Emulsion pump self-adaptive liquid supply method and system based on time series deep learning

Abstract

The application relates to the field of automation control technology and discloses an emulsion pump self-adaptive liquid supply method and system based on time sequence deep learning, which collects emulsion pump station outlet fluid state data and synchronously captures hydraulic support action instructions of a fully-mechanized coal mining face, constructs a multi-dimensional space-time feature matrix, inputs a BP-LSTM combined prediction model, uses an LSTM branch and a BP branch to respectively and parallelly predict the steady-state trend and the transient mutation of the flow, and outputs a total predicted flow demand value through a BP fusion network weighted output; the prediction value is reversely mapped into a feedforward basic target frequency, a fuzzy PID feedback compensation correction frequency based on real-time pressure deviation solving is combined, and a final control frequency is superimposed to drive the emulsion pump; the application can effectively solve the response lag problem caused by the missing of cross-modal semantics and physical inertia in traditional control, suppresses the transient pressure collapse and the water hammer effect, and realizes constant pressure liquid supply under the high-yield and high-efficiency working condition.

Description

Technical Field

[0001] This invention relates to the field of automation control technology, and more specifically, to an adaptive liquid supply method and system for emulsion pumps based on time-series deep learning. Background Technology

[0002] As the core power source of large-scale industrial hydraulic transmission systems, the stability of the output pressure and flow rate of emulsion pump stations directly affects the actuation accuracy of the end hydraulic actuators and the operational smoothness of the entire transmission network. With the development of industrial automation towards high productivity, high efficiency, heavy load, and high speed, higher requirements are placed on the hydraulic power source's fluid supply response speed, pressure stabilization accuracy, and adaptability to sudden load changes.

[0003] Chinese Patent CN116909148B discloses a control method and system for an emulsion pump station. This method constructs an operating parameter matrix by collecting historical real-time data from the system, establishes a hydraulic prediction model to obtain hydraulic prediction results, and combines measured hydraulic data with calculation and compensation prediction to attempt closed-loop control of the frequency converter speed. Chinese Patent CN102588244B discloses an emulsion pump station and its control system. This method adopts an architecture combining centralized control by a main controller and distributed control by sub-controllers. It can monitor the water quality, liquid level, temperature, and pressure of the fluid medium in real time, realizes the switching between automatic and manual modes and status monitoring, and improves the reliability and ease of maintenance of the pump station to a certain extent.

[0004] However, while existing technologies have improved the monitoring capabilities of pump stations to some extent or introduced basic predictive methods, they still struggle to resolve the fundamental contradiction of dynamic supply-demand matching when faced with complex operating conditions such as high-frequency concurrent actions of end-effectors or large-flow load switching under long-distance liquid supply conditions. Specifically, existing control systems suffer from severe problems of "cross-modal semantic deficiency" and "single-model dimension collapse" at the algorithm level. Because existing predictive and control logic often treats the actions of the end-load as unpredictable external random disturbances, it completely ignores the key modal information of "deterministic action intent" contained in the electro-hydraulic control system, resulting in the control system being "blind" to impending sudden changes in operating conditions. In addition, the single predictive model attempts to simultaneously fit the "steady-state trend" during the steady-state operation of the system and the "transient pulse" caused by the load action, leading to overfitting during the steady-state period and underfitting during the sudden change period.

[0005] This dual limitation of information and model makes it impossible for the supply side to effectively cope with the "ideal step function" type flow change caused by the opening of microsecond-level solenoid valves at the load end. The control response cannot overcome the physical lag window formed by the fluid transmission delay over long-distance pipelines and the mechanical rotational inertia of high-power motors. The direct consequence is that at the moment the actuator group moves, a millisecond-level "transient pressure collapse" inevitably occurs at the end of the fluid supply pipeline, resulting in insufficient driving force or reduced support stiffness of the actuators, affecting the accuracy of the mechanical action. When the pump station responds to full-speed fluid supply with a lag, the load action has often ended, and the huge fluid inertia immediately triggers a severe "water hammer effect," causing fatigue wear of high-pressure pipelines and precision sealing components, seriously affecting the continuous and stable operation of the hydraulic power system and the service life of the equipment. Summary of the Invention

[0006] To overcome the aforementioned deficiencies of existing technologies, this invention provides an adaptive fluid supply method and system for emulsion pumps based on temporal deep learning. By fusing fluid state data and hydraulic support action commands to construct a multi-dimensional spatiotemporal feature matrix, and using a BP-LSTM combined prediction model to analyze the trend and abrupt change characteristics of the flow rate in parallel, combined with feedforward inverse mapping and fuzzy PID feedback control, this method can effectively break down the information barriers between supply and demand, eliminate the physical lag window, and completely solve the problem of alternating oscillations between "weak legs" and "water hammer".

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

[0008] An adaptive liquid supply method for emulsion pumps based on temporal deep learning includes:

[0009] The system collects fluid state data from the emulsion pump station outlet and pipeline state parameters from the supply pipeline, and captures hydraulic support action commands. A multidimensional spatiotemporal feature matrix is ​​constructed based on the fluid state data and hydraulic support action commands. The fluid state data includes at least the real-time pump station outlet pressure value. The pipeline transmission lag time τ is calculated based on the pipeline state parameters.

[0010] A BP-LSTM combined prediction model is constructed. The multidimensional spatiotemporal feature matrix is ​​input into the LSTM trend prediction branch and the BP mutation prediction branch of the BP-LSTM combined prediction model for parallel prediction. The trend prediction value output by the LSTM trend prediction branch and the mutation prediction value output by the BP mutation prediction branch are input into the BP fusion network in the BP-LSTM combined prediction model for weighted fusion, and the total predicted flow demand value is output.

[0011] Pressure loss compensation is applied to the total predicted flow demand value. The compensated total predicted flow demand value is then mapped inversely to the feedforward base target frequency. The feedforward advance is determined based on the pipeline transmission lag time τ. Fuzzy deviation state variables are generated based on the real-time pump station outlet pressure value. The feedback compensation correction frequency is calculated using fuzzy PID rules. The feedforward base target frequency and the feedback compensation correction frequency are superimposed to generate the final control frequency to drive the emulsion pump.

[0012] The method for constructing a multidimensional spatiotemporal feature matrix includes:

[0013] Variational mode decomposition is performed on the fluid state data to generate a denoised fluid state time-series vector. The hydraulic support action command is encoded into a working condition semantic feature vector. The denoised fluid state time-series vector and the working condition semantic feature vector are fused to construct a multi-dimensional spatiotemporal feature matrix.

[0014] The method for generating the denoised fluid state time-series vector includes:

[0015] A sliding time window of length N is set to extract historical pressure and flow data from the past N moments from the fluid state data to construct a basic fluid state time series vector.

[0016] The variational mode decomposition algorithm is used to adaptively decompose the basic fluid state time series vector to obtain K eigenmode components. The alternating direction multiplier method is used to iteratively solve the problem. High-frequency noise components are removed from the K eigenmode components and low-frequency effective signal components are retained. The low-frequency effective signal components are then reconstructed to generate a denoised fluid state time series vector.

[0017] The method for encoding hydraulic support action commands into working condition semantic feature vectors includes:

[0018] The action type features and action duration features are extracted from the hydraulic support action commands. The action type features are mapped to the action type one-hot encoded vector. The action type one-hot encoded vector is combined with the action duration features to generate the working condition semantic feature vector. The action types include column lowering action, support moving action, column raising action and push sliding action.

[0019] The method for constructing the multidimensional spatiotemporal feature matrix includes:

[0020] The denoised fluid state time-series vector and the working condition semantic feature vector are timestamped. The aligned denoised fluid state time-series vector is used as the fluid state dimension and the aligned working condition semantic feature vector is used as the working condition semantic dimension. A multidimensional spatiotemporal feature matrix is ​​constructed through matrix concatenation.

[0021] The method for inversely mapping the compensated total predicted flow demand value to the feedforward base target frequency includes:

[0022] A flow-frequency characteristic relationship model for an emulsion pump is established. The total predicted flow demand value is input into the flow-frequency characteristic relationship model for inverse mapping, and the feedforward basic target frequency is output.

[0023] The method for generating the fuzzy deviation state quantity includes:

[0024] The pressure deviation between the real-time pump station outlet pressure value and the preset target working pressure value is calculated. The first derivative of the pressure deviation is obtained to get the pressure deviation change rate. The pressure deviation and the pressure deviation change rate are quantified and fuzzified and mapped into fuzzy linguistic variables to generate fuzzy deviation state variables.

[0025] The method for calculating the feedback compensation correction frequency includes:

[0026] Based on the preset fuzzy inference rule table, fuzzy inference is performed on the fuzzy deviation state quantity to dynamically calculate the proportional parameter correction, integral parameter correction, and derivative parameter correction of the PID controller. The preset PID parameters of the PID controller are updated in real time according to the proportional parameter correction, integral parameter correction, and derivative parameter correction. The updated PID parameters are used to calculate the pressure deviation and output the feedback compensation correction frequency.

[0027] Amplitude limiting protection is applied to the final control frequency.

[0028] An adaptive liquid supply system for emulsion pumps based on temporal deep learning is provided to implement the aforementioned adaptive liquid supply method for emulsion pumps based on temporal deep learning. The system includes:

[0029] Feature construction module: used to collect fluid state data at the outlet of emulsion pump station and pipeline state parameters of supply pipeline and capture hydraulic support action commands, and construct a multi-dimensional spatiotemporal feature matrix based on fluid state data and hydraulic support action commands; the fluid state data includes at least the real-time pump station outlet pressure value; and calculates the pipeline transmission lag time τ based on pipeline state parameters;

[0030] Traffic forecasting module: Used to construct BP-LSTM combined forecasting model. The multidimensional spatiotemporal feature matrix is ​​input into the LSTM trend forecasting branch and BP mutation forecasting branch of the BP-LSTM combined forecasting model for parallel forecasting. The trend forecast value output by the LSTM trend forecasting branch and the mutation forecast value output by the BP mutation forecasting branch are input into the BP fusion network in the BP-LSTM combined forecasting model for weighted fusion, and the total predicted traffic demand value is output.

[0031] Liquid supply control module: used to compensate for pressure loss of total predicted flow demand value, inversely map the compensated total predicted flow demand value to feedforward base target frequency, determine feedforward advance based on pipeline transmission lag time τ, generate fuzzy deviation state quantity based on real-time pump station outlet pressure value, solve feedback compensation correction frequency through fuzzy PID rules, and superimpose feedforward base target frequency and feedback compensation correction frequency to generate final control frequency to drive emulsion pump.

[0032] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0033] This invention constructs a multi-dimensional spatiotemporal feature matrix that integrates fluid state data and hydraulic support action commands, introducing discrete working condition semantic features into the control system. This breaks down the cross-modal semantic barriers between the supply and demand sides, enabling the system to perceive deterministic action intentions. It employs a BP-LSTM combined prediction model and its parallel branch architecture. The LSTM branch accurately captures the continuous steady-state trend of the fluid system, while the BP branch quickly analyzes discrete step load changes caused by support actions. Furthermore, through adaptive weighting of the BP fusion network, it effectively overcomes the overfitting of single prediction models during stable periods and the underfitting during abrupt changes. By addressing the dimensional collapse defect of the combined system, this invention achieves high-precision deconstruction of the coupling characteristics of "steady-state trend" and "transient pulse". Based on this, the invention establishes a dual closed-loop control mechanism that coordinates feedforward prediction coarse adjustment and fuzzy PID feedback fine adjustment. It utilizes high-precision flow prediction to generate feedforward control quantity, adjusting the pump station output in advance before load changes occur. This fundamentally overcomes the physical lag window formed by fluid transmission delay and mechanical inertia delay, eliminating the "blindness" and lag in liquid supply control. It effectively solves the problem of alternating oscillations of "soft legs" and "water hammer" under high-frequency support movements, improving the pressure stability and equipment reliability of the liquid supply system. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 A flowchart of an adaptive liquid supply method for emulsion pumps based on temporal deep learning provided in an embodiment of the present invention;

[0036] Figure 2 A schematic diagram illustrating the hydraulic support action types and their corresponding fluid consumption characteristics and encoding vectors provided in an embodiment of the present invention;

[0037] Figure 3This is a flowchart of the parallel prediction process of the BP-LSTM combined prediction model provided in an embodiment of the present invention.

[0038] Figure 4 The following is a flowchart of the feedforward-fuzzy PID control provided in an embodiment of the present invention;

[0039] Figure 5 A functional block diagram of an emulsion pump adaptive liquid supply system based on temporal deep learning provided in an embodiment of the present invention;

[0040] Figure 6 This is a scatter plot comparing the predicted values ​​and actual values ​​of different models provided in the embodiments of the present invention. Detailed Implementation

[0041] The technical solutions of 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.

