Pumping displacement control system based on submersible pump

Abstract

The invention relates to the technical field of fluid machinery control, in particular to a submersible pump-based pumping displacement control system, which comprises a submersible pump unit, a sensor network, an embedded controller cluster and a dynamic cluster management module, compared with the situation that an existing controller cluster usually adopts a fixed main node or a simple election mechanism, node performance fluctuation and network state change cannot be flexibly coped with, and control power cannot be distributed to an optimal node, a multi-factor-driven dynamic weight election and reconstruction mechanism is designed; according to the scheme, the dynamic weight is periodically calculated by integrating the health degree, the computing resource, the network quality and the topological position of the node, the main controller is elected in a non-centralized mode, it is ensured that the most healthy and reliable node serves as the leader right all the time, meanwhile, smooth transfer of the control right and abnormal self-healing of the cluster are achieved, and the reliability of the cluster is improved. Therefore, the overall toughness and the long-term operation reliability of the whole pump group control system are greatly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fluid machinery control, and in particular to a pumping displacement control system based on submersible pumps. BACKGROUND

[0002] The submersible pump displacement control system is a core device widely used in municipal drainage, industrial circulating water, agricultural irrigation, and mine rescue fields. Its core goal is to achieve precise, stable, and efficient pumping displacement output by coordinating the operation of multiple parallel submersible pumps. Traditional control systems mainly rely on physical flow sensors installed on each pump to obtain real-time displacement data, and combine pre-set fixed parameters or simple control logic (such as start-stop control based on liquid level) to adjust the operation state of the pump. Such systems usually adopt a centralized control architecture, i.e., a master controller collects data from each pump and issues instructions. Its monitoring dimension is relatively limited, focusing on basic electrical parameters and single flow readings.

[0003] The existing technology has several significant shortcomings. First, the system highly depends on physical flow meters that are high-cost, easily damaged in water, and easily clogged by debris, which not only increases hardware costs and maintenance burden, but also causes a sharp decline in system reliability when the sensor fails. Second, the control strategy is rigid and difficult to adapt to complex working condition changes, such as changes in medium density or viscosity, slow degradation of device performance, and sudden failures (such as impeller winding, air lock, etc.), often leading to decreased control quality, increased energy consumption, and even causing equipment damage. Third, the centralized control architecture has a single point of failure risk. Once the master controller fails, the entire system may be paralyzed. In addition, existing systems generally lack deep health state perception and predictive maintenance capabilities, with maintenance modes mainly being passive response or regular maintenance, which cannot effectively warn and intervene before failure occurs.

[0004] The present application proposes a pumping displacement control system based on submersible pumps to address the above problems. Its core is to generate high-reliability virtual displacement estimates and health state features by fusing multi-dimensional running signals through lightweight machine learning models, replacing or redundantly using physical flow meters with software algorithms; to reveal the internal physical mechanism changes of the device using physical information neural networks; and to achieve self-organizing optimization distribution of pump group load through distributed collaborative decision-making algorithms, while ensuring high reliability of the system through a decentralized dynamic cluster management mechanism, thereby realizing a fundamental leap from passive control to active perception, adaptive optimization, and intelligent decision-making. SUMMARY

[0005] To overcome the problems presented in the background art, the present application proposes a pumping displacement control system based on submersible pumps.

[0006] The technical scheme of the present application is: a pumping displacement control system based on a submersible pump, comprising: A submersible pump unit, comprising at least two submersible pumps operating in parallel; A sensor network, corresponding to each submersible pump, for real-time acquisition of multi-dimensional operating signals of the corresponding submersible pump, including electrical signals, rotational speed signals and vibration signals, the sensor network further comprising a flow sensor installed on the pump group output main pipeline for measuring the actual output flow of the system; An embedded controller cluster, comprising a plurality of embedded controllers respectively corresponding to each submersible pump, each embedded controller being interconnected through an industrial communication network; A dynamic cluster management module for assigning dynamic weights to each controller according to the local health score, network state and topological position of each embedded controller, and dynamically electing a master controller in a decentralized manner to realize dynamic reconstruction of the control cluster structure.

