag coefficient and the clean pipeline reference drag coefficient to obtain a nonlinear additional component characterizing the amount of dust adhesion and deterioration on the pipeline wall. Subsequently, edge computing module 40 updates the dynamic aerodynamic impedance coefficient of dust collection pipeline 13 based on net aerodynamic work and the nonlinear additional component, and writes the updated dynamic aerodynamic impedance coefficient into the shared data register or volatile data buffer of logic control module 30 for use in the next execution of step S3.

[0073] The specific implementation process of the above method will be further explained below with reference to the accompanying drawings.

[0074] Reference Figure 3Step S1 is used to obtain the dynamic response information of the grain-cutting unit 11 and the static physical property information of the current batch of materials before the operation, and organize the above information into a system state feature vector that can be called by the regression neural network model. Step S1 specifically includes the following sub-steps.

[0075] S101, establish the underlying data interaction link and obtain dynamic mechanical characteristics.

[0076] In one specific implementation, the edge computing module 40 establishes a low-level data interaction link with the variable frequency drive module 20 through the industrial real-time Ethernet protocol, and periodically reads the register address storing control variables in the main drive inverter 21.

[0077] When the main drive frequency converter 21 drives the grain-cutting unit 11 to perform pre-detection or formal cutting actions, its internal field-oriented control circuit outputs parameters characterizing the motor's operating state, such as electromagnetic torque and rotor angular velocity. The analytical process of electromagnetic torque and rotor angular velocity can be implemented using conventional algorithms in AC motor vector control, which will not be elaborated here. The edge computing module 40 uses the read electromagnetic torque and rotor angular velocity as the basic parameters of the dynamic mechanical characteristics.

[0078] S102 extracts the high-frequency harmonic components of the quadrature-axis current.

[0079] The main drive inverter 21 extracts the high-frequency harmonic components of the quadrature-axis current within the sampling period of its underlying control loop. Since the digital sampling frequency of the current loop inside the main drive inverter 21 is higher than the refresh frequency of the external communication bus, the high-frequency characteristics are preferentially pre-extracted inside the main drive inverter 21, and then the extraction results are read by the edge computing module 40.

[0080] In the AC motor control model, the quadrature-axis current is related to the output torque at the motor shaft end. When the grain-cutting unit 11 contacts and cuts the grain pile, changes in the local hardness, moisture aggregation state, and degree of crusting of the material are transmitted back to the motor rotor through mechanical stress waves, causing high-frequency waveform distortion in the quadrature-axis current. Based on this physical correlation, the high-frequency harmonic components in the quadrature-axis current can be used to characterize the transient resistance ripple during the cutting process.

[0081] Specifically, the main drive inverter 21 is equipped with a programmable digital signal processing unit. The programmable digital signal processing unit can access the quadrature-axis current sampling sequence in the field-oriented control loop, and extract the discrete-time sequence of the quadrature-axis current within a set time window, and perform a fast Fourier transform on the discrete-time sequence.

[0082] The time window is determined based on the rated speed and number of pole pairs of the main drive motor, and its length is an integer multiple of one mechanical rotation cycle of the main drive motor to maintain the periodic integrity of the captured signal. After Fast Fourier Transform, the quadrature-axis current signal is converted from the time domain to the frequency domain, and the main drive inverter 21 extracts the harmonic amplitude values ​​within a specific high-frequency band. The specific high-frequency band is selected as a non-characteristic harmonic frequency band that deviates from the inverter's fundamental frequency by an integer multiple to avoid the conventional electromagnetic noise generated by the inverter's switching devices.

[0083] Grain may form a crusty layer after long-term storage or local fermentation. The hardness and moisture content of this crust can alter the friction coefficient and shear yield strength on the cutting surface. When the cutting tool overcomes the non-uniform yield strength, it generates a periodic excitation force. This periodic excitation force corresponds to the amplitude of the high-frequency harmonic components of the quadrature-axis current. The main drive inverter 21 writes the extracted amplitude of a single high-frequency harmonic into the communication register, and the edge computing module 40 reads this amplitude through the data communication bus 50. In this way, high-frequency physical characteristics can be transmitted to the edge computing module 40 without consuming a large amount of bus bandwidth.

[0084] S103, obtain the static physical property parameters of the material and construct the state feature vector.

[0085] The edge computing module 40 sends a data retrieval request to the external warehouse management system through its network communication interface. Upon responding to the request, the warehouse management system distributes the static physical property parameters of the materials in the current batch. These static physical property parameters include the average storage period data of the bulk materials in the target warehouse, the most recently measured sampling moisture content data, and at least one parameter selected from the material type, particle size class, and bulk density class to determine aerodynamic properties.

[0086] The average storage cycle data and sampling moisture content data serve as static physical property parameters input to the regression neural network model; material type, particle size class, and bulk density class are used to determine the aerodynamic characteristics of the material being processed. The edge computing module 40, based on the material type, particle size class, bulk density class, and sampling moisture content, looks up the critical suspension velocity of the material being processed in a pre-stored critical suspension velocity calibration table. .

[0087] If a network communication anomaly causes data retrieval failure, the edge computing module 40 performs an exception takeover. When the material category of the current work batch can be determined by the local work task record, the edge computing module 40 calls the historical average value of the same type of material in the previous work batch recorded in the local non-volatile memory as a backup static physical property parameter; when the current material category cannot be determined, the edge computing module 40 calls the preset conservative static physical property parameter group and marks the control system as a degraded operation state.

[0088] If the pre-detection action fails to generate effective dynamic mechanical characteristic parameters, the edge computing module 40 classifies the reasons for the failure. When the failure is due to an emergency stop signal, a hardware fault in the main drive inverter 21, or a safety interlock signal trigger, the logic control module 30 stops executing subsequent formal mechanical cutting commands and outputs a fault alarm. When the failure is due to missing torque sampling, failure to extract high-frequency harmonic components of the quadrature-axis current, or failure to pass the dynamic mechanical characteristic validity verification, the edge computing module 40 calls the average dynamic mechanical characteristic of the same type of material in the previous effective batch as a backup dynamic mechanical characteristic parameter and marks the current control as a degraded operation state. In the degraded operation state, the logic control module 30 increases the overall pneumatic delay time according to the preset safety delay margin, so that the dust suction hood opening has more time to establish negative pressure before the grain cutting unit 11 starts.

[0089] After the edge computing module 40 obtains the basic parameters of dynamic mechanical features, the amplitude of high-frequency harmonic components of cross-axis current, and the static physical property parameters, it does not directly perform simple linear splicing. Instead, it constructs a heterogeneous input structure containing a time-series dynamic matrix and a static attribute vector, and performs feature dimensionality reduction and high-order representation extraction.