[0042] Example 1

[0043] Please see Figure 1 As shown, this embodiment provides an adaptive liquid supply method for emulsion pumps based on temporal deep learning, including:

[0044] Step S10: Collect fluid state data at the outlet of the emulsion pump station and pipeline state parameters of the supply pipeline, and capture hydraulic support action commands. Perform variational mode decomposition on the fluid state data to generate a denoised fluid state time-series vector. Encode the hydraulic support action commands into a working condition semantic feature vector. Merge the denoised fluid state time-series vector and the working condition semantic feature vector to construct a multi-dimensional spatiotemporal feature matrix. Calculate the pipeline transmission lag time τ based on the pipeline state parameters.

[0045] As the power source for hydraulic supports in fully mechanized coal mining faces, emulsion pump stations drive hydraulic supports to perform actions such as lowering, moving, raising, and pushing conveyors using high-pressure emulsion. There is a direct causal relationship between the fluid state at the pump station outlet and the actions of the hydraulic supports: when the hydraulic supports perform actions, the consumption of emulsion leads to a decrease in pump station outlet pressure and an increase in flow demand; when the hydraulic supports stop operating, the reduced emulsion consumption leads to a rebound in pressure and a decrease in flow demand. Fluid state data includes two core parameters: pump station outlet pressure and pump station outlet flow rate. The pump station outlet pressure reflects the pressure level at the output of the emulsion pump, and the pump station outlet flow rate reflects the volume of emulsion delivered by the pump to the pipeline system per unit time. Both together characterize the real-time fluid supply status of the emulsion pump station. Traditional fluid supply control systems only collect pressure or flow values ​​as control inputs and process them as isolated time series. This approach ignores the fundamental cause of flow changes—the actions of the hydraulic supports—causing the control system to respond only after the flow change has occurred, failing to anticipate upcoming sudden changes in flow demand, resulting in control lag. Step S10 simultaneously collects fluid state data and hydraulic support action commands, and merges the two into a unified multidimensional spatiotemporal feature matrix. This enables the deep learning prediction model in the subsequent step S20 to simultaneously acquire information from two dimensions: the current state of the flow and the cause of the flow change. This gives the prediction model the ability to perform causal reasoning on the flow change, providing holographic data input for the accurate prediction in step S20.

[0046] The pipeline state parameters of the fluid supply pipeline include pipeline length, pipeline inner diameter, and pipeline friction coefficient. These parameters determine the transmission time required for the emulsion to be transferred from the pump station outlet to the hydraulic support of the fully mechanized mining face, as well as the pressure loss during the transmission process. The longer the pipeline, the smaller the pipeline inner diameter, and the larger the friction coefficient, the longer the transmission lag time and the greater the pressure loss. The fluid supply pipeline of the fully mechanized mining face typically consists of a main fluid supply pipe, branch pipes, and support connecting hoses, with a total length ranging from hundreds to thousands of meters. The pressure loss caused by pipeline flow resistance is not negligible, and the response delay caused by pipeline transmission lag is an important factor affecting the fluid supply control performance. If pipeline flow resistance and transmission lag are not considered, feedforward control will not be able to accurately compensate for pressure fluctuations when flow demand changes, and the response speed of the control system will be limited by the physical characteristics of the pipeline. Step S10 collects pipeline state parameters and calculates pipeline transmission lag time, providing a basis for pipeline lag compensation for the feedforward control in the subsequent step S30, enabling the control system to issue adjustment commands in advance based on the actual physical characteristics of the pipeline.

[0047] Further, step S10 includes:

[0048] Step S11: The fluid state data includes the real-time pump station outlet pressure value and the real-time pump station outlet flow rate value. A sliding time window of length N is set to extract historical pressure data and historical flow rate data from the past N moments from the fluid state data to construct a basic fluid state time series vector.

[0049] The real-time pump station outlet pressure value refers to the instantaneous pressure value of the emulsion measured by the pressure sensor installed at the outlet pipeline of the emulsion pump station. This pressure value directly reflects the pressure level of the pipeline system when the pump station delivers emulsion to the working face. The pressure value decreases when the hydraulic support performs its fluid-consuming action and rises back to the system-set constant pressure value when the hydraulic support stops operating. The real-time pump station outlet flow value refers to the instantaneous volumetric flow rate of the emulsion measured by the flow sensor installed at the outlet pipeline of the emulsion pump station. This flow rate reflects the volume of emulsion pumped into the pipeline system per unit time. The flow rate rises sharply when multiple hydraulic supports operate simultaneously and falls accordingly when the hydraulic support operation decreases. Both the pressure sensor and the flow sensor adopt an intrinsically safe design for mining applications, operate at low DC voltage, and use current signals. The measurement data is transmitted to the control system in real time via an RS485 interface or industrial Ethernet. Sliding time window is a time series data processing technique. The system sets a fixed time interval of N and extracts the outlet pressure and outlet flow values ​​of N times before the current time at each sampling time to form a historical pressure data sequence and a historical flow data sequence.

[0050] For example, if the sampling period is Δt and N is set to 30, the sliding time window captures historical data from 30 sampling periods backward from the current moment, and the actual time span corresponding to the window length is 30 × Δt. The determination of the N value needs to comprehensively consider the dynamic response characteristics of the downhole fluid supply system and the typical operating cycle of the hydraulic support. If the N value is too small, the amount of data in the window will be insufficient to reflect the changing trend of the fluid state. If the N value is too large, too much outdated data will be introduced, thereby reducing the model's sensitivity to recent changes in operating conditions. Therefore, the N value should be set to the number of sampling points that can cover two to three complete operating cycles of the hydraulic support.

[0051] The basic fluid state time-series vector is constructed as follows: historical pressure data at N time points within the window are arranged chronologically to form a pressure time-series sub-vector; historical flow data at N time points within the window are arranged chronologically to form a flow time-series sub-vector; and the pressure and flow time-series sub-vectors are concatenated column-wise to form a basic fluid state time-series vector with dimension N×2. The use of a sliding time window allows the system to represent the fluid state using a continuous historical data sequence instead of relying on isolated single-point measurements. The time-series data within the window contains the changing trends and dynamic characteristics of the fluid state, laying a data foundation for the subsequent extraction of long short-term memory patterns by the LSTM neural network. The system continuously receives sensor data at a fixed sampling period Δt, constructing a "first-in, first-out" circular data buffer in memory. This means that at any current sampling time t, the system memory retains a continuous data sequence from time tN to time t-1. Therefore, the truncation operation is essentially an instantaneous indexing and extraction of the existing data queue in memory. This data acquisition method, which requires no additional waiting time, ensures the real-time performance of the algorithm, making the construction of a dataset containing historical information completely controllable in terms of computational resources. Because single-point measurements are significantly affected by sensor noise and transient disturbances, direct use would lead to unstable quality of subsequent model input data. A sliding time window, by capturing continuous time-series data, allows the model input to reflect the continuous characteristics of fluid dynamics, enhancing the input data's resistance to transient random disturbances. Pipeline state parameters are obtained as follows: pipeline length is obtained from the design drawings of the fully mechanized mining face's fluid supply system or through on-site measurements; the pipeline inner diameter is read from pipeline specifications; and the pipeline friction coefficient is obtained from engineering manuals or calibrated through on-site pressure testing based on the pipeline material and inner wall roughness. The formula for calculating the pipeline transmission lag time τ based on the pipeline state parameters is τ = L / v, where L is the total length of the fluid supply pipeline, and v is the average flow velocity of the emulsion in the pipeline. The average flow velocity v can be determined based on the current pump station output flow rate Q. pump The ratio of v to the pipe cross-sectional area A is calculated as v=Q pump / A. The pipeline transmission lag time τ characterizes the pure delay time required for the change in pump station output flow to be transmitted to the end of the working face after the control command is issued. This time is an incompressible physical lag and must be compensated for in feedforward control.

[0052] Step S12: The variational mode decomposition algorithm is used to adaptively decompose the basic fluid state time series vector to obtain K eigenmode components. The alternating direction multiplier method is used to iteratively solve the problem. High-frequency noise components are removed from the K eigenmode components and low-frequency effective signal components are retained. The low-frequency effective signal components are reconstructed to generate a denoised fluid state time series vector.

[0053] Step S12 employs a variational mode decomposition (VMD) algorithm to denoise the basic fluid state time series vector. Since the basic fluid state time series vector contains two channels—pressure and flow rate—the system independently performs VMD on the pressure and flow rate time series sub-vectors respectively. Then, the denoising results of the two channels are concatenated column-wise to form the denoised fluid state time series vector. VMD is an adaptive signal decomposition method. Its core idea is to decompose the original signal (i.e., the time series sub-vector of a single channel) into several eigenmode components with different center frequencies. Each eigenmode component has compact support characteristics in the frequency domain, and the spectra of each component are mutually separated. The working principle of VMD is based on the construction and solution of a constrained variational model: transforming the signal decomposition problem into an optimization problem, the goal of which is to find K eigenmode components such that the sum of the estimated bandwidths of each component is minimized, while constraining the sum of all components to equal the original signal. The constrained variational model is transformed into an unconstrained variational model by introducing Lagrange multipliers. An iterative solution is then performed using the alternating direction multiplier method, updating each eigenmode component and its center frequency in each iteration until a convergence condition is met. The convergence condition is determined by the ratio of the energy difference between the eigenmode components obtained in two adjacent iterations to the energy of the eigenmode component in the current iteration being less than a preset convergence threshold ε.

[0054] The alternating direction multiplier method is an iterative algorithm suitable for separable convex optimization problems. By decomposing a complex joint optimization problem into multiple simpler subproblems that are solved alternately, it can reduce computational complexity while ensuring convergence. The number K of intrinsic modal components needs to be preset according to the spectral characteristics of the signal. If the K value is too small, effective information will be missed; if the K value is too large, over-decomposition will result in redundant components. The K value can be determined by the center frequency monitoring method: gradually increase the K value and observe the center frequency distribution of each component. Stop increasing when center frequency overlap occurs or component energy is too small. The criterion for center frequency overlap is that the difference between the center frequencies of two adjacent components is less than half the sum of their bandwidths. The criterion for component energy being too small is that the component energy accounts for less than 1% of the total energy of the original signal. Based on the spectral characteristics of the fluid state signal of the emulsion pump station, the typical range of K value is 4 to 8. In practical applications, it can be adjusted within this range according to the number of hydraulic supports at the working face and the complexity of the pipeline system. The underground electromagnetic environment is complex. Electromagnetic interference from equipment such as high-voltage cables, frequency converters, and coal mining machines can couple into sensor signals, forming high-frequency noise. The frequencies of these noise components are usually much higher than the effective signal frequencies of fluid state changes. After variational mode decomposition, the system determines the attributes of each intrinsic mode component based on its center frequency and energy percentage: components with a center frequency higher than a preset frequency threshold or an energy percentage lower than a preset energy threshold are identified as high-frequency noise components and are removed; the remaining components are identified as low-frequency effective signal components and are retained. The frequency threshold is determined based on the bandwidth characteristics of the fluid dynamics response. The flow and pressure changes in the emulsion pump station are mainly driven by the movement of hydraulic supports. The movement frequency of hydraulic supports is usually within a few times per second; therefore, the frequency threshold should be set to several times the upper limit of the hydraulic support movement frequency to ensure that effective signals are not mistakenly removed. The energy threshold is determined based on the energy ratio characteristics of noise components and effective signal components. Components with an energy percentage lower than 1% of the total energy are usually identified as noise components.

[0055] The signal reconstruction process involves superimposing and summing the retained low-frequency effective signal components in the time domain to obtain a denoised fluid state time-series vector. Compared with traditional low-pass filtering methods, variational mode decomposition (VMD) has the advantages of strong adaptability and accurate frequency band separation. It can automatically determine the center frequency of each component based on the signal's own characteristics without pre-setting a cutoff frequency, avoiding the loss of effective signal or noise residue caused by a fixed cutoff frequency. Compared with empirical mode decomposition (EMD), VMD incorporates the decomposition process into an optimization problem through a variational framework, thus avoiding the mode aliasing phenomenon in EMD and giving each component a clear physical meaning. If the denoising process in step S12 is missing, the high-frequency noise mixed in the raw data collected in step S11 will directly enter the subsequent neural network model, causing the model to learn noise features rather than the true laws of flow changes during training, resulting in a significant decrease in prediction accuracy and a weakening of the model's generalization ability.

[0056] Step S13: Capture the hydraulic support action commands of the fully mechanized mining face in real time, extract action type features and action duration features from the hydraulic support action commands, map the action type features into an action type one-hot encoded vector, and combine the action type one-hot encoded vector with the action duration features to generate a working condition semantic feature vector.