[0007] As a preferred, the embedded controllers are all built-in with the same software function modules, the software function modules comprising: A signal processing and virtual sensing module for fusion analysis based on operating signals and actual output flow through a lightweight machine learning model to calculate real-time virtual displacement and health status characteristics of the corresponding submersible pump; A physical information constraint module built-in with a lightweight physical information neural network fusing pump hydraulic mechanism equation and motor mechanism equation for calculating virtual physical quantities of the corresponding submersible pump, including hydraulic efficiency, mechanical loss and medium state change quantity, with easily measured electrical quantities in the operating signals as input; An adaptive control module for dynamically adjusting the parameters of the PID controller according to the real-time virtual displacement, virtual physical quantities and preset displacement target, and generating a rotational speed control instruction to drive the motor of the corresponding submersible pump; A local health assessment module for calculating the local health score of the corresponding controller based on health status characteristics, dynamic load parameters and device historical data; A distributed collaborative decision-making module for generating a globally optimal load distribution scheme through distributed negotiation based on a cooperative game algorithm, real-time state information of each submersible pump and a simplified system hydraulic model within the embedded controller cluster.

[0008] As a preferred, the signal processing and virtual sensing module, when working, specifically comprises: S11: Data acquisition and synchronization, acquiring multi-dimensional operating signals of the submersible pump and system total flow signals, and performing time synchronization based on time stamps; S12: Signal preprocessing and feature engineering, cleaning, transforming the operating signals, and extracting time domain, frequency domain and time-frequency domain features to construct high-representative feature vectors; S13: Virtual flow estimation, input the feature vector into the lightweight machine learning model, take the readings of the flow sensor under standard working conditions as the supervision benchmark, and calculate the real-time virtual displacement of the corresponding submersible pump; S14: Online fault diagnosis, based on the feature vector, real-time identify the specific fault mode and health degradation trend of the pump through the lightweight fault diagnosis model; S15: Confidence fusion and output, evaluate the reliability of the virtual displacement, and integrate the virtual displacement and the fault diagnosis result to generate the final virtual displacement output value and the multi-dimensional health state feature vector.

[0009] As preferred, the physical information neural network in the physical information constraint module adopts an encoder-decoder structure, including: A11: Input layer, used to receive the normalized electrical feature vector containing active power, apparent power, power factor and rotating speed; A12: Encoder network, composed of multiple fully connected layers, used to extract high-dimensional latent features from input features; A13: Physical constraint fusion layer, a hidden layer of the network, the design of its neuron activation function constrains the mechanism relationship between pump head, flow, shaft power and motor electromagnetic torque; A14: Output layer, output virtual physical quantities, including real-time hydraulic efficiency, total mechanical loss power and dimensionless factor representing medium state change.

[0010] As preferred, in the training process of the physical constraint fusion layer, the composite physical loss function applied is defined as the sum of mechanism equation residuals: ; Wherein, is the composite physical loss function, and are weighting coefficients, is the pump hydraulic mechanism loss term, defined as: ; Wherein, is the network-derived pump shaft power, is the standard medium density, is the acceleration of gravity, is the network-derived head according to the similarity law and the current rotating speed and medium factor, is the predicted flow of the pump under the current working condition, is the network-predicted pump hydraulic efficiency; is the motor mechanism loss term, defined as: ; Wherein, is the electromagnetic torque calculated from the input electrical quantities according to motor theory, is the angular velocity, is the total mechanical loss torque output by the network.

[0011] As preferred, the adaptive control module specifically comprises: A21: a parameter dynamic adjustment unit, configured to calculate a basic parameter group of the PID controller in real time according to the medium state change factor and the hydraulic efficiency in the virtual physical quantity; A22: a control performance evaluation and fine-tuning unit, configured to fine-tune and optimize the basic parameter group online according to the tracking error and error change rate characteristics of the virtual displacement; A23: a fault state control strategy switching unit, configured to switch between normal adaptive PID control, degraded robust PID control and safety shutdown protocol according to the health state feature vector; A24: a speed instruction generation unit, configured to calculate and output a speed control instruction for driving the submersible pump motor based on the finally determined PID parameters and the preset displacement target.