[0090] Specifically, the edge computing module 40 is internally configured with a feature dimensionality reduction preprocessing network. For discrete time-domain sequences and transient electromagnetic torque sequences containing high-frequency harmonic components of quadrature-axis current, due to their high sampling frequency, there is a large amount of redundant information and background electrical noise from the equipment. The edge computing module 40 combines the above high-frequency sequences into a two-dimensional feature matrix and inputs it into a pre-deployed one-dimensional convolutional autoencoder for dimensionality reduction processing. The encoder part of the one-dimensional convolutional autoencoder contains multiple dilated causal convolutional layers to expand the receptive field without losing temporal resolution and extract high-dimensional temporal features characterizing cutting resistance ripple. Subsequently, a global max pooling layer compresses the high-dimensional temporal features into a low-dimensional dynamic latent space vector, thereby achieving dimensionality reduction of the feature vector while preserving the physical characteristics of high-frequency mechanical collisions.

[0091] The edge computing module 40 concatenates the dynamic mechanical characteristic parameters, the amplitude of the high-frequency harmonic components of the quadrature-axis current, and the static physical property parameters input to the regression neural network model according to a preset dimensional order to construct an initial state array. The initial state array contains both electromagnetic torque values ​​on the order of hundreds of Newton-meters and moisture content values ​​expressed as decimals or percentages. To avoid excessive differences in dimensions and numerical ranges between different physical quantities, which could affect subsequent model inference, the edge computing module 40 performs linear normalization on the initial state array, mapping the data of each dimension to a unified dimensionless interval.

[0092] The normalization process follows the mathematical formula below:

[0093] ;

[0094] In the formula, Represents the normalized dimensionless system state characteristic value; This represents the initial extracted static physical dimension measured data, which includes the storage period time and the percentage of sampled moisture content. This represents the upper limit baseline value allowed for the corresponding physical dimension within the system's set operating range; This represents the lower limit of the allowable baseline value for the corresponding physical dimension within the system's set operating range.

[0095] and The values ​​are determined based on the motor rated parameter boundaries of the physical execution module 10 and the statistical range of local conventional material properties, and are fixed in the configuration file of the edge computing module 40. After normalization, the edge computing module 40 generates a system state feature vector. The data dimension of this system state feature vector is consistent with the number of nodes in the input layer of the regression neural network model, forming a 1×5 one-dimensional input tensor in this embodiment, which is used as the pre-input for subsequent impedance prediction and operating condition assessment.

[0096] See attached document Figure 4 Step S2 executes data mapping-based working condition prediction and anti-clogging pipe control, specifically including the following sub-steps:

[0097] The edge computing module 40 fuses the multimodal features input from the system state feature input unit into the internally stored data and performs nonlinear data mapping with the predictive neural network model. This model differs from conventional fully connected networks, employing a two-stream network topology based on a cross-attention mechanism.

[0098] This dual-stream network topology includes a dynamic feature processing stream, a static feature processing stream, and an attention fusion module. The dynamic feature processing stream receives dynamic latent space vectors and further mines the temporal evolution of mechanical stress waves during the pre-detection action through a one-layer bidirectional long short-term memory network, outputting a dynamic feature sequence, i.e., a key vector and a value vector. The static feature processing stream receives static physical property parameters and maps them to the same latent layer dimension as the dynamic features through a multilayer perceptron, outputting a static feature vector, i.e., a query vector.

[0099] In the attention fusion module, a multi-head cross-attention mechanism is introduced. This mechanism uses a static feature vector representing the essential physical properties of the current batch of materials as the query vector, and a dynamic feature sequence representing the transient cutting condition as the key and value vectors. An attention weight matrix is ​​generated by calculating the dot product of the query vector and the key vector and applying a normalized exponential function. This attention weight matrix dynamically measures the sensitivity correlation of different material properties to transient high-frequency mechanical vibrations. Subsequently, the attention weight matrix and the value vector are weighted and summed to obtain the fused high-order cross-feature tensor. This cross-attention mechanism allows the neural network to autonomously focus its attention on mechanical excitation frequency bands that easily lead to dust adhesion during inference, thereby improving the prediction accuracy for complex nonlinear conditions.

[0100] The higher-order cross-feature tensor is finally input to the output layer, which contains two continuous value nodes, and outputs the expected dust generation index and the dust adhesion tendency index to the pipe wall, respectively. The output layer uses a type function activation to limit the output value to a dimensionless range of zero to one. The larger the value, the higher the corresponding dust generation intensity or adhesion risk.

[0101] For the construction and training of the multimodal feature fusion and prediction neural network model, the training steps are deployed on a cloud server or a local high-performance training cluster, and the weights are distributed to the edge computing module 40 via the network. The specific training steps are as follows:

[0102] First, sample set construction and data augmentation. Heterogeneous input structures synchronously recorded during multiple historical batches of operations were collected as sample data. To address the scarcity of samples for extreme bag-clogging failure conditions, a synthetic minority oversampling technique was introduced into the extracted minority class dynamic latent space vectors. Virtual high-risk condition samples were generated by interpolation in the feature space, achieving a balanced distribution of positive and negative samples.

[0103] Next, a custom asymmetric loss function was constructed. Since the engineering cost of failing to predict dust adhesion leading to pipe blockage far outweighs the cost of falsely reporting dust adhesion causing fan frequency reduction, the system constructed an asymmetric weighted smoothed average absolute error loss function. This loss function, when calculating the gradient backpropagation of the dust adhesion tendency index, assigns a penalty amplification coefficient greater than one to errors where the predicted value is lower than the true label value, forcing the model to tend to output a more conservative adhesion tendency index when facing uncertain boundaries. A weight decay regularization term is also superimposed on the loss function to prevent overfitting.

[0104] Then, gradient descent training is performed in stages. First, the feature fusion module and output layer are frozen, and the one-dimensional convolutional autoencoder is pre-trained unsupervised using historical unlabeled wind turbine idle operation data. The weights of the dimensionality-reduced encoder are optimized using reconstruction loss to give it generalizable high-frequency noise filtering capabilities. Subsequently, the entire network structure is unfrozen, and an adaptive moment estimation optimization algorithm with momentum decay is introduced. A cosine annealing learning rate scheduling strategy is adopted, and end-to-end supervised fine-tuning training is performed using a sample set containing sensor-measured dust concentration conversion values ​​and bag filter pressure differential degradation rate labels.

[0105] Finally, model quantization and edge deployment. After training convergence, to accommodate the computing power limitations of the edge computing module 40, the floating-point weight matrix in the network model is quantized using 8-bit integers. The quantized model weight file, the scaling factor of the attention mechanism, and the activation function configuration file are stored in the non-volatile memory of the edge computing module 40 for online forward inference. The edge computing module 40 no longer performs backpropagation training during the online inference phase.

[0106] S202, the edge computing module 40, based on a preset control domain mapping table, converts the expected dust generation index, which is fused with multimodal features and output by the predictive neural network model, into the initial target operating frequency of the fan inverter 22. The control domain mapping table is pre-stored in the non-volatile memory of the edge computing module 40, and internally records the mapping coordinates between discrete dust generation indices and the inverter operating frequencies corresponding to the matched suction air volume.