[0057] The electro-hydraulic control system of the longwall mining face sends action commands to each hydraulic support via an industrial communication bus. Each action command includes information such as the target support number, action type, and action parameters. The system connects to the communication bus of the electro-hydraulic control system, listens in real time to the action command data frames transmitted on the bus, and extracts key information related to flow rate changes. The action type characteristic refers to the specific type of action performed by the hydraulic support. The hydraulic supports in the longwall mining face mainly perform four basic types of actions. See also... Figure 2 This is a schematic diagram of the hydraulic support action type and corresponding unique thermal encoding vector provided in the embodiments of this application. Figure 2 The text uses simplified mechanical diagrams and motion symbols to illustrate the physical forms of four types of movements: such as... Figure 2 As shown, the lowering action is represented by a downward-pointing motion indicator, indicating the hydraulic support column descends to release its supporting force on the roof. The shifting action is represented by a horizontal motion indicator, indicating the hydraulic support as a whole moves in the direction of the coal mining machine. The raising action is represented by a motion indicator pointing upwards, indicating the hydraulic support column rises to re-support the roof. The pushing action is represented by a motion indicator pointing towards the bottom scraper structure, indicating the hydraulic support pushes the scraper conveyor towards the coal face. Different action types have significantly different emulsion consumption: the raising action requires injecting a large amount of emulsion into the lower cavity of the column to generate supporting force, making it a high-emulsion-consumption action; the shifting action requires continuous emulsion supply to drive the traveling mechanism, making it a continuous-emulsion-consumption action; the lowering action mainly releases the emulsion in the upper cavity of the column, with relatively low emulsion consumption; the pushing action's emulsion consumption is between that of the raising and lowering actions. The action duration characteristic refers to the preset execution time of the action command. The duration of different actions directly determines the total amount of emulsion consumed by that action; the longer the duration, the greater the cumulative emulsion consumption. The action type feature is encoded using one-hot encoding, which is a method of converting discrete categorical variables into numerical vectors. For M possible categories, each category is encoded as a binary vector of length M, in which only one element is 1 and the rest are 0. The position of the element 1 corresponds to the index of the category in the category list.

[0058] For example, in combination Figure 2The one-hot encoding vectors shown at the bottom, as shown in Table 1, indicate that if there are four types of actions, numbered in the order of lowering the column, moving the support, raising the column, and pushing the slide, then the lowering column action is encoded as vector [1,0,0,0], the moving support action as vector [0,1,0,0], the raising column action as vector [0,0,1,0], and the pushing slide action as vector [0,0,0,1]. The use of one-hot encoding allows the neural network to treat all types of actions equally without artificial priority bias due to differences in numerical values, avoiding the misleading influence of numerical distance on model learning when using integer encoding. The encoding mapping table is constructed based on the actual hydraulic support types deployed on the working face and the instruction protocol of the control system. The correspondence between each action type code and the one-hot encoding vector is determined by parsing the communication protocol document of the electro-hydraulic control system.

[0059] Table 1. One-hot encoding vectors for the four action types

[0060] Action type Action Code One-hot encoded vector Falling column A01 [1,0,0,0] Moving shelf A02 [0,1,0,0] Lifting column A03 [0,0,1,0] push away A04 [0,0,0,1]

[0061] The duration of actions is expressed in units of time and normalized. The normalization method is to divide the actual duration by the standard duration of that type of action, making the duration characteristics of different types of actions comparable. The standard duration is determined by collecting historical execution data of each type of action under normal production conditions in the fully mechanized mining face, statistically analyzing the actual execution time of each type of action, and taking the median execution time of that type of action as the standard duration.

[0062] For example, the standard duration of the column lowering action can be set to 3 seconds, the standard duration of the support shifting action can be set to 8 seconds, the standard duration of the column raising action can be set to 5 seconds, and the standard duration of the push-slide action can be set to 4 seconds. Specific values ​​should be calibrated according to the actual hydraulic support model and electro-hydraulic control system parameters used at the working face. The semantic feature vector of the working condition is generated by concatenating the one-hot encoded vector of the action type with the normalized action duration feature column-wise to form a semantic feature vector of length M+1. Step S13 transforms the physical hydraulic support action into a digital semantic feature representation, enabling the control system to understand the root cause of flow fluctuations rather than merely observing the result of flow fluctuations, thus realizing a technological shift from passive response to active prediction. The action type information contained in the semantic feature vector of the working condition provides a direct basis for analyzing transient changes in the BP neural network branch of the subsequent step S20, while the action duration information provides a quantitative reference for predicting the magnitude of flow changes. Without the semantic feature encoding in step S13, the subsequent prediction model will only rely on historical traffic data for time-series extrapolation, unable to perceive upcoming actions and thus unable to predict sudden traffic changes in advance. The prediction model's ability to respond to sudden operating conditions will be greatly reduced.

[0063] Step S14: Perform timestamp alignment on the denoised fluid state time-series vector and the working condition semantic feature vector. Use the aligned denoised fluid state time-series vector as the fluid state dimension and the aligned working condition semantic feature vector as the working condition semantic dimension. Construct a multidimensional spatiotemporal feature matrix through matrix concatenation operation.

[0064] Step S14 performs spatiotemporal data alignment and feature fusion to generate a multidimensional spatiotemporal feature matrix. The denoised fluid state time-series vector generated in step S12 is continuously sampled data, with its sampling time determined by a fixed sampling period, generating one data point every sampling period Δt. The working condition semantic feature vector generated in step S13 is event-triggered data, with its generation time determined by the issuance time of the hydraulic support action command. The issuance of the action command is random and not synchronized with the sampling time of the fluid state data. The asynchrony of the two types of data on the time axis will lead to mismatch during feature fusion, causing fluid state data at a certain moment to be associated with an action command at an incorrect moment, disrupting the causal relationship between flow rate changes and actions. Timestamp alignment addresses the issue of time asynchrony in heterogeneous data from different sources. The process involves assigning an absolute timestamp, accurate to milliseconds, to each data point in the denoised fluid state time-series vector and each data point in the operating condition semantic feature vector. Using the sampling time of the denoised fluid state time-series vector as the reference time axis, the operating condition semantic feature vector with the closest timestamp to each reference time is matched. If no action command is given near a reference time, the operating condition semantic feature vector for that time is set to zero, indicating a state of no action. The timestamp accuracy must be higher than one-tenth of the sampling period to ensure negligible alignment errors. The timestamp source can be the unified time synchronization module of the control system or a synchronous clock based on a network time protocol.

[0065] The construction process of the multidimensional spatiotemporal feature matrix is ​​as follows: the aligned and denoised fluid state time-series vector is used as the fluid state dimension of the matrix, which includes two channels: historical pressure data and historical flow data; the aligned working condition semantic feature vector is used as the working condition semantic dimension of the matrix, which includes two channels: action type one-hot encoding and action duration; the data of the two dimensions are merged in the feature axis direction through matrix concatenation operation.

[0066] For example, if the sliding time window length N is 30 and the number of action types M is 4, then the multidimensional spatiotemporal feature matrix has a dimension of 30×7, where the first column is historical pressure data, the second column is historical flow data, the third to sixth columns are unique thermal encodings of action types, and the seventh column is the normalized action duration. Each row of the multidimensional spatiotemporal feature matrix corresponds to a historical moment, containing complete information on the fluid state and action semantics at that moment. The entire matrix represents the spatiotemporal evolution of the fully mechanized mining face fluid supply system within the time span covered by the sliding time window. If the timestamp alignment processing in step S14 is missing, the time misalignment between fluid state data and action commands will cause the model to learn incorrect causal relationships, making it unable to correctly understand which action will lead to which flow rate change, and the prediction accuracy will drop significantly due to the disordered causal relationship. The construction of the multidimensional spatiotemporal feature matrix enables the LSTM trend prediction branch and the BP mutation prediction branch in the subsequent step S20 to receive the same and complete input data. The two branches can extract long-term trend patterns and transient mutation features from the data respectively, providing comparable prediction results for the weight allocation of the fusion network in step S23.

[0067] Step S10 acquires the real-time operating status of the emulsion pump station by collecting fluid state data, obtains the physical characteristics of the supply pipeline by collecting pipeline state parameters and calculates the pipeline transmission lag time, obtains the driving factors of flow rate changes by capturing hydraulic support action commands, removes high-frequency noise affecting prediction accuracy through variational mode decomposition, transforms physical actions into computable digital representations through working condition semantic feature encoding, and constructs a multi-dimensional spatiotemporal feature matrix that contains both temporal evolution information and causal semantic information through spatiotemporal data fusion. This multi-dimensional spatiotemporal feature matrix holographically represents the operating status and changing driving factors of the fully mechanized mining face's fluid supply system, enabling the BP-LSTM combined prediction model in the subsequent step S20 to simultaneously perform temporal pattern mining and causal relationship inference, rather than blindly extrapolating from historical data alone. The semantic dimension information of the operating conditions contained in the multidimensional spatiotemporal feature matrix generated in step S10 enables the BP mutation prediction branch in step S22 to establish a direct mapping between action features and flow mutations. When the system detects a new action command, it can immediately predict the flow change caused by the action. This predictive ability based on causal reasoning is not available in traditional methods that rely solely on temporal extrapolation. The historical pressure and historical flow information contained in the denoised fluid state time-series vector output in step S10 enables the LSTM trend prediction branch in step S21 to extract long-cycle liquid supply patterns and periodic pulsation features through a gating mechanism, providing a baseline trend for flow prediction under stable operating conditions. The variational mode decomposition preprocessing in step S10 and the neural network prediction in step S20 form a synergy between signal processing and machine learning: variational mode decomposition removes narrowband noise interference that is difficult for neural networks to filter automatically from the signal processing level, while neural networks extract complex nonlinear mapping relationships that are difficult to describe by traditional signal processing methods from the machine learning level. The two techniques complement each other in their respective areas of expertise, enabling the overall prediction system to have both noise resistance and nonlinear modeling capabilities.

[0068] In step S10, variational mode decomposition eliminates high-frequency noise components through adaptive frequency band separation, significantly improving the signal-to-noise ratio of the input data entering the prediction model. Noise interference with model training and inference is greatly suppressed, allowing the model to focus on the true patterns of flow changes rather than noise features. The stability and reliability of the prediction results are correspondingly enhanced due to the improved quality of the input data. The introduction of working condition semantic feature vectors expands the input data from a single fluid state time series to multi-dimensional spatiotemporal features containing causal semantics. The prediction model gains the ability to understand the causes of flow changes. When the hydraulic support issues an action command, the model can immediately determine whether the action will cause the flow to increase or decrease, and the magnitude of the change. This causal reasoning ability significantly accelerates the prediction model's response speed to sudden working conditions. Timestamp alignment ensures the accuracy of time registration between fluid state data and action commands, avoiding causal relationship confusion caused by time misalignment during heterogeneous data fusion. The mapping relationship between actions and flow changes learned by the model remains consistent with the real relationship in the physical world, ensuring the physical rationality of the prediction results. The sliding time window dilutes the impact of a single outlier sampling point on the overall input by truncating continuous time-series data. Variational mode decomposition eliminates the influence of transient electromagnetic interference by removing high-frequency noise. Semantic feature encoding for operating conditions decouples the representation of action type from numerical magnitude through one-hot encoding. These multiple safeguards ensure that the entire data acquisition and feature construction process is highly adaptable to the complex and variable electromagnetic environment and operating conditions downhole. The system can still output high-quality feature data under harsh conditions. The multidimensional spatiotemporal feature matrix provides the LSTM trend prediction branch and the BP mutation prediction branch with their respective strengths in handling different data types. The LSTM branch extracts long-term patterns from the temporal dimension, while the BP branch analyzes mutation correlations from the semantic dimension. The capabilities of both branches are fully realized through multidimensional data input, and the overall performance of the prediction model reaches a high level due to the completeness of the input data.

[0069] Step S20: Construct a BP-LSTM combined prediction model. Input the multidimensional spatiotemporal feature matrix into the LSTM trend prediction branch and the BP mutation prediction branch of the BP-LSTM combined prediction model for parallel prediction. Input the trend prediction value output by the LSTM trend prediction branch and the mutation prediction value output by the BP mutation prediction branch into the BP fusion network in the BP-LSTM combined prediction model for weighted fusion and output the total predicted traffic demand value.

[0070] Specifically, the flow rate changes in emulsion pump stations exhibit two distinct patterns: when hydraulic supports are not performing any actions or only a few supports are performing low-fluid-consumption actions, the pump station flow rate shows a steady fluctuation trend. This trend is characterized by temporal continuity and periodic regularity, and is related to the inherent mechanical pulsations of the pump station and the fluid dynamics characteristics of the pipeline system. When hydraulic supports are performing high-fluid-consumption actions or multiple supports are operating simultaneously, the pump station flow rate exhibits a step-like transient characteristic. This abrupt change is directly related to the type and number of actions performed by the hydraulic supports, with large amplitude and short duration. A single prediction model cannot accurately capture both of these significantly different patterns simultaneously. If only models that are good at handling time-series data are used, the response to abrupt changes will be slow; if only models that are good at nonlinear mapping are used, they lack the ability to remember historical patterns. Step S20 constructs a BP-LSTM combined prediction model, using an LSTM neural network to extract long-term trend patterns and a BP neural network to analyze transient mutation correlations. The BP fusion network is used to achieve intelligent weighted fusion of the two types of prediction results, enabling the prediction model to adaptively allocate weights between trend prediction and mutation prediction according to the current operating conditions, thereby taking into account both the trend tracking capability under stable operating conditions and the rapid response capability under mutation operating conditions.

[0071] Further, see Figure 3 Step S20 includes:

[0072] Step S21: Construct an LSTM trend prediction branch. Input the multidimensional spatiotemporal feature matrix into the LSTM trend prediction branch. Use the gate structure composed of the LSTM forget gate, input gate and output gate of the LSTM trend prediction branch to filter the short-term random fluctuations in the multidimensional spatiotemporal feature matrix, extract the long-period fluid operation law, and output the trend prediction value.

[0073] The LSTM trend prediction branch in step S21 uses a Long Short-Term Memory (LSTM) neural network as its core architecture. LSTM is a variant of recurrent neural networks specifically designed for processing sequential data. Its core feature is the introduction of a gating mechanism to control the flow of information over time, thus overcoming the gradient vanishing problem faced by traditional recurrent neural networks when processing long sequences of data. The network structure of the LSTM trend prediction branch consists of three layers: an input layer, hidden layers, and an output layer. The hidden layers are composed of multiple stacked LSTM units, each containing four core components: a forget gate, an input gate, a cell state, and an output gate. The number of hidden layers in the LSTM trend prediction branch is set to L. LSTM The number of neurons in each hidden layer is set to N. LSTMThe number of hidden layers and neurons is determined using a grid search method. During the pre-training phase, different parameter combinations are traversed, and the prediction error index on the validation set is used as the evaluation criterion to select the parameter combination that achieves a low prediction error index. For example, the LSTM trend prediction branch can be set as a two-layer hidden layer structure, with each hidden layer having 128 neurons.

[0074] The function of the forget gate is to determine how much historical information should be retained and how much outdated information should be forgotten in a cell's state. The forget gate receives the hidden state h from the previous time step. (t-1) and the input x at the current time t The sigmoid activation function outputs a forgetting coefficient vector f between 0 and 1. t The sigmoid activation function is a non-linear function that maps any real number to the interval between 0 and 1, and the forgetting coefficient vector f t Each element in the LSTM corresponds to the degree of information retention at a specific location within the cell state. A value closer to 1 indicates that historical information at that location should be retained more, while a value closer to 0 indicates that historical information at that location should be forgotten more. The forgetting gate achieves soft selection of historical information through the output characteristics of the sigmoid function, avoiding information gaps caused by hard thresholding. This makes information retention and forgetting a differentiable process, allowing for optimization using gradient descent. In the scenario of emulsion pump station flow prediction, the forgetting gate can identify and filter out short-term random fluctuations caused by sensor sampling noise or occasional disturbances, while retaining long-term trend information reflecting the inherent operating rules of the supply system. This selective forgetting mechanism enables the LSTM trend prediction branch to extract stable periodic flow change patterns from the multi-dimensional spatiotemporal feature matrix.

[0075] The function of the input gate is to determine how much of the input information at the current moment should be written into the cell state. The input gate consists of two parallel computational paths: one path outputs the input coefficient vector i through the sigmoid activation function. t One path controls the intensity of new information writing; the other path outputs the candidate cell state vector C̃ through the tanh activation function. t This is used to generate new information to be written. The tanh activation function is a nonlinear function that maps any real number to the interval -1 to 1, and its output range covers both positive and negative directions, allowing candidate cell states to enhance or suppress information expression. The input coefficient vector i t With the candidate cell state vector C̃ t The element-wise multiplication result is the increment of new information actually written to the cell state at the current moment. The cell state is updated by multiplying the output of the forget gate with the cell state C from the previous moment. (t-1)The element-wise multiplication result is added to the output of the input gate to complete the update of the cell state C. t It integrates filtered historical information and filtered new information. Cell state, as the core memory carrier of LSTM unit, transmits information in an approximately linear manner over time. This linear transmission characteristic allows gradients to be stably backpropagated over long time spans without exponential decay or explosion, thus giving LSTM the ability to process long sequence dependencies.

[0076] The output gate determines the hidden state at the current time step. It outputs the coefficient vector o through a sigmoid activation function. t The updated cell state C t After processing with the tanh activation function, it is compared with the output coefficient vector o. t Multiply each element to obtain the hidden state h at the current time step. t Hidden state h t The hidden state serves both as the output of the current LSTM unit, passed to the next network layer or output layer, and as one of the inputs to the next LSTM unit, participating in the gating calculation of the next time step. The output gate selectively reads the cell state, ensuring that the hidden state contains only information relevant to the current prediction task, avoiding interference from irrelevant information. The output layer of the LSTM trend prediction branch receives the hidden state from the last time step and maps it to a scalar value through a fully connected layer; this scalar value is the trend prediction value Q. LSTM This characterizes the flow rate prediction of the liquid supply system at the next moment under the assumption of steady operating conditions.

[0077] The input to the LSTM trend prediction branch is the multidimensional spatiotemporal feature matrix generated in step S14. This matrix is ​​input to the LSTM unit row by row in chronological order, with each row corresponding to the fluid state dimension data and the working condition semantic dimension data at a historical moment. Since the number of rows in the multidimensional spatiotemporal feature matrix equals the sliding time window length N, the LSTM trend prediction branch needs to process the sequence data of N time steps. The input dimension of each time step equals the number of columns in the multidimensional spatiotemporal feature matrix. The LSTM trend prediction branch accumulates and extracts temporal patterns over N time steps through a gating mechanism. The hidden state of the last time step condenses the temporal evolution information within the entire sliding time window. Considering the periodic characteristics of emulsion pump station flow rate changes, the LSTM trend prediction branch uses historical data from the previous complete production cycle as training samples during the training phase, enabling the model to learn the flow rate change patterns at different production stages. The training of the LSTM trend prediction branch uses the mean squared error loss function to measure the deviation between the predicted and actual values, and employs gradient descent and its variants to iteratively optimize the network parameters. The learning rate is set to η. LSTM Batch size set to B LSTMThe training continues until the loss function value on the validation set no longer decreases significantly. For example, the learning rate can be set to 10. -4 The quantity and batch size can be set to 32.

[0078] Step S22: Construct the BP mutation prediction branch, input the multidimensional spatiotemporal feature matrix into the BP mutation prediction branch, use the multi-layer nonlinear mapping capability of the BP mutation prediction branch to analyze the nonlinear relationship between the working condition semantic feature vector and the flow change in the multidimensional spatiotemporal feature matrix, and output the mutation prediction value.

[0079] The BP mutation prediction branch in step S22 uses a backpropagation neural network as its core architecture. A backpropagation neural network is a multi-layer feedforward neural network that optimizes parameters through an error backpropagation algorithm. Its core characteristic lies in its ability to establish a complex mapping relationship between input and output through multiple nonlinear transformations. The network structure of the BP mutation prediction branch consists of three layers: an input layer, hidden layers, and an output layer. The hidden layers are composed of multiple fully connected layers stacked together. Each fully connected layer contains several neurons, which are not connected to each other but are fully connected to all neurons in adjacent layers. The number of hidden layers in the BP mutation prediction branch is set to L. BP The number of neurons in each hidden layer is set to N. BP The number of hidden layers and neurons are also determined using a grid search method. For example, the BP mutation prediction branch can be set as a three-layer hidden layer structure, with the number of neurons in each hidden layer set to 64, 64 and 16 respectively, forming a layer-by-layer compressed feature extraction structure.

[0080] The computation process of a single neuron in the BP mutation prediction branch is as follows: It receives the output values ​​of all neurons in the previous layer, multiplies each output value by its corresponding weight coefficient, sums the results, adds a bias term, and then applies a nonlinear transformation through an activation function to obtain the neuron's output value. The weight coefficients and bias term are learnable parameters of the neural network, continuously adjusted during training using gradient descent to make the network output approximate the true label. The activation function introduces nonlinearity, breaking the limitation that multiple layers of linear transformations are still equivalent to a single-layer linear transformation, allowing the network to approximate arbitrarily complex nonlinear functions. The BP mutation prediction branch can use either the ReLU or Tanh activation function. The ReLU function keeps the input value unchanged when the input is positive and outputs zero when the input is negative, offering advantages such as computational simplicity and mitigating gradient vanishing. The Tanh function maps the input to the interval between -1 and 1, offering the advantage of a centrally symmetric output. The output layer of the BP mutation prediction branch uses linear activation, directly outputting a scalar value as the mutation prediction value Q. BP , which represents the incremental change in transient flow rate caused by the movement of the hydraulic support.

[0081] The input to the BP mutation prediction branch is also the multidimensional spatiotemporal feature matrix generated in step S14. However, unlike the temporal processing method of the LSTM trend prediction branch, the BP mutation prediction branch flattens the multidimensional spatiotemporal feature matrix into a one-dimensional vector before using it as the input to the input layer. The flattening operation arranges all elements of the multidimensional spatiotemporal feature matrix in row-major order into a vector of length N multiplied by the number of columns.

[0082] For example, if N is 30 and the number of columns is 7, the flattened vector length is 210, which includes historical pressure data, historical flow data, one-hot encoding of action type, and normalized action duration for 30 time points. This processing method allows the BP mutation prediction branch to treat the multidimensional spatiotemporal feature matrix as a whole feature set and directly establish the mapping relationship between the feature set and flow mutations through multi-layer nonlinear transformations. Although the BP mutation prediction branch receives the complete multidimensional spatiotemporal feature matrix as input, it automatically learns during training to prioritize the weight allocation of the working condition semantic dimension data in the multidimensional spatiotemporal feature matrix, because the one-hot encoding vector of action type and the action duration feature directly reflect the hydraulic support action, which is the root cause of flow mutations. When a one-hot encoding vector of a certain action type appears in the multidimensional spatiotemporal feature matrix, the BP mutation prediction branch can learn the correspondence between the action type and the magnitude of flow change through training, thereby outputting a mutation prediction value that matches the action.

[0083] The training of the BP mutation prediction branch also employs the mean squared error loss function and gradient descent, but the training samples are constructed with an emphasis on data segments containing the moments when the action occurs. During the forward propagation phase, data is passed from the input layer through each hidden layer to the output layer. The output of each layer is obtained by linear transformation and nonlinear activation of the previous layer's output. During the backpropagation phase, the prediction error of the output layer is used to calculate the gradient of each layer's parameters layer by layer using a chain rule. The gradient information indicates the direction and magnitude of parameter adjustment. During the parameter update phase, the parameters of each layer are adjusted by a step size along the negative direction of the gradient, with the step size controlled by the learning rate. The learning rate of the BP mutation prediction branch is set to η. BP Batch size set to B BP For example, the learning rate can be set to 2.5 × 10⁻⁶. -4 The batch size can be set to 64. Because the BP mutation prediction branch has a relatively simple network structure and no recurrent connections, its training and inference speeds are faster than the LSTM trend prediction branch, meeting the real-time requirements for rapid response to transient mutations.

[0084] The LSTM trend prediction branch in step S21 and the BP mutation prediction branch in step S22 complement each other in terms of network structure and processing mechanism: the LSTM trend prediction branch accumulates information in the time dimension through a gating mechanism, excelling at capturing the periodic patterns and gradual trends inherent in fluid state data; however, the information filtering characteristics of the gating mechanism cause a lag in its response to sudden signals. The BP mutation prediction branch directly establishes an input-output mapping through multi-layer nonlinear transformations, excelling at analyzing the immediate correlation between working condition semantic features and flow rate changes; however, its lack of historical sequence memory prevents it from utilizing temporal evolution information. Both branches receive the same multi-dimensional spatiotemporal feature matrix as input, but extract information from different perspectives: the LSTM trend prediction branch focuses on the temporal patterns in the fluid state dimension, while the BP mutation prediction branch focuses on the causal relationships in the working condition semantic dimension. The parallel prediction mechanism of the two branches ensures that the generation of the total predicted flow demand value does not depend on the prediction capability of a single model, but rather achieves complementary capabilities by comprehensively utilizing the advantages of both models.

[0085] Step S23: Construct a BP fusion network, input the trend prediction value and the mutation prediction value into the BP fusion network, and output the total predicted flow demand value for the next time step.

[0086] The BP fusion network in step S23 uses a lightweight backpropagation neural network as its core architecture. Its function is to convert the trend prediction value Q output by the LSTM trend prediction branch into a fusion network. LSTM The mutation prediction value Q output by the BP mutation prediction branch BP Perform intelligent weighted fusion to output the total predicted traffic demand value Q for the next time step. pred The input to the BP fusion network is the trend prediction value Q. LSTM and mutation prediction value Q BP The resulting two-dimensional vector outputs the total predicted flow demand value Q. pred The number of hidden layers in the BP fusion network is set to L. fusion The number of neurons in each hidden layer is set to N. fusion Since the input dimension of the fusion network is small and the fusion task is relatively simple, the number of hidden layers and neurons can be set to small values ​​to reduce computational complexity. For example, the BP fusion network can be configured with a two-layer hidden layer structure, with the number of neurons in each hidden layer set to 16 and 8 respectively, and the learning rate set to 10. -5 The quantity and batch size can be set to 32.