[0107] The control domain mapping table is established through on-site calibration. During on-site calibration, under different material moisture contents, different cutting loads, and different target dust collection requirements, the minimum fan operating frequency that meets the dust collection requirements of the dust hood at the work point is recorded, and the dust generation index and the corresponding fan operating frequency are used to form discrete mapping coordinate points. The edge computing module 40 uses a piecewise linear interpolation algorithm to find the frequency value corresponding to the current expected dust generation index in the control domain mapping table, and uses it as the basic suction frequency required to maintain the cleanliness of the current cutting surface, i.e., the initial target operating frequency.

[0108] S203, the edge computing module 40 compares the dust adhesion tendency index on the pipe wall with the preset adhesion risk threshold. The adhesion risk threshold is set based on the statistical data of bag blockage accumulated over a long period of operation of the dust removal system, and the lower quartile of the statistical distribution of the tendency index corresponding to the historical bag blockage failure is taken as the adhesion risk threshold.

[0109] If the dust adhesion tendency index on the pipe wall does not reach the preset adhesion risk threshold, the system will issue the initial target operating frequency as the target operating frequency.

[0110] If the dust adhesion tendency index on the pipe wall reaches or exceeds the adhesion risk threshold, it indicates that the current cutting operation has come into contact with a high-moisture, high-viscosity fermented and hardened grain layer. The edge computing module 40 triggers a nonlinear suppression strategy, which introduces a suppression coefficient into the calculation logic to lower the upper limit of the initial target operating frequency and calculates and generates the corrected target operating frequency.

[0111] In the specific configuration, the initial target operating frequency is multiplied by a descent penalty coefficient ranging from 0.6 to 0.8 to obtain the corrected target operating frequency. The descent penalty coefficient is determined based on the dust pipe wall adhesion tendency index. The higher the dust pipe wall adhesion tendency index, the smaller the descent penalty coefficient. The edge computing module 40 determines the descent penalty coefficient between 0.6 and 0.8 through segmented table lookup or linear interpolation.

[0112] The control principle is that by reducing the frequency to weaken the airflow energy, gravity can overcome the weakened airflow drag in the expansion and settling section or the end dust collection hopper of the dust collection network 13, causing highly viscous dust to settle naturally in advance, avoiding the impact and adhesion of wet and sticky dust to the dust collector bags, which would cause bag clogging. The early natural settling occurs in the expansion and settling section or the end dust collection hopper, rather than in the horizontal main pipe; the edge computing module 40 limits the corrected target operating frequency by limiting the minimum critical frequency threshold, ensuring that the actual conveying wind speed in the horizontal main pipe is not lower than the minimum critical suspension wind speed of the current material, thereby reducing the risk of bag clogging and preventing powder accumulation and blockage in the horizontal main pipe.

[0113] The nonlinear suppression strategy is equipped with a physical safety net mechanism. Before outputting the final control command, the edge computing module 40 executes a maximum value comparison logic to ensure that the corrected target operating frequency is greater than or equal to the minimum critical frequency threshold. The minimum critical frequency threshold corresponds to the minimum critical suspension velocity required to prevent the current material from settling at the bottom of the horizontal pipe. Setting the minimum critical frequency threshold ensures that the fluid in the dust collection network 13 has the capacity to transport bulk materials, preventing powder accumulation and blockage in the horizontal main pipe due to excessively low wind speed.

[0114] The minimum critical frequency threshold is calculated using the following mathematical formula:

[0115] ;

[0116] In the formula, This represents the calculated lowest critical frequency threshold. The critical suspension velocity representing the material being processed is determined experimentally based on the aerodynamic characteristics of the bulk material and entered into the system. This represents the effective internal cross-sectional area of ​​the horizontal main pipe in dust collection network 13; This represents the rated volumetric flow rate of dust collection fan unit 12 under rated operating conditions. The nameplate shows the rated operating frequency parameters of the wind turbine inverter 22; The system safety margin coefficient is used to compensate for air leakage along the pipeline and wind speed loss caused by local flow field disturbances. Its value is set between 1.1 and 1.3, and is calibrated by on-site commissioning personnel based on the pipeline length and the number of bends. The edge computing module 40 outputs the value after being subject to dual constraints of nonlinear suppression strategy and physical safety mechanism as the final target operating frequency.

[0117] Reference Figure 5 Step S3 is used to determine the waiting time for the fan to start ahead of schedule based on the fan's electrical response capability and the current aerodynamic state of the dust collection network 13. Step S3 specifically includes the following sub-steps.

[0118] S301, read the equipment electrical parameters and pipeline status parameters.

[0119] The logic control module 30 reads the acceleration slope setting value from the internal parameter register of the wind turbine inverter 22 through the underlying communication protocol. The acceleration slope is the rate at which the output frequency of the wind turbine inverter 22 changes with time during the startup or frequency ramp-up phase, in Hz / s.

[0120] The logic control module 30 also accesses the shared data register, volatile data buffer, or non-volatile data storage area to retrieve the latest recorded dynamic aerodynamic resistance coefficient of the dust collection network 13. The dynamic aerodynamic resistance coefficient is used to characterize the degree of fluid resistance degradation caused by residual dust from historical operations in the dust collection network 13. This coefficient is obtained in subsequent step S5 through decoupling of the fan electromechanical data and written into the shared data register or volatile data buffer of the logic control module 30; it is then synchronously saved to the non-volatile data storage area when the current batch of operations ends, the system is shut down, or the preset storage period is reached.

[0121] During the initial power-on of the system, or in the first operating cycle after the dust collection network 13 has just been cleaned and maintained, if the impedance data updated in step S5 has not yet been generated, the logic control module 30 calls the preset initial reference impedance coefficient in the firmware as the current dynamic aerodynamic impedance coefficient. The initial reference impedance coefficient is set to 1.0, which is used to represent the reference aerodynamic resistance level of the dust collection network 13 in a clean state.

[0122] After step S5 completes an effective impedance update, the logic control module 30 prioritizes calling the latest dynamic aerodynamic impedance coefficient in the shared data register or volatile data buffer; if the latest dynamic aerodynamic impedance coefficient does not exist or the verification is abnormal, the most recent effective impedance coefficient stored in the non-volatile data storage area is called.

[0123] S302, calculate the overall aerodynamic delay time.

[0124] The logic control module 30 calculates the overall aerodynamic delay time based on the target operating frequency, the acceleration slope of the fan inverter 22, and the current dynamic aerodynamic impedance coefficient. The overall aerodynamic delay time refers to the time required from the time the fan inverter 22 receives the start-up command or frequency adjustment command until a stable negative pressure flow field reaching the preset working negative pressure threshold is established at the dust collection hood opening at the work point.

[0125] The establishment of fluid negative pressure is not instantaneous. The process includes two parts: first, the electromechanical response time required for the fan motor to overcome rotational inertia and reach the target speed during startup or frequency adjustment; second, the aerodynamic conduction time required for air to overcome friction resistance, replace the original air in the dust collection pipe network 13, and form a stable flow field.

[0126] When the wind turbine inverter 22 starts from zero frequency to the target operating frequency, the total aerodynamic delay time is calculated according to the following relationship:

[0127] ;

[0128] In the formula, Represents the total aerodynamic delay time, in seconds; This represents the target operating frequency determined in step S2, in Hz. The acceleration slope of the wind turbine inverter 22 represents the acceleration slope. The basic fluid time constant represents the dust collection pipe network 13; This represents the current dynamic aerodynamic resistance coefficient.

[0129] When the wind turbine inverter 22 is already in a low-speed standby or coasting state before receiving the start-up command or frequency adjustment command, the logic control module 30 reads the current output frequency of the wind turbine inverter 22. The electrical response time is determined based on the difference between the target operating frequency and the current output frequency. If... If it is zero, then it is calculated according to the method of starting from zero frequency; if Greater than or equal to If the electrical response time term is zero, only the aerodynamic conduction time required for the dust collection network 13 to establish a stable negative pressure flow field is calculated.

[0130] In the above calculation relationship, the first part corresponds to the response time of the fan's electrical layer, and the second part corresponds to the additional delay caused by the fluid resistance of the pipeline network. As the amount of deposits on the inner wall of the dust collection pipeline 13 increases, the effective flow cross-sectional area of ​​the pipeline decreases, the surface roughness increases, and the frictional resistance that needs to be overcome to draw the air column in the pipe to a steady flow velocity increases accordingly, thus prolonging the establishment time of the stable negative pressure flow field at the dust collection hood opening.

[0131] Basic fluid time constant Related to the spatial volume, pipeline length, air density, and pipeline propagation characteristics of the dust collection pipeline 13, the test back-calculation calibration can be established by the closed pipeline under no-load negative pressure during the system installation and commissioning stage, and then solidified in the logic control module 30.

[0132] By introducing a dynamic aerodynamic impedance coefficient, the logic control module 30 can adjust the fan's advance start time according to the pipeline deterioration status. When the dynamic aerodynamic impedance coefficient increases, the calculated overall aerodynamic delay time is correspondingly extended. The logic control module 30 writes this overall aerodynamic delay time into the timer interrupt register as the timing reference for releasing the mechanical cutting command in step S4, so that when the grain-cutting unit 11 starts cutting, a negative pressure flow field matching the current working conditions has been formed at the dust suction hood opening of the work point.

[0133] Reference Figure 6 Step S4 is used to establish a dust collection negative pressure before the start of the grain-cutting unit 11, and to use this delay period to obtain fan response data that is not affected by the current cutting dust. Step S4 specifically includes the following sub-steps.

[0134] S401, latch the mechanical cutting command and initiate the fan-side response.

[0135] The logic control module 30 receives mechanical cutting commands from external operators or the host computer. Without timing-based coordinated control, mechanical cutting commands typically directly trigger the main drive inverter 21 to start. In this embodiment, upon receiving a mechanical cutting command, the logic control module 30 first triggers an internal timer interrupt program and then uses a timing latch interception mechanism to set the execution status register of the mechanical cutting command to a suspended waiting state, thus temporarily preventing the main drive inverter 21 from starting.

[0136] Within the same control cycle of the latched mechanical cutting command, the logic control module 30 sends a start-up command or frequency adjustment command carrying the target operating frequency to the wind turbine inverter 22. To reduce the impact of industrial communication latency and equipment underlying response time on timing accuracy, the logic control module 30 polls the status feedback word of the wind turbine inverter 22; when it receives a start-up confirmation signal, frequency adjustment confirmation signal, or target frequency hold confirmation signal returned by the wind turbine inverter 22, the logic control module 30 records this moment as the initial physical timestamp. And simultaneously start the countdown program.

[0137] The timer interrupt, status register latching, and instruction release processes can be implemented by the programmable logic controller according to the industrial control language. The specific instruction format can be determined according to the controller model and communication protocol used.

[0138] S402 constructs a clean timing window and releases the mechanical cutting command with a delay.

[0139] The logic control module 30 uses the total aerodynamic delay time calculated in step S3 as the countdown threshold. From the start of the dust collection fan unit 12 or the frequency adjustment until the end of the countdown, the main drive frequency converter 21 remains locked and stopped. The mechanical cutting tool of the grain-shearing cutting unit 11 does not perform cutting action, and the bulk material on the working surface is not damaged or agitated by this formal cutting action.

[0140] During this delay period, the dust collection hood at the end of the dust collection network 13 was not affected by the new dust input generated by the current grain-cutting action. The electromechanical response on the fan side mainly reflects changes in ambient airflow, the network's inherent resistance, and resistance caused by historical deposits. This delay period is defined as the clean time window, and its time interval is expressed as follows:

[0141] ;

[0142] In the formula, Represents a pure time series window interval; The initial physical timestamp of the logic control module 30 after sending a start-up command, frequency adjustment command, or target frequency hold command to the wind turbine frequency converter 22 is used to obtain confirmation feedback. This represents the total aerodynamic delay time calculated in step S3.

[0143] The purpose of setting a clean timing window is to exclude dust disturbances generated by the current cutting action from the data acquisition stage. Since the grain-raking cutting unit 11 does not perform continuous cutting actions within this window, there is no new dust load generated by the current continuous cutting action in the dust collection network 13. Subsequent steps can use the fan electrical parameters collected within this window to decouple the background aerodynamic resistance of the network and the historical dust adhesion deterioration state.

[0144] Once the countdown reaches the total aerodynamic delay time, the logic control module 30 releases the latched state of the mechanical cutting command and sends a start signal to the main drive frequency converter 21, enabling the grain-cutting unit 11 to enter the formal cutting operation. At this time, a negative pressure flow field matching the current target operating frequency and pipeline impedance state has been established at the dust suction hood opening of the work point, thereby enabling suction and collection of dust at the initial stage of dust generation.

[0145] During the countdown, if the logic control module 30 receives a system emergency stop command or a job cancellation command, it immediately resets the countdown program, clears the execution status register in the pending state, and issues a free stop command to the wind turbine inverter 22 to prevent the mechanical cutting command from being released accidentally after the countdown ends.

[0146] Reference Figure 7Step S5 is used to estimate the resistance degradation state of the dust collection pipeline 13 using the fan electromechanical response data within the clean timing window without additional pipeline differential pressure sensors, and to feed back the updated dynamic aerodynamic impedance coefficient to the subsequent timing control process. Step S5 specifically includes the following sub-steps.

[0147] S501 collects the wind turbine electromechanical response data within the clean timing window.