[0087] The BP fusion network learns the mapping relationship between two predicted values ​​(trend prediction and abrupt change prediction) and the actual traffic value during the training phase. Based on the combined characteristics of the two predictions, it dynamically adjusts the weight allocation of the trend prediction and abrupt change prediction. The weight learning mechanism of the BP fusion network enables it to dynamically adjust the trend prediction Q.LSTM With mutation prediction value Q BP In the total predicted flow demand value Q pred The proportion of [something]. During the training phase, the BP fusion network uses the outputs of the LSTM trend prediction branch and the BP mutation prediction branch as inputs, and the actual flow value as the label. It optimizes the network parameters through error backpropagation, enabling the network to automatically learn the weight ratio that should be assigned to the two predicted values ​​under different input combinations. When the trend prediction value Q... LSTM The mutation prediction value Q is closer to the true value. BP When the value deviates from the true value, the BP fusion network tends to increase the trend prediction value Q. LSTM The weights; when the mutation prediction value Q BP The value Q is closer to the true value and is a trend prediction. LSTM When deviating from the true value, the BP fusion network tends to increase the mutation prediction value Q. BP The weighting. This dynamic weighting mechanism makes the total predicted flow demand value Q... pred It can adaptively rely more on trend prediction under stable operating conditions and more on mutation prediction under sudden operating conditions, without the need to manually set fixed weight coefficients.

[0088] Training the BP fusion network should be performed after the LSTM trend prediction branch and the BP mutation prediction branch have been trained. The training process is as follows: input the training dataset into the trained LSTM trend prediction branch and BP mutation prediction branch to obtain the prediction output of each branch for each sample; pair the prediction outputs of the two branches to form the input samples of the BP fusion network, using the corresponding real flow values ​​as labels; train the BP fusion network using the mean squared error loss function and gradient descent method, iteratively optimizing the parameters of the fusion network until convergence. The output layer of the BP fusion network uses linear activation, and the output total predicted flow demand value Q is... pred This is a continuous scalar value, its physical meaning being the predicted flow rate that the emulsion pumping station should provide at the next moment. Total predicted flow rate demand value Q pred This will be used as input to step S30 to generate a feedforward base target frequency to enable an early response to changes in flow.

[0089] The overall training of the BP-LSTM combined prediction model in step S20 can be conducted using either end-to-end training or phased training. End-to-end training treats the LSTM trend prediction branch, the BP mutation prediction branch, and the BP fusion network as a single network for joint optimization. Gradients are fed back from the output of the fusion network to the inputs of the two branches, allowing the parameters of the three modules to be coordinated and adjusted under the guidance of the same loss function. Phased training first trains the LSTM trend prediction branch and the BP mutation prediction branch independently. Once the prediction performance of these two branches stabilizes, the BP fusion network is then trained. This approach reduces training difficulty and facilitates independent evaluation of the performance of each module. Regardless of the training method used, the training data must come from actual production data from the fully mechanized mining face, including fluid state data and hydraulic support action data under various working conditions to ensure the model's generalization ability across different working conditions. After model training, the prediction performance needs to be evaluated on an independent test dataset. Evaluation metrics can include commonly used regression prediction metrics in this field, such as the coefficient of determination, root mean square error, and mean absolute percentage error.

[0090] In step S20, the parallel architecture of the LSTM trend prediction branch and the BP mutation prediction branch allows for the division of labor and optimization of the model's response capability to different types of working condition changes. When the longwall mining face is in a stable production state, the hydraulic support operates at a low frequency, and the operation type is mainly low-fluid-consumption, the flow rate change mainly exhibits a stable periodic fluctuation. At this time, the long-period pattern extracted by the LSTM trend prediction branch based on the gating mechanism can accurately predict the gradual trend of flow rate change, and the BP fusion network automatically increases the trend prediction value Q. LSTM The weights make the total predicted flow demand value Q pred It relies more heavily on trend prediction results. When the longwall face enters the rapid advancement stage and the hydraulic supports perform high-fluid-consumption actions such as group support shifting or simultaneous lifting of multiple supports, the flow rate changes exhibit a step-like abrupt change. At this time, the BP abrupt change prediction branch, based on the action-flow correlation of multi-layer nonlinear mapping analysis, can quickly predict the magnitude of the flow rate abrupt change, and the BP fusion network automatically increases the abrupt change prediction value Q. BP The weights make the total predicted flow demand value Q pred It relies more on mutation prediction results. This dynamic weight allocation mechanism enables the BP-LSTM combined prediction model to adaptively cope with complex and ever-changing downhole operating conditions without the need for manual switching of prediction strategies or adjustment of model parameters. Step S20 improves the prediction accuracy of the emulsion pump station flow prediction system compared to a single model by constructing the BP-LSTM combined prediction model and realizing the intelligent fusion of trend prediction and mutation prediction. The gating mechanism of the LSTM trend prediction branch enables it to extract long-term fluid supply patterns from the time dimension of the multi-dimensional spatiotemporal feature matrix, avoiding interference from short-term random fluctuations on trend prediction, and outputting the trend prediction value Q. LSTMIt can reflect the baseline flow rate level of the liquid supply system under stable operating conditions. The multi-layer nonlinear mapping capability of the BP mutation prediction branch enables it to analyze the nonlinear correlation between action type and flow rate changes from the operating condition semantic dimension of the multi-dimensional spatiotemporal feature matrix, outputting the mutation prediction value Q. BP It can quickly respond to changes in flow demand indicated by hydraulic support operation commands. The dynamic weight learning capability of the BP fusion network enables it to adaptively allocate trend prediction values ​​Q based on current operating conditions. LSTM and mutation prediction value Q BP The fusion weights output the total predicted traffic demand value Q. pred It comprehensively utilizes the advantages of the two prediction branches while avoiding their respective limitations.

[0091] The LSTM trend prediction branch employs a gating mechanism to process time-series data. The forget gate filters irrelevant information, the input gate selects valid information, and the output gate selects the output content. This triple gating mechanism enables the LSTM trend prediction branch to maintain effective information transmission over long time spans, thereby capturing the inherent periodic pulsations and gradual trend characteristics of the fluid supply system, and ensuring the predicted trend value Q... LSTM The stability and reliability are guaranteed. The BP mutation prediction branch uses a multi-layer fully connected structure to process multi-dimensional features. Each hidden layer introduces nonlinear transformations through activation functions, enabling the network to fit the complex nonlinear mapping relationship between action type and traffic change. When a new action instruction appears in the multi-dimensional spatiotemporal feature matrix, the BP mutation prediction branch can immediately output the corresponding traffic change prediction value through the learned mapping relationship, making the mutation prediction value Q... BP The response speed is guaranteed. The BP fusion network adopts a data-driven weight learning mechanism. By optimizing on training data containing multiple operating conditions, the network automatically learns the confidence differences between the outputs of the two prediction branches under different operating conditions. This allows for dynamic adjustment of the fusion weights during the inference phase, ensuring that the total predicted traffic demand value Q is within acceptable limits. pred High prediction accuracy is achieved under various operating conditions. The parallel operation of the two prediction branches in step S20, without interference, ensures that prediction errors from a single branch are not directly transmitted to the final output, but are suppressed through the weight adjustment mechanism of the BP fusion network. When the LSTM trend prediction branch experiences a large prediction lag error due to a sudden change in operating conditions, the BP sudden change prediction branch simultaneously outputs the sudden change prediction value Q. BP The BP fusion network can provide more timely predictive references by reducing the trend prediction value Q. LSTM The weighting is used to mitigate the impact of lag error on the total predicted flow demand value Q. pred The impact of this. When the BP mutation prediction branch produces a large prediction bias due to missing or abnormal semantic features of the operating condition, the trend prediction value Q output by the LSTM trend prediction branch will also be affected. LSTMThe BP fusion network can provide a more robust predictive reference by reducing the mutation prediction value Q. BP The weighting is used to mitigate the bias on the total predicted flow demand value Q. pred The redundant and complementary prediction architecture enhances the robustness of the BP-LSTM combined prediction model to abnormal operating conditions and sensor failures, enabling the emulsion pump station flow prediction system to output reliable prediction results even in complex downhole electromagnetic environments and variable production conditions. Without the combined prediction mechanism in step S20, the feedforward control in subsequent step S30 would lack reliable flow prediction basis. The control system might over-adjust or under-adjust due to excessive deviation between predicted and actual values, leading to significant fluctuations in the fluid supply pressure. This would negatively impact the smoothness of the hydraulic support operation and the production efficiency of the working face.

[0092] Step S30: Compensate for pressure loss on the total predicted flow demand value, inversely map the compensated total predicted flow demand value to the feedforward base target frequency, determine the feedforward advance based on the pipeline transmission lag time τ, read the real-time pump station outlet pressure value in the fluid state data, generate fuzzy deviation state quantity based on the real-time pump station outlet pressure value, solve the feedback compensation correction frequency through fuzzy PID rules, and superimpose the feedforward base target frequency and the feedback compensation correction frequency to generate the final control frequency to drive the emulsion pump.

[0093] Specifically, the emulsion pump station's supply control system faces two conflicting performance requirements: on the one hand, it needs to respond quickly to changes in flow demand caused by hydraulic support movements to avoid transient pressure collapse at the end of the supply pipeline; on the other hand, it needs to accurately eliminate pressure deviations caused by various disturbances to maintain a constant pressure supply state. Traditional pure feedback control strategies inherently exhibit response lag when flow demand changes abruptly. This lag stems from the transmission delay of the emulsion in the pipeline and the mechanical rotational inertia of the motor and pump. The control system must wait for the pressure sensor to detect the pressure deviation before initiating adjustment. The entire process from the generation of the pressure deviation to the inverter's response and then to the change in motor speed involves an incompressible physical lag time. While pure feedforward control strategies can eliminate response lag by adjusting the speed in advance based on prediction, the total predicted flow demand value Q output by the prediction model... pred There is always a certain prediction error. Furthermore, the presence of unpredictable disturbances in the hydraulic system, such as pipeline leakage and internal valve leakage, means that relying solely on feedforward control cannot guarantee that the pressure will remain stable near the target value. Step S30 combines feedforward control with fuzzy PID feedback control to construct a dual closed-loop control architecture of feedforward coarse adjustment and feedback fine adjustment: the feedforward control is based on the total predicted flow demand value Q output in step S20. predThe frequency of the inverter is adjusted in advance to speed up the system response. Fuzzy PID feedback control dynamically tunes the PID parameters based on the real-time pressure deviation to eliminate residual errors and suppress various disturbances. The synergistic effect of the two control strategies enables the emulsion pump station to meet the dual requirements of fast response and precise pressure stabilization.

[0094] Further, see Figure 4 Step S30 includes:

[0095] Step S31: Compensate for pressure loss on the total predicted flow demand value to obtain the compensated total predicted flow demand value; establish a flow-frequency characteristic relationship model for the emulsion pump, input the compensated total predicted flow demand value into the flow-frequency characteristic relationship model for inverse mapping, determine the feedforward advance based on the pipeline transmission lag time, output the feedforward basic target frequency and send it to the frequency converter for pre-adjustment.

[0096] Step S31 establishes a flow-frequency characteristic model of the emulsion pump and performs inverse mapping to generate a feedforward target frequency. The flow-frequency characteristic model describes the correspondence between the flow output capacity of the emulsion pump and the drive frequency at different inverter output frequencies. This model is based on the inherent mechanical characteristics of the emulsion pump and the speed-frequency conversion relationship of the motor. As a positive displacement reciprocating pump, the theoretical displacement of the emulsion pump is proportional to the crankshaft speed, which is determined by the speed of the drive motor. The inverter adjusts the motor speed by changing the motor power supply frequency, thereby controlling the pump's output flow. Under constant pressure supply conditions, the pump station outlet pressure is determined by the balance between the flow consumed by the hydraulic support operation and the pump station output flow. When the flow demand increases while the pump station output flow remains constant, the pressure decreases; when the flow demand decreases while the pump station output flow remains constant, the pressure increases. The flow-frequency characteristic relationship model can be constructed in two ways: analytical model or data-driven model. The analytical model is derived based on the theoretical displacement formula of the pump and the speed-frequency relationship of the motor, while the data-driven model is obtained by measuring the flow output of the pump station under different frequency setpoints and fitting the data.