[0148] The edge computing module 40 performs data acquisition within the clean timing window formed in step S4. During this time period, the grain-cutting unit 11 has not yet entered the formal continuous cutting state; residual dust that may have been brought into the dust collection network 13 by the pre-detection action has been discharged or reduced to the preset allowable range through the preset evacuation process. Therefore, there is no new dust input in the dust collection network 13 generated by the current formal cutting action. The fan response data collected in this stage mainly reflects the background aerodynamic resistance of the network and the resistance changes caused by historical deposits.

[0149] The wind turbine inverter 22 records the electromagnetic torque and actual rotor angular velocity of the dust collection fan unit 12 according to a preset sampling period in its internal control loop, and writes the sampled data into the internal ring buffer. The edge computing module 40 reads the electromagnetic torque time series and actual rotor angular velocity time series in the ring buffer in batches through the high-speed communication bus. Its reading refresh frequency is higher than the control cycle of the logic control module 30, so as to obtain complete transient response data during the fan start-up or frequency adjustment phase.

[0150] Before performing impedance updates, the edge computing module 40 determines whether the range of rotor angular velocity change within the clean time window exceeds a preset effective speed change threshold. If the range of rotor angular velocity change does not reach the preset effective speed change threshold, it indicates that the current data is insufficient to support subsequent quadratic polynomial fitting. The edge computing module 40 does not write the current fitting result into the dynamic aerodynamic impedance coefficient and keeps the previous effective dynamic aerodynamic impedance coefficient unchanged. When it is necessary to generate operating status prompts, the edge computing module 40 can use the difference between the steady-state torque and the steady-state reference value of the clean pipeline network as an auxiliary monitoring quantity for maintenance prompts or abnormal warnings.

[0151] S502, decoupling yields net aerodynamic work.

[0152] The edge computing module 40 performs rigid body dynamics decoupling calculations based on the electromagnetic torque and actual rotor angular velocity collected within the clean timing window. The total mechanical work output by the dust collection fan unit 12 during the start-up or frequency adjustment phase mainly includes two parts: one part is used to overcome the rotational inertia of the motor rotor and impeller body to complete the speed change; the other part is used to overcome the fluid resistance in the dust collection pipe network 13 to drive the air column flow.

[0153] When the wind turbine inverter 22 is in a near-constant speed operating state within the pure timing window, the kinetic energy increment corresponding to the moment of inertia approaches zero. At this time, the edge computing module 40 can use the total mechanical work or steady-state torque data in this state as an auxiliary monitoring quantity for the pipeline resistance state; if the near-constant speed state results in the rotor angular velocity variation range being insufficient to support quadratic polynomial fitting, then the current dynamic aerodynamic impedance coefficient overwrite will not be performed, and the previous effective dynamic aerodynamic impedance coefficient will continue to be used.

[0154] Under the condition of valid sampling, the edge computing module 40 integrates the product of electromagnetic torque and actual rotor angular velocity within the pure time window to obtain the total mechanical work; then, it subtracts the kinetic energy increment corresponding to the moment of inertia within the same time interval to obtain the net aerodynamic work used to overcome fluid resistance. During the calculation, the edge computing module 40 constructs a cumulative net aerodynamic work function according to the sampling time. This function represents the time from the start of the pure timing window. Up to any sampling time Between these, the cumulative energy used to overcome the fluid resistance within the dust collection pipe network 13.

[0155] Net work done by pneumatic motion is calculated according to the following relationship:

[0156] ;

[0157] In the formula, This represents the net aerodynamic work done within the pure time window; when used for subsequent transient power calculations... Expressed as a function of cumulative net aerodynamic work over time Its final value is equal to the total net aerodynamic work done at the end of the pure time window; The start timestamp representing the pure timing window; Represents the duration of the pure timing window; Represents the time series function of electromagnetic torque; A time series function representing the actual rotor angular velocity; The equivalent moment of inertia of the dust collection fan unit 12; The rotor angular velocity represents the end of the pure timing window; The rotor angular velocity at the start of the pure timing window is used to subtract the initial kinetic energy already present in the fan's coasting or low-speed standby state.

[0158] Equivalent moment of inertia During the initial commissioning phase of the equipment, the motor parameters can be determined by the self-identification program built into the fan inverter 22 and then written into the edge computing module 40.

[0159] S503, extract the nonlinear additional component.

[0160] Edge computing module 40 pairs of cumulative net aerodynamic work functions Differential or differential calculations are performed to obtain the transient net aerodynamic power. This transient net aerodynamic power is then divided by the actual rotor angular velocity at the corresponding moment to obtain the aerodynamic drag torque sequence characterizing fluid resistance. To avoid calculation divergence caused by low-speed division, the edge computing module 40 discards the corresponding sampling point when the actual rotor angular velocity is lower than a preset effective speed threshold. To reduce the impact of high-frequency electromagnetic interference in the industrial environment on the differential results, the edge computing module 40 performs first-order low-pass filtering on the relevant sequences.

[0161] In an ideal clean pipeline, the aerodynamic resistance torque of the fan typically exhibits a stable relationship with the square of the rotational speed. However, as sticky dust adheres to the inner wall of the dust collection network 13 during historical operations, the wall roughness increases, and the local effective flow cross-sectional area decreases. Consequently, the actual aerodynamic resistance torque will be superimposed with an additional nonlinear upward component on top of the basic relationship of the clean pipeline network.

[0162] Based on the above relationship, the edge computing module 40 performs a quadratic polynomial fitting on the filtered aerodynamic drag torque sequence relative to the actual rotor angular velocity to obtain the fitting relationship between aerodynamic drag torque and rotational speed under the current pipeline network condition, and extracts the current drag coefficient from it. The edge computing module 40 will use the current drag coefficient. Compared with the pre-stored clean pipeline reference resistance coefficient Compare and according to Normalization yields a dimensionless nonlinear additional component characterizing the degree of historical dust adhesion degradation. .

[0163] In one specific implementation, the quadratic polynomial fitting relationship is expressed as:

[0164] ;

[0165] In the formula, Represents aerodynamic drag torque; Represents the actual rotor angular velocity; Represents the coefficient of the quadratic term related to the fluid resistance of the pipeline network; and It is used to absorb mechanical friction, low-speed bias, and measurement deviation. The edge computing module 40 will use the coefficients of the quadratic term obtained from the current fitting. As the current resistance coefficient .

[0166] Clean pipeline reference resistance coefficient The quadratic term reference fitting coefficient is obtained by performing a quadratic polynomial fitting on the aerodynamic drag torque sequence and rotor angular velocity using the same test procedure after the initial installation or deep cleaning of the dust collection pipe network 13. The edge computing module 40 calculates the nonlinear additional components according to the following relationship:

[0167] ;

[0168] when Not greater than hour, Take zero; when Greater than hour, It increases as the deviation between the two increases. Through this calculation process, the resistance degradation characterization quantity of the dust collection pipeline 13 can be obtained using the electromechanical data on the fan side without installing a pipeline differential pressure sensor.

[0169] S504, update the dynamic aerodynamic impedance coefficient and write it to the control side data area.

[0170] The edge computing module 40 obtains the nonlinear additional components from step S503. Update the dynamic aerodynamic impedance coefficient of dust collection pipe network 13. Before the update, edge computing module 40 first... Perform amplitude limiting; when When the preset maximum additional component threshold is exceeded, the preset maximum additional component threshold is used in subsequent calculations.

[0171] In one specific implementation, the edge computing module 40 employs an exponentially weighted moving average algorithm to fuse the previous effective dynamic aerodynamic impedance coefficient with the current nonlinear additional component, obtaining candidate values ​​for the dynamic aerodynamic impedance coefficient. Subsequently, the edge computing module 40 applies upper and lower bound constraints to these candidate values, ensuring that the final updated dynamic aerodynamic impedance coefficient... Limited to 1.0 to the preset maximum impedance coefficient Between. Where 1.0 corresponds to the baseline state of the clean piping network, The corresponding upper limit for the allowable degradation of pipeline resistance.

[0172] The dynamic aerodynamic drag coefficient is updated according to the following relationship:

[0173] ;

[0174] In the formula, This represents the updated current dynamic aerodynamic drag coefficient; This represents the previous effective dynamic aerodynamic resistance coefficient; λ represents the dimensionless nonlinear additional component; λ represents the smoothing filter coefficient, which ranges from 0.1 to 0.3 and can be tuned according to the on-site sampling frequency and pipeline length to suppress parameter abrupt changes caused by single measurement errors.

[0175] After the calculation is completed, the edge computing module 40 will update the dynamic aerodynamic impedance coefficient. The data is written to the shared data register or volatile data buffer of the logic control module 30; it is then synchronously saved to the non-volatile data storage area when the current batch of work ends, the system stops, or the preset save period is reached. The updated dynamic aerodynamic impedance coefficient is called in the next execution of step S3 to correct the overall aerodynamic delay time.

[0176] If the edge computing module 40 detects a fault in the wind turbine inverter 22, an emergency stop signal, missing speed sampling, or excessive torque sampling within the clean timing window, or if the calculated nonlinear additional components exceed the physically reasonable range, then the current impedance update is deemed invalid, and the previous valid dynamic aerodynamic impedance coefficient remains unchanged. If the number of valid sampling points is lower than a preset threshold, or the rotor angular velocity variation range is lower than a preset effective speed variation threshold, causing the quadratic polynomial fitting to fail to meet the validity condition, the edge computing module 40 also will not overwrite the dynamic aerodynamic impedance coefficient.

[0177] Therefore, the system completes the closed-loop correction of the aerodynamic resistance state of the pipeline network through the fan electromechanical data in the pure timing window, so that the fan running time in the next working cycle can be adaptively adjusted according to the actual deterioration state of the dust collection pipeline network 13.

[0178] Taking the outgoing operation of high-moisture corn in a large grain reserve as an example, this paper details the data flow and instruction generation process of the start-stop control system for grain unloading and dust collection based on neural network prediction.

[0179] The target material for the current batch is set as corn with a storage period of 180 days and a sampled moisture content of 18.5%. After the system powers on and establishes the underlying data interaction link, the logic control module 30 sends a low-frequency preparatory command to the wind turbine inverter 22, and simultaneously controls the main drive inverter 21 to drive the grain-shearing and cutting unit 11 to perform a pre-detection action. The edge computing module 40 periodically reads the quadrature-axis current sampling sequence of the main drive inverter 21 and extracts the amplitude of high-frequency harmonic components through fast Fourier transform. The current detected amplitude reaches 4.2 amperes, indicating that the bottom material has formed a physical agglomeration due to long-term stacking and high humidity, and there is high-frequency stress distortion at the contact interface of the cutting tool.

[0180] The edge computing module 40 inputs high-frequency harmonic features into a one-dimensional convolutional autoencoder to perform spatial dimensionality reduction, and concatenates them with static features such as a 180-day storage period and 18.5% moisture content to reconstruct a system state feature input unit. After receiving this input unit, the multimodal feature fusion and prediction neural network model maps feature weights through a multi-head cross-attention mechanism. The output node analyzes the expected dust generation index of the current working condition as 0.85 and the dust adhesion tendency index of the pipe wall as 0.78.

[0181] The adhesion risk threshold set in the non-volatile memory is 0.70. The system compares the current tendency index of 0.78, which exceeds the limit, and determines that the dust contains highly moist adhesive components, triggering the frequency reduction suppression boundary condition.

[0182] In the frequency reconstruction phase, the control system extracts the initial target operating frequency based on the expected dust generation index interpolation, denoted as . The current value is 45Hz. For high-risk adhesion, the corresponding degradation penalty coefficient is determined by looking up the table and recorded as follows: The current value is 0.75. Simultaneously, the system pre-calibrates the critical suspension wind speed of corn within the horizontal pipe, corresponding to the lower limit of the frequency reference, denoted as... The current value is 30Hz. The final target operating frequency is denoted as... The solution model is as follows:

[0183] ;

[0184] Substitute the specific operating parameters into the logic gate of the above physical safety net mechanism for calculation:

[0185] ;

[0186] The edge computing module 40 ultimately outputs 33.75Hz as the target operating frequency. By actively reducing the frequency, the flow field drag is reduced, forcing highly viscous dust to settle naturally in advance in the dust collection hopper expansion zone, thus cutting off the risk chain of dust collector bag clogging at the aerodynamic source.

[0187] Entering the timing-coordinated delay calculation stage. Logic control module 30 reads the acceleration slope set by the wind turbine inverter 22 and records it as... The value is taken as 5Hz / s; the time constant of the solidified base fluid is recorded as... The value is 2.5s; and the dynamic aerodynamic impedance coefficient recorded in the previous cycle in the volatile buffer is read and recorded as... The current value is 1.2. The total aerodynamic delay time is denoted as... Its solution model integrates both electrical response and aerodynamic transmission dimensions:

[0188] ;

[0189] Substituting the current state parameters into the calculation yields the specific timing sequence:

[0190] ;

[0191] Upon receiving the external mechanical cutting command, the logic control module 30 executes a timing latch to intercept and suspend the cutting start signal, simultaneously sending a 33.75Hz operation control word to the wind turbine inverter 22 and triggering a 9.75s countdown interrupt. Within this clean timing window, the cutting surface remains stationary, preventing new disturbances. The edge computing module 40 performs inertia stripping calculations on the wind turbine's electromechanical parameters to determine the net aerodynamic work, and fits and outputs the current true drag coefficient, denoted as... The current calculated value is 0.070. The system retrieves the baseline resistance coefficient of the clean pipeline network and records it as... The calibration value is 0.050. The control unit performs differential normalization on the two, and calculates the nonlinear additional component, denoted as... :

[0192] ;

[0193] The system then introduces a smoothing filter coefficient, denoted as... The set value is 0.2. Based on the exponentially weighted moving average algorithm, the dynamic aerodynamic impedance coefficient for the current period is denoted as... Perform state space update:

[0194] ;

[0195] ;

[0196] The updated impedance coefficient of 1.24 is written back to the buffer as the feedforward input for the next batch of work. The 9.75s interrupt countdown resets to zero, the logic control module 30 releases the suspended start signal, and the main drive inverter 21 outputs torque. At the moment the grain-cutting unit 11 contacts the grain pile, a steady-state negative pressure flow field capable of carrying the 33.75Hz suction kinetic energy has been established at the dust extraction hood, perfectly achieving seamless capture of the initial dust generation peak.