[0097] The analytical expression of the flow-frequency characteristic model is derived based on the electromechanical transmission characteristics of the emulsion pump station. The output flow rate Q of the emulsion pump... pump The relationship between Q and the inverter output frequency f can be expressed as Q pump =Vd×n×ηv, where Vd is the theoretical displacement of the pump in a single stroke, n is the crankshaft speed, and ηv is the volumetric efficiency of the pump. The relationship between the crankshaft speed n and the inverter output frequency f is n=60×f / p, where p is the number of pole pairs of the motor. Combining the two equations, we get Q. pump =Vd×(60×f / p)×ηv=(Vd×60×ηv / p)×f, which can be simplified to Q pump =kQ ×f, where k Q k is the flow-to-frequency conversion coefficient. Q =(Vd×60×ηv / p). Flow-frequency conversion coefficient k Q The determination of the flow rate needs to be combined with the pump's technical parameters and actual calibration: the theoretical single-stroke displacement Vd is obtained from the pump's nameplate or technical manual, the number of pole pairs p is obtained from the motor nameplate, and the volumetric efficiency ηv needs to be obtained through actual testing by measuring the ratio of the actual flow rate to the theoretical flow rate under rated operating conditions. Because of frictional resistance and local resistance along the pipeline during the transport of the emulsion, pipeline flow resistance will cause a certain pressure loss. To maintain the target pressure at the end of the working face, the pump station needs to output an additional flow to overcome the pipeline pressure loss. To compensate for the pressure loss caused by pipeline flow resistance, flow resistance compensation is needed for the total predicted flow demand value, Q. comp =Q pred ×(1+k pipe ), where k pipe This is the pipe flow resistance compensation coefficient, which is related to the pipe length, pipe inner diameter, and pipe friction coefficient collected in step S11, k. pipe The value of k typically ranges from 0.05 to 0.15, with the specific value determined through on-site calibration based on the pipe length and flow resistance characteristics. The longer the pipe, the smaller the inner diameter, and the greater the friction coefficient, the higher the value of k. pipe The larger the value, the better.

[0098] Reverse mapping refers to calculating the required inverter frequency in reverse based on the known flow demand. The reverse mapping formula is f. base =Q comp / k Q , where f base The target frequency for feedforward is defined as follows. Furthermore, due to the lag time τ in pipeline transmission, feedforward control needs to issue adjustment commands τ time in advance so that changes in the pump station's output flow rate accurately reach the end of the working face at the moment the hydraulic support operates. The feedforward advance is achieved as follows: Step S20 predicts the future flow demand; combined with the pipeline transmission lag time τ, the feedforward control command issued at the current time t corresponds to the predicted flow demand at time t+τ, thus achieving advance compensation for the pipeline transmission lag. The physical meaning of the inverse mapping is: when the predicted flow demand at time τ is Q... pred At this time, the inverter frequency needs to be set to f base This ensures that the pump station outputs a flow rate that matches the demand, allowing the pump station to be adjusted to the expected operating state before the actual flow rate change occurs. At the same time, flow resistance compensation ensures that the end of the working face receives sufficient pressure.

[0099] In step S31, the feedforward fundamental target frequency f baseAfter the calculation is completed, the control system will... base The first-level control input is sent directly to the frequency setting port of the frequency converter. Upon receiving the frequency setting, the frequency converter uses pulse width modulation (PWM) technology to generate an AC voltage corresponding to the set frequency to drive the motor. The motor speed changes accordingly, which in turn changes the crankshaft speed of the emulsion pump, adjusting the pump station's output flow accordingly. The core value of feedforward control lies in its predictive capability: traditional feedback control must wait for a pressure deviation to occur before initiating adjustment. However, the occurrence of a pressure deviation means that the flow supply and demand are already unbalanced, making adjustment too late. Feedforward control, on the other hand, adjusts the speed in advance based on predicted flow demand. When the flow demand change actually occurs, the pump station is already in a ready state, effectively shifting the start time of the control action forward, eliminating the physical lag time window caused by fluid transmission delay and mechanical inertia delay. In particular, by introducing the pipeline transmission lag time τ as a feedforward advance, feedforward control can accurately compensate for the response delay caused by the length of the supply pipeline and flow resistance, ensuring that the execution effect of the control command is precisely matched in time with the hydraulic support's fluid demand. The longer the pipeline and the greater the flow resistance, the greater the feedforward lead, and the earlier the control system's predictive action. This adaptive feedforward mechanism based on the pipeline's physical characteristics allows the method of this invention to adapt to liquid supply pipeline systems of different lengths and structures. The accuracy of the total predicted flow demand value output in step S20 directly determines the rationality of the feedforward base target frequency. The more accurate the prediction, the better the effect of the feedforward control; conversely, if the prediction deviation is too large, the feedforward control may cause over-adjustment or under-adjustment. Due to the objective existence of prediction errors and the influence of unmeasurable disturbances in the system, feedforward control cannot guarantee pressure stability at the target value on its own; it must be combined with subsequent fuzzy PID feedback control to achieve precise constant pressure.

[0100] Step S32: Read the real-time pump station outlet pressure value, calculate the pressure deviation between the real-time pump station outlet pressure value and the preset target working pressure value, obtain the first derivative of the pressure deviation to get the pressure deviation change rate, and perform quantization and fuzzification processing on the pressure deviation and the pressure deviation change rate to map them into fuzzy linguistic variables to generate fuzzy deviation state variables.

[0101] Real-time pump station outlet pressure value P t The pressure sensor acquisition channel in step S11 of the read multiplexing process converts the current pump station outlet pressure into a current signal, which is then input to the control system via analog-to-digital conversion. The target working pressure value P... target This is the set pressure for the emulsion pump station to operate under constant pressure. Its value is determined comprehensively based on the working pressure rating of the hydraulic supports in the longwall mining face and the pressure loss of the pipeline system. It is usually set slightly higher than the rated working pressure of the hydraulic supports to compensate for pipeline pressure drop. The formula for calculating the pressure deviation E is E=P. target -P t , where Pt The pressure deviation E represents the real-time pump station outlet pressure. A positive pressure deviation E indicates that the current pressure is lower than the target pressure, requiring an increase in the liquid supply. A negative pressure deviation E indicates that the current pressure is higher than the target pressure, requiring a decrease in the liquid supply. A zero pressure deviation E indicates that the current pressure is exactly equal to the target pressure, requiring no adjustment. The pressure deviation change rate EC is calculated using the first-order difference method: EC = (E(t) - E(t - Δt)) / Δt, where E(t) is the pressure deviation at the current moment, E(t - Δt) is the pressure deviation at the previous sampling moment, and Δt is the sampling period. The pressure deviation change rate EC reflects the trend of pressure deviation change: a positive EC indicates that the pressure deviation is increasing, i.e., the pressure is deviating from the target value; a negative EC indicates that the pressure deviation is decreasing, i.e., the pressure is approaching the target value. The larger the absolute value of EC, the more drastic the change trend. The pressure deviation E and the pressure deviation change rate EC together constitute the dual input signals of the fuzzy PID controller. The combination of the two can comprehensively characterize the degree and trend of deviation between the current pressure state and the target state.

[0102] Quantization processing converts the continuous pressure deviation E and the rate of change of pressure deviation EC into discrete quantized values ​​for subsequent fuzzification. Quantization factors ke and kec are used for the quantization conversion of pressure deviation E and the rate of change of pressure deviation EC, respectively. The physical meaning of the quantization factors is to map the range of variation of the actual physical quantity to the domain specified by the fuzzy controller. The method for determining the quantization factor ke is as follows: [The text abruptly ends here, so the translation stops.] max Dividing by the positive boundary value ne of the fuzzy universe of discourse, we get ke = E max / ne. The method for determining the quantification factor kec is similar: the maximum allowable range of the pressure deviation change rate EC is calculated. max Dividing by the positive boundary value nec of the fuzzy universe of discourse, we get kec = EC max / nec. The quantized pressure deviation e' = E / ke, the quantized pressure deviation change rate ec = EC / kec, e' is an integer value within the domain of -ne to +ne, and ec is an integer value within the domain of -nec to +nec. Fuzzification mapping maps the quantized values ​​e' and ec to a set of fuzzy linguistic variables, which use semantic labels to represent the state interval of the values.

[0103] The fuzzy linguistic variable sets for pressure deviation E and pressure deviation change rate EC are typically divided into multiple levels, for example, seven levels, as shown in Table 2:

[0104] Table 2. Fuzzy Linguistic Variable Classification of Pressure Deviation and Rate of Change of Pressure Deviation

[0105] Language variable symbols Semantic meaning Corresponding domain interval NB Large burden Negative boundary of the domain to the negative mid-region NM Negative From negative central area to negative small area NS Negative small Negative cell to zero cell boundary ZO zero Near Zone Zero PS Just small Zero zone boundary to positive cell PM middle From the main residential area to the central area PB Zhengda From the central region to the positive boundary of the domain of discourse

[0106] Fuzzy processing uses membership functions to map quantized values ​​to the membership degrees of each fuzzy linguistic variable. Common forms of membership functions include triangular membership functions and Gaussian membership functions. The triangular membership function is determined by three parameters: the left boundary *a*, the peak point *b*, and the right boundary *c*. When the input value *x* is between *a* and *c*, the membership degree continuously varies between 0 and 1, reaching 1 at the peak point *b*. The output of fuzzy processing is the fuzzy linguistic variable to which the pressure deviation *E* belongs and its membership degree, and the fuzzy linguistic variable to which the pressure deviation change rate *EC* belongs and its membership degree. Together, these constitute the fuzzy deviation state variable. The use of fuzzy processing allows the control system to focus on the degree range of the deviation rather than the precise numerical value of the pressure deviation. This qualitative approach is similar to the thinking patterns of human engineers, enhancing the robustness of the control system to sensor measurement noise and transient disturbances.

[0107] Step S33: Perform fuzzy inference on the fuzzy deviation state quantity according to the preset fuzzy inference rule table, dynamically calculate the proportional parameter correction, integral parameter correction and derivative parameter correction of the PID controller, update the preset PID parameters of the PID controller in real time according to the proportional parameter correction, integral parameter correction and derivative parameter correction, and use the updated PID parameters to calculate the pressure deviation and output feedback compensation correction frequency.

[0108] Step S33 performs fuzzy inference on the fuzzy deviation state quantity according to the preset fuzzy inference rule table, dynamically calculates the parameter correction of the PID controller and updates the PID parameters in real time, and calculates the feedback compensation correction frequency using the updated PID parameters. The fuzzy inference rule table is the core knowledge base of the fuzzy controller, which stores the PID parameter correction that should be output under various combinations of the fuzzy states of pressure deviation E and the fuzzy states of pressure deviation change rate EC. The design of the fuzzy inference rule table follows expert experience and the dynamic characteristics of the control system;

[0109] For example, when the pressure deviation E is negative and the deviation change rate EC is negative, it indicates that the pressure has dropped sharply and is still decreasing rapidly. In this case, the proportional parameter Kp needs to be increased significantly to achieve strong compensation. When the pressure deviation E is close to zero and the deviation change rate EC is also close to zero, it indicates that the pressure has stabilized. In this case, the proportional parameter Kp needs to be reduced and the integral parameter Kw needs to be increased appropriately to eliminate static error. When the pressure deviation E is positive but the deviation change rate EC is negative, it indicates that the pressure is higher than the target value but is decreasing towards the target. In this case, it is not advisable to over-adjust to avoid overshoot.

[0110] The fuzzy inference rule table is organized in two-dimensional matrix form. The row index is the fuzzy linguistic variable of pressure deviation E, the column index is the fuzzy linguistic variable of pressure deviation change rate EC, and the matrix elements are the PID parameter corrections under the corresponding conditions. Since the PID controller includes three adjustable parameters: proportional parameter Kp, integral parameter Kw, and derivative parameter Kd, the fuzzy inference rule table actually contains three two-dimensional matrices corresponding to the correction rules for the three parameters. For example, the fuzzy inference rule table for the proportional parameter correction ΔKp is shown in Table 3:

[0111] Table 3. Fuzzy Inference Rules for Proportional Parameter Correction ΔKp

[0112]

[0113] The execution process of fuzzy inference is as follows: Based on the fuzzy linguistic variables of pressure deviation E and pressure deviation change rate EC output in step S32, the corresponding rule activation conditions are searched in the fuzzy inference rule table. The activation intensity of each rule is calculated using the fuzzy inference algorithm. The outputs of all activated rules are weighted and synthesized to obtain the fuzzy output. Then, the fuzzy output is converted into a precise numerical output through defuzzification. Common defuzzification methods include the centroid method and the maximum membership method. The centroid method calculates the weighted average of the outputs of all activated rules as the final output, while the maximum membership method selects the rule output with the highest membership degree as the final output. After defuzzification, the precise values ​​of the proportional parameter correction ΔKp, integral parameter correction ΔKw, and differential parameter correction ΔKd are obtained.

[0114] The PID parameters are updated in real time using an incremental superposition method. The updated proportional parameter Kp equals the preset proportional parameter Kp0 plus the proportional parameter correction ΔKp; the updated integral parameter Kw equals the preset integral parameter Kw0 plus the integral parameter correction ΔKw; and the updated derivative parameter Kd equals the preset derivative parameter Kd0 plus the derivative parameter correction ΔKd. The preset PID parameters Kp0, Kw0, and Kd0 are the reference parameters for the fuzzy PID controller, and their values ​​are determined using the classic PID tuning method based on the dynamic characteristics of the emulsion pump station pressure control system. The emulsion pump station pressure control system can be approximated as a second-order inertial plus pure time-delay system, with the transfer function being... Where e is the natural constant, and the transfer function G(s) is a mathematical model in control system theory that describes the input-output relationship of a linear time-invariant system. It is defined as the ratio of the Laplace transform of the system output to the Laplace transform of the system input. In this embodiment, G(s) represents the dynamic mapping relationship between the inverter frequency setpoint as input and the pump station outlet pressure value as output. In the formula, s is the complex frequency domain variable in the Laplace transform, k1 is the liquid supply system gain, which represents the change in pump station output flow caused by a unit frequency change. It is obtained by measuring the output flow corresponding to different frequency setpoints under steady-state conditions and calculating the slope through linear fitting; k2 is the motor gain, which represents the proportional relationship between the inverter output frequency and the actual motor speed. It is obtained by reading the ratio of rated speed to rated frequency from the motor nameplate; k3 is the sensor gain, which represents the sensitivity coefficient of the pressure sensor. It is obtained by reading the ratio of the range to the output signal range from the sensor technical manual; T1 is the inertial time constant of the liquid supply system, which represents the pump station's... The characteristic quantity of the time required for the output flow rate to transition from one steady state to another is obtained by measuring the flow response curve under a step input and calculating the time taken for the flow rate to reach 63.2% of the steady state value; T2 is the motor inertia time constant, which represents the characteristic quantity of the time required for the motor speed to transition from one steady state to another. It is obtained by calculating based on the ratio of the motor rotor moment of inertia to the electromagnetic torque constant or by measuring the step response; τ is the pipeline transmission lag time, which is consistent with the pipeline transmission lag time calculated based on the pipeline state parameters in step S11, reflecting the direct influence of pipeline length and flow resistance on the dynamic response of the system.

[0115] The transfer function model adopts a structure of second-order inertial elements in series with a pure time-delay element. It is based on the following physical mechanism: the dynamic response process of the emulsion pump station pressure control system includes two relatively independent inertial elements and one pure time-delay element. The first inertial element (T1×s+1) describes the electromechanical coupling dynamic process of the motor speed change driving the pump output flow change after the frequency converter adjusts the frequency. The second inertial element (T2×s+1) describes the fluid dynamic process of the fluid kinetic energy and pressure potential energy in the pipeline system mutually converting to reach a new equilibrium after the flow change. The pure time-delay element e -τ×sThis model describes the incompressible time delay generated by the transmission of emulsions in pipelines. By separating the dynamic links corresponding to different physical mechanisms, it can accurately describe the complete dynamic process of an emulsion pump station from frequency regulation to pressure response, providing a precise system characteristic description for PID parameter tuning. This transfer function model provides a theoretical basis for tuning preset PID parameters. By analyzing the pole distribution and frequency characteristics of the transfer function, the initial range of PID parameters that ensures system stability and good dynamic performance can be determined. It also provides a reference for designing fuzzy PID parameter corrections. When changes in system operating conditions cause changes in the inertial time constant T1 or lag time τ, fuzzy inference rules can adjust the PID parameters accordingly to maintain control performance. Preset PID parameters can be determined using classic tuning methods such as the Ziegler-Nichols method or the Cohen-Kuhn method, based on the system's step response characteristics.

[0116] Feedback compensation correction frequency f comp The calculation uses either positional PID control or incremental PID control. The expression for positional PID control is: The integral of E is approximated by the rectangular integral method as the sum of the pressure deviations at all historical moments multiplied by the sampling period. The derivative of E is approximated by the backward difference method as the difference between the pressure deviation at the current moment and the pressure deviation at the previous moment divided by the sampling period. The design of the positional PID control law is based on the error-driven principle in control theory: the proportional term Kp×E is proportional to the current pressure deviation; the larger the pressure deviation, the larger the output frequency correction. This term is used to quickly respond to the pressure deviation and make the system move closer to the target pressure in time, but proportional control alone cannot completely eliminate steady-state error. The integral term Kw×∫Edt accumulates the historical pressure deviation. As long as the pressure deviation exists, the integral term will continue to increase until the deviation is eliminated to zero. This term is used to eliminate the steady-state error of the system and achieve error-free regulation, but the integral action will reduce the system response speed and may cause overshoot. The derivative term Kd×dE / dt reflects the changing trend of the pressure deviation. When the pressure deviation changes rapidly, the derivative term outputs a large correction to suppress the further development of the deviation in advance. This term is used to improve the dynamic performance of the system, reduce the overshoot, and accelerate the convergence speed of the regulation process. The three control actions work together: the proportional term provides basic adjustment force, the integral term ensures steady-state accuracy, and the derivative term optimizes the dynamic process, together achieving rapid and accurate elimination of pressure deviation.

[0117] The expression for the incremental PID control law is: The current feedback compensation correction frequency f comp (t) = the value f at the previous moment comp (t-Δt)+Δf compThe incremental PID control law is designed based on the differential transformation of the positional PID control law: subtracting the positional PID outputs from two adjacent moments yields the incremental expression, where Kp×[E(t)-E(t-Δt)] corresponds to the proportional adjustment of the pressure deviation increment, Kw×E(t) corresponds to the integral adjustment of the current pressure deviation, and Kd×[E(t)-2×E(t-Δt)+E(t-2Δt)] corresponds to the derivative adjustment of the pressure deviation rate of change increment. The incremental PID control law has three advantages over the positional PID control law: First, the incremental output is the increment of the control quantity rather than its absolute value. Even if a deviation occurs in the calculation at a certain moment or a sensor malfunctions momentarily, it only affects the increment at that moment, without causing a large jump in the control quantity, resulting in higher system safety. Second, the incremental control law does not require calculating the complete historical integral of the deviation, avoiding integral saturation problems. When the system is in a saturated state for a long time, it can quickly exit saturation and resume normal regulation. Third, the incremental control law facilitates a seamless switch from manual to automatic operation; during the switch, simply setting the increment to zero maintains continuous control. In emulsion pump station pressure control applications, the incremental PID control law functions as follows: when pressure fluctuates suddenly, the frequency increment of the incremental output can be smoothly superimposed on the current frequency to achieve gradual adjustment. This avoids the sudden changes in control quantity caused by excessive accumulation of the integral term in position control, making the inverter and motor run more smoothly and extending the service life of the equipment. Feedback compensation corrects the frequency f. comp The physical meaning is: the frequency correction amount calculated based on the current pressure deviation state and PID control strategy, which is used to eliminate residual errors that the feedforward control could not fully compensate for, as well as various disturbances generated during system operation.

[0118] The fuzzy PID control strategy in step S33 exhibits significant adaptive capability compared to the traditional fixed-parameter PID control strategy. Traditional PID controllers maintain fixed parameters throughout operation, and their parameter values ​​can only be optimized for a specific operating condition. When operating conditions change, fixed parameters often struggle to simultaneously meet the requirements of rapid response and avoidance of overshoot. The operating conditions of emulsion pump stations vary widely: when the hydraulic supports are not in operation, the pump station only needs to maintain low flow and pressure; when a few supports are in operation, a medium flow rate is required; and when moving or rapidly advancing a group of supports, a large flow rate is needed. The dynamic characteristics of the pressure control system differ significantly under different operating conditions, making it difficult for fixed-parameter PID controllers to adapt to such a broad range of operating conditions. The fuzzy PID controller achieves online self-tuning of PID parameters through fuzzy inference rules, dynamically adjusting parameters based on the current pressure deviation: increasing the proportional parameter to accelerate response when the deviation is large; decreasing the proportional parameter and increasing the integral parameter to eliminate static error and avoid overshoot when the deviation approaches zero; and increasing the derivative parameter to suppress oscillations when the deviation changes drastically. This parameter self-tuning mechanism enables the fuzzy PID controller to maintain good control performance under different operating conditions, with faster response speed and smaller overshoot.

[0119] Step S34: The feedforward base target frequency and the feedback compensation correction frequency are superimposed and summed to obtain the final control frequency. The final control frequency is subjected to amplitude limiting protection processing. The final control frequency after amplitude limiting protection processing is output to the variable frequency drive of the emulsion pump to control the motor speed to complete the liquid supply control of the current cycle. Then, return to step S11 to enter the next control cycle.

[0120] Step S34 will feed forward the base target frequency f base With feedback compensation correction frequency f comp The final control frequency f is obtained by superimposing and summing the results. final After performing amplitude limiting protection, the output is sent to the frequency converter driver. The final control frequency f final The calculation formula is f final =f base +f comp This formula reflects the synergistic effect of feedforward control and feedback control: the feedforward base target frequency f base Based on the predicted flow demand, a predictive coarse adjustment amount is provided for the control system, enabling the pumping station to be adjusted to near the target operating state before changes in flow demand occur; feedback compensation correction frequency f comp The pressure deviation, calculated based on real-time measurements, provides precise fine-tuning for the control system, eliminating prediction errors and residual deviations caused by external disturbances. The superposition of the two frequency components results in the final control frequency f. final It possesses both the fast response characteristics of feedforward control and the precise stability characteristics of feedback control, achieving a balance between speed and accuracy.

[0121] Limiting protection is applied to the final control frequency f. final Upper and lower limit constraints are applied to ensure the output frequency remains within the inverter's safe operating range. The logic for limiting protection is as follows: if f final Greater than the upper frequency limit f max Then f final Set as f max If f final Less than the lower frequency limit f min Then f final Set as f min Otherwise keep f final Unchanged. Frequency upper limit f max The frequency limit is determined based on the mechanical strength limitations of the motor and pump body. Operating beyond the upper frequency limit will lead to motor overheating and excessive wear of the pump body; the lower frequency limit f min The frequency limit is determined based on the pump's stable operation requirements. Operating below the lower frequency limit will lead to increased pump output pulsation and even cavitation. For example, the upper frequency limit can be set to a multiple of the power frequency, such as 1.2 times the power frequency, and the lower frequency limit can be set to a fraction of the power frequency, such as 0.2 times the power frequency. This frequency limiting protection prevents the control system from outputting abnormal frequencies exceeding the equipment's capacity under extreme operating conditions, ensuring the safe operation of the emulsion pump station.

[0122] The final control frequency f after amplitude limiting protection processing final The frequency is sent to the variable frequency drive (VFD) of the emulsion pump via a communication interface or analog output port. After receiving the frequency setpoint, the VFD generates a three-phase AC voltage corresponding to the frequency using vector control or V / F control algorithms to drive the motor. The motor speed changes with the frequency, and the crankshaft speed of the emulsion pump connected to the motor's output shaft changes accordingly, adjusting the pump station's output flow rate and pressure. When the final control frequency f... final When the frequency is increased, the motor speed increases, the pump station output flow increases, and the pipeline pressure rises. When the final control frequency f final When the flow rate decreases, the motor speed drops, the pump station output flow rate decreases, and the pipeline pressure decreases. Compared with the traditional power frequency operation loading and unloading valve control method, the variable frequency fluid supply method has higher adjustment accuracy and a wider adjustment range, enabling stepless adjustment of the fluid supply flow rate to precisely match the fluid demand of the hydraulic support. After the fluid supply control of the current control cycle is completed, the system returns to step S11 to enter the next control cycle. The newly collected pressure and flow values ​​are again entered into the multi-dimensional spatiotemporal feature matrix construction process, the new total predicted flow demand value is again entered into the feedforward control process, and the new pressure deviation is again entered into the fuzzy PID feedback control process, forming a continuously operating closed-loop control system.

[0123] Step S30 constructs a dual closed-loop control architecture based on feedforward prediction coarse adjustment and fuzzy PID feedback fine adjustment, effectively resolving the technical contradiction between response speed and pressure regulation accuracy in traditional control strategies. This step establishes a flow-frequency characteristic model for the emulsion pump, fully considering the impact of supply pipeline length and flow resistance on system response. Sufficient pressure supply to the working face end is ensured through pipeline flow resistance compensation, and precise compensation for pipeline delay is achieved through feedforward advance based on pipeline transmission lag time. The total predicted flow demand value is inversely mapped to the feedforward base target frequency, allowing the frequency converter to be adjusted in advance before the actual flow demand changes, thereby eliminating physical response lag caused by fluid transmission and mechanical inertia. Simultaneously, the system introduces a fuzzy PID control strategy. By calculating the real-time pressure deviation and its rate of change and performing fuzzification processing, fuzzy inference rules are used to dynamically tune the PID parameters online to output the feedback compensation correction frequency. This not only enhances the system's robustness to sensor noise but also adaptively eliminates prediction errors and residual deviations caused by external disturbances according to changes in operating conditions. By superimposing the feedforward base target frequency and the feedback compensation correction frequency, and then generating the final control frequency after amplitude limiting protection, the emulsion pump is driven. This achieves rapid forward response to sudden changes in working conditions and precise constant pressure maintenance in stable working conditions. This enables the control system to adapt to liquid supply pipelines of different lengths and flow resistance characteristics, significantly improving the production efficiency of the fully mechanized mining face while ensuring the safe operation of the equipment.