[0197] The engineering verification of this solution was carried out using a dual-line parallel testing platform with equivalent physical parameters. The target area was a high-humidity fermenting corn zone with a moisture content of 17.0% to 19.5%, and 20 consecutive batches were continuously recorded. Production line 1 used a fixed 5-second start-up delay and a 50Hz fixed-frequency suction strategy. Production line 2 deployed the prediction and timing-based collaborative architecture disclosed in this invention. Multi-channel data acquisition front-ends simultaneously recorded environmental dust concentration and differential pressure degradation data.

[0198] The detailed comparison of test environment indicators is shown in the table below:

[0199] <![CDATA[Average daily dust concentration in test area mg / m 3 > 12.4 4.1 Dust collector bag pressure differential deterioration rate (Pa / day) 185 42 Dust collection fan consumption per ton of grain (kWh / t) 0.85 0.52 Number of unplanned downtimes triggered by filter screen caking 3 0

[0200] Physical calculations demonstrate that this control architecture effectively overcomes the technical bottlenecks of traditional open-loop dust removal systems. The aerodynamic delay time, dynamically reconstructed from the mathematical model, achieves sub-second timing alignment between the fan's pressure build-up process and the dust generation peak from mechanical cutting, completely suppressing reactive suction caused by premature startup and initial dust escape caused by delayed startup. The polarization difference in the rate of pressure differential degradation proves that the neural network possesses high-dimensional analytical capabilities for heterogeneous features. Precisely guided frequency reduction actions allow high-risk dust particles to escape airflow entrainment, successfully ensuring the long-term permeability of the core filter element.

[0201] Figure 8 This paper presents the heterogeneous node computing power allocation and state switching constraints of this architecture. The top-level graph domain depicts the dual-stream input topology of the edge-side neural network nodes, where dynamic mechanical features and static physical parameters undergo feature dimensionality reduction and hidden layer encoding through multi-head cross-attention. The middle-level graph domain presents the nonlinear suppression decision tree of the control decision layer. After the logic gate judges the risk trend, the branch calls the frequency reconstruction mathematical model mentioned above to perform down-pressure frequency reduction. The lower-level timing axis strictly limits the instruction lifecycle, marks the start and end limits of the pure timing window triggered by the timing latch interception along the initial physical timestamp, and indicates the interlock execution logic of pipeline impedance observation and mechanical cutting release in vector form.

[0202] Figure 9 The system's resistance to degradation is verified using a multidimensional time series joint distribution. The horizontal axis represents the cumulative amount of long-term continuous operation, and the vertical axis represents the absolute value of the fluid impedance pressure drop transmitted from the sensor. The red band manifold maps the pressure difference transition trajectory under traditional control of Line 1. Its shape exhibits an extremely high time gradient, and the step-like steep rise indicates that the constant high wind speed causes deep intrusion and embedding of high-humidity dust. The sharp drop at the peak represents the system failure shutdown and forced manual dust removal intervention. The dark blue narrow band envelope maps the operating condition of the linkage control strategy of Line 2. This manifold is continuously maintained at the low-damping safety lower bound, with a very small trajectory slope and a convergent confidence interval, mathematically confirming the effective protection of the filter bed's microscopic permeability by the adaptive frequency reduction and pure time-series closed-loop synergistic mechanism.

[0203] Figure 10A kernel density distribution model was constructed based on the state space of equipment work and aerodynamic drag. The horizontal dimension represents the total mechanical work consumed by each test batch, while the vertical dimension maps the independently solved dynamic aerodynamic impedance coefficient within a pure time window. Warm and cool color gradations render the probability density clustering kernel of the multi-batch solution data. Discrete feature points in the figure form an asymptotically connected manifold representing the increase in surface roughness of the pipeline network along the diagonal direction. The contour bands distributed outside the high-density core accurately map the mathematical boundary of the convergence of electromechanical high-frequency sampling noise by the exponentially weighted moving average filtering algorithm. This topology rigorously proves that the dynamic decoupling model of this invention has a highly realistic aerodynamic friction inversion capability.

[0204] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for the coordinated start-up and shutdown of grain harvesting and dust collection based on neural network prediction, applied to a grain harvesting and dust collection system including a grain harvesting and cutting unit, a dust collection fan unit, and a dust collection pipeline network, characterized in that... Includes the following steps: The grain-cutting unit is controlled to perform a pre-detection action, and the high-frequency harmonic components of the cross-axis current are extracted as dynamic mechanical characteristic parameters. The dynamic mechanical characteristic parameters are combined with the obtained static physical property characteristic parameters to generate a system state characteristic vector. The system state feature vector is input into the predictive neural network model to output the expected dust generation index and the dust pipe wall adhesion tendency index. The expected dust generation index is converted into the initial operating frequency. When the dust pipe wall adhesion tendency index reaches the preset risk threshold, the initial operating frequency is corrected to the target operating frequency. When the target operating frequency is less than the minimum critical frequency threshold, the minimum critical frequency threshold is used as the target operating frequency. When the dust pipe wall adhesion tendency index does not reach the preset risk threshold, the initial operating frequency is used as the target operating frequency. The overall aerodynamic delay time is calculated based on the target operating frequency and the current dynamic aerodynamic impedance coefficient of the dust collection network. The dust collection fan unit is started according to the target operating frequency. After the overall aerodynamic delay time, the grain cutting unit is started. The timing phase from the start of the dust collection fan unit to the start of the grain cutting unit is taken as the timing window. The electromagnetic torque sequence and actual rotor angular velocity sequence of the dust collection fan unit within the time window are obtained for dynamic decoupling, and the nonlinear additional components are extracted to update the current dynamic aerodynamic impedance coefficient. Upon receiving a shutdown command, the system controls the grain-cutting unit to stop. After a preset shutdown delay has elapsed since the grain-cutting unit stopped, the system controls the dust collection fan unit to stop.