[0124] Example 2

[0125] This embodiment, based on Embodiment 1, provides an adaptive liquid supply system for emulsion pumps based on temporal deep learning, such as... Figure 5 As shown, it includes:

[0126] Feature construction module: used to collect fluid state data at the outlet of emulsion pump station and pipeline state parameters of supply pipeline and capture hydraulic support action commands, and construct a multi-dimensional spatiotemporal feature matrix based on fluid state data and hydraulic support action commands; the fluid state data includes at least the real-time pump station outlet pressure value; and calculates the pipeline transmission lag time τ based on pipeline state parameters;

[0127] Traffic forecasting module: Used to construct BP-LSTM combined forecasting model. The multidimensional spatiotemporal feature matrix is ​​input into the LSTM trend forecasting branch and BP mutation forecasting branch of the BP-LSTM combined forecasting model for parallel forecasting. The trend forecast value output by the LSTM trend forecasting branch and the mutation forecast value output by the BP mutation forecasting branch are input into the BP fusion network in the BP-LSTM combined forecasting model for weighted fusion, and the total predicted traffic demand value is output.

[0128] Liquid supply control module: used to compensate for pressure loss of total predicted flow demand value, inversely map the compensated total predicted flow demand value to feedforward base target frequency, determine feedforward advance based on pipeline transmission lag time τ, generate fuzzy deviation state quantity based on real-time pump station outlet pressure value, solve feedback compensation correction frequency through fuzzy PID rules, and superimpose feedforward base target frequency and feedback compensation correction frequency to generate final control frequency to drive emulsion pump.

[0129] Example 3

[0130] Based on Example 1, this embodiment provides experimental verification of the adaptive liquid supply method for emulsion pumps based on temporal deep learning in Example 1. In order to objectively evaluate the performance of the BP-LSTM combined prediction model proposed in this invention, this embodiment selects the actual production data of an emulsion pump station in a fully mechanized mining face that has been running continuously for one month as the dataset, and constructs a SARIMA model, a single BP neural network model, and a single LSTM neural network model as the comparison group, and conducts a comparative experiment with the BP-LSTM combined prediction model described in this invention.

[0131] In this embodiment, the raw acquired data is first processed according to steps S10 to S14 of Embodiment 1. Fluid state data and working condition semantic data are acquired, time-series data are extracted using a sliding time window, high-frequency noise is removed through variational mode decomposition, and hydraulic support action commands are converted into working condition semantic feature vectors. Finally, a multi-dimensional spatiotemporal feature matrix is ​​constructed as the input to each model. To ensure optimal model performance, this embodiment uses a grid search method to optimize the hyperparameters of each neural network model. When constructing the model, the specific structural parameters of each model are determined based on the principle of minimizing the mean absolute percentage error (MAPE) on the validation set.

[0132] The specific branch and fusion network parameter settings for the BP-LSTM combined prediction model described in Example 1 are shown in Table 4:

[0133] Table 4. Parameter settings for branches and fusion networks in the BP-LSTM combined prediction model.

[0134]

[0135] (1) BP mutation prediction branch: The network structure contains 3 hidden layers, with the number of neurons in each hidden layer configured as [64, 64, 16]. During training, the initial learning rate is set to 2.5 × 10⁻⁶. -4 The batch size is set to 64. This branch utilizes multi-layer nonlinear mapping capabilities to analyze the semantic features of operating conditions in the multi-dimensional spatiotemporal feature matrix, capturing transient changes in flow.

[0136] (2) LSTM Trend Prediction Branch: The network structure contains two hidden layers, with the number of neurons in each hidden layer configured as [128, 128]. During training, the initial learning rate is set to 10. -4 The batch size is set to 32. This branch uses a gating mechanism to extract long-period fluid motion patterns from a multidimensional spatiotemporal feature matrix.

[0137] (3) Backpropagation (BP) Fusion Network: This network fuses the trend predictions and mutation predictions from the two branches mentioned above. It contains two hidden layers with the number of neurons in each layer configured as [16, 8]. During training, the initial learning rate is set to 10. -5 The batch size is set to 32.

[0138] To quantitatively evaluate the prediction accuracy of each model, this embodiment selects the coefficient of determination R. 2 As a key performance indicator, this embodiment uses the four pre-trained models (SARIMA, BP, LSTM, and BP-LSTM) to predict the test set data and compares the prediction results with the true values. Please refer to [link / reference]. Figure 6 The scatter plot shows the comparison between the predicted and actual values ​​from different models. The horizontal axis represents the actual flow rate, and the vertical axis represents the predicted flow rate. The red solid line y=x represents the ideal state where the prediction is completely accurate.

[0139] SARIMA model: its coefficient of determination R for prediction results 2 The value is 0.938. As a traditional time series model, although it can capture certain linear trends, its performance is generally poor when dealing with the complex nonlinear conditions of fully mechanized mining faces.

[0140] Single LSTM neural network model: its prediction coefficient R 2 The value is 0.933. The scatter plot shows that its fitting effect is acceptable in the low flow range, but there is a certain deviation near the high flow abrupt change point, which is consistent with its characteristic of being good at memorizing long-term patterns but slow to respond to transient changes.

[0141] Single BP neural network model: its prediction coefficient R 2 The value is 0.952. Thanks to its powerful nonlinear mapping capability, the BP model performs well in handling sudden changes in flow and its accuracy is better than that of the LSTM model.

[0142] BP-LSTM combined prediction model: The coefficient of determination R of the prediction results of the method described in this invention 2 It achieves a score of 0.954, making it the best performing of the four models.

[0143] From the perspective of the scatter distribution characteristics, the predicted values ​​of the BP-LSTM combined prediction model closely surround both sides of the y=x reference line, exhibiting a highly symmetrical distribution characteristic and showing no systematic tendency to overestimate or underestimate. This balanced error distribution characteristic indicates that the BP fusion network described in Example 1 effectively combines the advantages of the LSTM trend prediction branch and the BP mutation prediction branch: under stable operating conditions, the LSTM branch maintains accurate tracking of periodic patterns, while the BP branch enables rapid response to transient demands when the hydraulic support movement causes sudden changes in flow.

[0144] Experimental data show that the BP-LSTM combined prediction model proposed in this invention effectively solves the "dimensional collapse" problem of a single model through the construction of a multi-dimensional spatiotemporal feature matrix and a dual-branch parallel prediction mechanism, significantly improving the accuracy and robustness of emulsion pump station flow prediction. Based on this high-precision total predicted flow demand value, the feedforward control strategy in Example 1 can obtain a more accurate basic target frequency, thereby more effectively eliminating physical lag in actual control and achieving the goals of on-demand liquid supply and constant pressure control.

[0145] The methods and systems of this application may be implemented in many ways. For example, they may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order of steps for the method is for illustrative purposes only, and the steps of the method of this application are not limited to the order specifically described above, unless otherwise specifically stated.

[0146] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0147] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An adaptive liquid supply method for emulsion pumps based on temporal deep learning, characterized in that, The method includes: The system collects fluid state data from the emulsion pump station outlet and pipeline state parameters from the supply pipeline, and captures hydraulic support action commands. A multidimensional spatiotemporal feature matrix is ​​constructed based on the fluid state data and hydraulic support action commands. The fluid state data includes at least the real-time pump station outlet pressure value. The pipeline transmission lag time τ is calculated based on the pipeline state parameters. A BP-LSTM combined prediction model is constructed. The multidimensional spatiotemporal feature matrix is ​​input into the LSTM trend prediction branch and the BP mutation prediction branch of the BP-LSTM combined prediction model for parallel prediction. The trend prediction value output by the LSTM trend prediction branch and the mutation prediction value output by the BP mutation prediction branch are input into the BP fusion network in the BP-LSTM combined prediction model for weighted fusion, and the total predicted traffic demand value is output. Pressure loss compensation is performed on the total predicted flow demand value. The pressure loss compensation method is as follows: set a pipeline flow resistance compensation coefficient, and perform flow resistance compensation on the total predicted flow demand value based on the pipeline flow resistance compensation coefficient to obtain the compensated flow demand value; inversely map the compensated total predicted flow demand value to the feedforward basic target frequency, determine the feedforward advance based on the pipeline transmission lag time τ, generate a fuzzy deviation state quantity based on the real-time pump station outlet pressure value, solve the feedback compensation correction frequency through fuzzy PID rules, and superimpose the feedforward basic target frequency and the feedback compensation correction frequency to generate the final control frequency to drive the emulsion pump. The method for generating the fuzzy deviation state quantity includes: calculating the pressure deviation between the real-time pump station outlet pressure value and the preset target working pressure value; obtaining the first derivative of the pressure deviation to get the pressure deviation change rate; and performing quantization and fuzzification processing on the pressure deviation and the pressure deviation change rate to map them into fuzzy linguistic variables to generate the fuzzy deviation state quantity. The method for calculating the feedback compensation correction frequency includes: performing fuzzy inference on the fuzzy deviation state quantity according to a preset fuzzy inference rule table, dynamically calculating the proportional parameter correction, integral parameter correction, and derivative parameter correction of the PID controller, updating the preset PID parameters of the PID controller in real time according to the proportional parameter correction, integral parameter correction, and derivative parameter correction, and using the updated PID parameters to calculate the pressure deviation and output the feedback compensation correction frequency.

2. The adaptive liquid supply method for emulsion pumps based on temporal deep learning according to claim 1, characterized in that, The method for constructing a multidimensional spatiotemporal feature matrix includes: Variational mode decomposition is performed on the fluid state data to generate a denoised fluid state time-series vector. The hydraulic support action command is encoded into a working condition semantic feature vector. The denoised fluid state time-series vector and the working condition semantic feature vector are fused to construct a multi-dimensional spatiotemporal feature matrix.

3. The adaptive liquid supply method for emulsion pumps based on temporal deep learning according to claim 2, characterized in that, The method for generating the denoised fluid state time-series vector includes: A sliding time window of length N is set to extract historical pressure and flow data from the past N moments from the fluid state data to construct a basic fluid state time series vector. The variational mode decomposition algorithm is used to adaptively decompose the basic fluid state time series vector to obtain K eigenmode components. The alternating direction multiplier method is used to iteratively solve the problem. High-frequency noise components are removed from the K eigenmode components and low-frequency effective signal components are retained. The low-frequency effective signal components are then reconstructed to generate a denoised fluid state time series vector.

4. The adaptive liquid supply method for emulsion pumps based on temporal deep learning according to claim 3, characterized in that, The method for encoding hydraulic support action commands into working condition semantic feature vectors includes: The action type features and action duration features are extracted from the hydraulic support action commands. The action type features are mapped to the action type one-hot encoded vector. The action type one-hot encoded vector is combined with the action duration features to generate the working condition semantic feature vector. The action types include column lowering action, support moving action, column raising action and push sliding action.

5. The adaptive liquid supply method for emulsion pumps based on temporal deep learning according to claim 4, characterized in that, The method for constructing the multidimensional spatiotemporal feature matrix includes: The denoised fluid state time-series vector and the working condition semantic feature vector are timestamped. The aligned denoised fluid state time-series vector is used as the fluid state dimension and the aligned working condition semantic feature vector is used as the working condition semantic dimension. A multidimensional spatiotemporal feature matrix is ​​constructed through matrix concatenation.

6. The adaptive liquid supply method for emulsion pumps based on temporal deep learning according to claim 5, characterized in that, The method for inversely mapping the compensated total predicted flow demand value to the feedforward base target frequency includes: A flow-frequency characteristic relationship model for an emulsion pump is established. The total predicted flow demand value is input into the flow-frequency characteristic relationship model for inverse mapping, and the feedforward basic target frequency is output.

7. The adaptive liquid supply method for emulsion pumps based on temporal deep learning according to claim 6, characterized in that, Amplitude limiting protection is applied to the final control frequency.

8. An adaptive liquid supply system for an emulsion pump based on temporal deep learning, used to implement the adaptive liquid supply method for an emulsion pump based on temporal deep learning as described in any one of claims 1-7, characterized in that, The system includes: Feature construction module: used to collect fluid state data at the outlet of emulsion pump station and pipeline state parameters of supply pipeline and capture hydraulic support action commands, and construct a multi-dimensional spatiotemporal feature matrix based on fluid state data and hydraulic support action commands; the fluid state data includes at least the real-time pump station outlet pressure value; and calculates the pipeline transmission lag time τ based on pipeline state parameters; Traffic forecasting module: Used to construct BP-LSTM combined forecasting model. The multidimensional spatiotemporal feature matrix is ​​input into the LSTM trend forecasting branch and BP mutation forecasting branch of the BP-LSTM combined forecasting model for parallel forecasting. The trend forecast value output by the LSTM trend forecasting branch and the mutation forecast value output by the BP mutation forecasting branch are input into the BP fusion network in the BP-LSTM combined forecasting model for weighted fusion, and the total predicted traffic demand value is output. Liquid supply control module: used to compensate for pressure loss of total predicted flow demand value, inversely map the compensated total predicted flow demand value to feedforward base target frequency, determine feedforward advance based on pipeline transmission lag time τ, generate fuzzy deviation state quantity based on real-time pump station outlet pressure value, solve feedback compensation correction frequency through fuzzy PID rules, and superimpose feedforward base target frequency and feedback compensation correction frequency to generate final control frequency to drive emulsion pump.

Images