2. The method for coordinated start-up and shutdown of grain harvesting and dust collection based on neural network prediction according to claim 1, characterized in that, The operation process of a multimodal feature fusion and prediction neural network model includes: Dynamic mechanical feature parameters containing high-frequency harmonic components of cross-axis current are received through a dynamic feature processing stream. The dynamic mechanical feature parameters are processed by a bidirectional long short-term memory network to output a dynamic feature sequence. The key vector and value vector are generated from the dynamic feature sequence. Static physical property parameters are received through a static feature processing stream, and a static feature vector is output by processing the static physical property parameters using a multilayer perceptron. The query vector is then generated from the static feature vector. Calculate the dot product of the query vector and the key vector, generate the attention weight matrix using the normalized exponential function, and then perform a weighted summation of the attention weight matrix and the value vector to generate a high-order cross-feature tensor. The higher-order cross-feature tensor is input into the output layer to map and output the expected dust generation index and the dust adhesion tendency index to the pipe wall.

3. The method for coordinated start-up and shutdown of grain harvesting and dust collection based on neural network prediction according to claim 1, characterized in that, The training steps for a multimodal feature fusion and prediction neural network model include: Obtain unlabeled idle running data, and perform unsupervised pre-training on the dimensionality reduction coding network in the multimodal feature fusion and prediction neural network model based on the unlabeled idle running data; Virtual operating condition samples are generated by interpolation in the feature space using synthetic minority class oversampling techniques. A weighted smoothed average absolute error loss function is constructed. When calculating the gradient backpropagation of the dust adhesion tendency index on the pipe wall, the error where the predicted value is lower than the true label value is multiplied by a preset penalty amplification factor. Supervised fine-tuning training is performed using an adaptive moment estimation optimization algorithm with momentum decay. After training convergence, integer quantization is performed on the network weight matrix.

4. The method for coordinated start-up and shutdown of grain harvesting and dust collection based on neural network prediction according to claim 1, characterized in that, The high-frequency harmonic components of the quadrature-axis current extracted based on the pre-detection action are obtained as dynamic mechanical characteristic parameters, and static physical property characteristic parameters are obtained. The dynamic mechanical characteristic parameters and static physical property characteristic parameters are combined to form a system state characteristic vector, including: Obtain the discrete-time sequence and transient electromagnetic torque sequence containing the high-frequency harmonic components of the quadrature-axis current, and combine the discrete-time sequence and transient electromagnetic torque sequence into a two-dimensional feature matrix; The two-dimensional feature matrix is ​​input into a one-dimensional convolutional autoencoder for feature dimensionality reduction, and the dynamic latent space vector is extracted. The dynamic latent space vector is used as the dynamic mechanical feature parameter. The dynamic mechanical characteristic parameters and the static physical property characteristic parameters, which include data on the storage cycle and sampled moisture content, are concatenated according to a preset dimension order, and linear normalization is performed to generate a system state feature vector.

5. The method for coordinated start-up and shutdown of grain harvesting and dust collection based on neural network prediction according to claim 1, characterized in that, The initial target operating frequency is corrected to generate a target operating frequency. The target operating frequency is compared with a minimum critical frequency threshold. If the target operating frequency is less than the minimum critical frequency threshold, the minimum critical frequency threshold is used as the target operating frequency. This includes: The preset reduction penalty coefficient is determined by looking up a preset mapping table based on the dust adhesion tendency index of the pipe wall. The initial target running frequency is multiplied by a preset descent penalty coefficient to generate candidate target running frequencies; The candidate target operating frequency is compared with the minimum critical frequency threshold. If the candidate target operating frequency is greater than the minimum critical frequency threshold, the candidate target operating frequency is taken as the target operating frequency; if the candidate target operating frequency is not greater than the minimum critical frequency threshold, the minimum critical frequency threshold is taken as the target operating frequency.

6. The method for coordinated start-up and shutdown of grain harvesting and dust collection based on neural network prediction according to claim 1, characterized in that, The overall aerodynamic delay time is calculated based on the target operating frequency and the current dynamic aerodynamic impedance coefficient of the dust collection network, including: Divide the target operating frequency by the acceleration slope set value to generate the wind turbine electrical response time parameters; Multiply the current dynamic aerodynamic impedance coefficient by the basic fluid time constant to generate the aerodynamic conduction time parameters of the pipeline network; The total aerodynamic delay time is generated by summing the electrical response time parameters of the wind turbine and the aerodynamic conduction time parameters of the pipeline.

7. The method for coordinated start-up and shutdown of grain harvesting and dust collection based on neural network prediction according to claim 1, characterized in that, The electromagnetic torque sequence and actual rotor angular velocity sequence of the dust collector fan unit within the timing window are obtained for dynamic decoupling calculation, including: The total mechanical work is generated by integrating the product of the electromagnetic torque sequence and the actual rotor angular velocity sequence over time. Calculate the kinetic energy increment caused by rotational inertia between the start and end times of the time window; The net aerodynamic work is generated by subtracting the kinetic energy increment from the total mechanical work, and the cumulative net aerodynamic work function is constructed according to the sampling time.

8. The method for coordinated start-up and shutdown of grain harvesting and dust collection based on neural network prediction according to claim 7, characterized in that, Extracting nonlinear additional components to update the current dynamic aerodynamic impedance coefficient, including: The transient net aerodynamic power is obtained by differential calculation of the cumulative net aerodynamic work function. The transient net aerodynamic power is divided by the actual rotor angular velocity at the corresponding moment to generate the aerodynamic drag torque sequence. The aerodynamic drag torque sequence is fitted with a quadratic polynomial relative to the actual rotor angular velocity sequence, and the coefficients of the quadratic term in the fitting result are extracted as the current drag coefficient. Calculate the differential normalized ratio of the current resistance coefficient to the reference resistance coefficient of the clean pipeline network, and use the differential normalized ratio as an additional nonlinear component.

9. The method for coordinated start-up and shutdown of grain harvesting and dust collection based on neural network prediction according to claim 8, characterized in that, Extracting nonlinear additional components to update the current dynamic aerodynamic impedance coefficient also includes: An exponentially weighted moving average algorithm is used to calculate the fusion value of the previous dynamic aerodynamic impedance coefficient and the nonlinear additional component; When the fusion value does not exceed the preset maximum impedance coefficient, the fusion value is used as the current dynamic aerodynamic impedance coefficient; when the fusion value exceeds the preset maximum impedance coefficient, the preset maximum impedance coefficient is used as the current dynamic aerodynamic impedance coefficient.

10. The method for coordinated start-up and shutdown of grain harvesting and dust collection based on neural network prediction according to claim 1, characterized in that, If the high-frequency harmonic component of the cross-axis current is not extracted during the pre-detection action, the historical average value of the dynamic mechanical characteristics of the same type of material in the previous batch recorded in the local non-volatile memory is called as the dynamic mechanical characteristic parameter, and the total aerodynamic delay time is added to the preset safety delay margin as the updated total aerodynamic delay time.

Images