[0012] As preferred, the local health assessment module specifically comprises: A31: a multi-dimensional health index calculation unit, configured to process the health state feature vector, the virtual physical quantity and the real-time dynamic load parameter, and calculate a standardized basic health index; A32: a historical data analysis and trend extraction unit, configured to access device historical operation data, and calculate a health degradation trend index based on time series analysis; A33: a comprehensive health degree fusion calculation unit, configured to fuse and calculate the basic health index and the trend index through a weighted evaluation model to obtain a final health degree score.

[0013] As preferred, the distributed collaborative decision-making module specifically comprises: A41: a local benefit calculation unit, configured to construct and calculate a local benefit function based on the real-time state, health degree and energy consumption characteristics of the pump; A42: a hydraulic model calculation unit, configured to internally build a simplified system hydraulic model, for evaluating the working point change of the pump and its influence on the head-flow of the system pipeline under different load distribution schemes; A43: a neighbor information processing unit, configured to periodically receive and analyze the state and decision proposal information of the adjacent controllers through the communication network; A44: a distributed negotiation engine, configured to perform iterative calculation to seek a Nash equilibrium point of the cluster based on the local information, neighbor information and hydraulic model calculation results, by using a cooperative game algorithm; A45: decision output and consistency verification unit, used for outputting the final rotation speed set point of the pump after negotiation convergence, and performing final consistency verification with other nodes of the cluster.

[0014] As preferred, the dynamic cluster management module specifically comprises: A51: dynamic weight calculation unit, used for periodically calculating the dynamic weight value of the node, and specifically, the dynamic weight value integrates local node health degree, calculation resource state, network connection quality and topology position information; A52: cluster state perception and synchronization unit, used for maintaining the dynamic weight list and role information of all online nodes in the cluster through periodic heartbeat messages and state broadcast; A53: distributed consensus election unit, adopting a competition-response mechanism based on dynamic weight, so as to make the cluster reach consensus on the master controller identity without a central coordinator; A54: smooth control right transfer unit, used for managing and controlling the transfer and synchronization process related to control strategy and historical state data when the master controller is changed; A55: abnormality processing and cluster reconstruction unit, used for detecting node failure and network partition abnormality, and triggering topology reconstruction and weight recalculation of the cluster.

[0015] As preferred, the workflow of the distributed consensus election unit is a broadcast-response mechanism based on priority, and specifically comprises the following steps: S81: each node periodically broadcasts its heartbeat message, and the message at least contains its node, current dynamic weight, current role and a logical timestamp; S82: if a node detects that its weight is higher than that of the current known master controller and the difference exceeds a threshold value, the node broadcasts a master controller declaration message to the cluster; S83: within a preset competition window period, if a node receives a master controller declaration from another node with higher weight, the node withdraws its own declaration and becomes a slave controller, and if no higher weight declaration is received, the node becomes the master controller after the window period ends; S84: the new master controller broadcasts a role confirmation message, all slave controllers update the master controller information maintained by them, and send a confirmation response to the master controller.

[0016] The present application has the following beneficial effects: 1. Compared to existing controller clusters that often use fixed master nodes or simple election mechanisms, which cannot flexibly cope with node performance fluctuations and network status changes, potentially leading to control not being allocated to the optimal node, this invention designs a multi-factor driven dynamic weight election and reconstruction mechanism. This scheme integrates node health, computing resources, network quality, and topology location to periodically calculate dynamic weights, and elects the master controller in a decentralized manner, ensuring that leadership is always held by the healthiest and most reliable node. At the same time, it achieves smooth transfer of control and cluster anomaly self-healing, thereby greatly improving the overall resilience and long-term operational reliability of the entire pump group control system. 2. Compared to existing purely data-driven models that often lack physical constraints and are prone to producing outputs that do not conform to physical laws when operating conditions change drastically, limiting their generalization ability and reliability, this invention innovatively embeds the hydraulic mechanism of the pump and the motor mechanism equations as hard constraints into a lightweight neural network to construct a physical information constraint module. This scheme ensures that the model output, such as hydraulic efficiency and mechanical loss, strictly follows physical laws. In particular, it can calculate a dimensionless factor characterizing the change in the state of the medium, thereby significantly improving the model's extrapolation ability under unknown operating conditions and the physical reliability of the output results, providing a solid basis for adaptive control. 3. Compared with existing technologies that mostly use controllers with fixed parameters, which are difficult to adapt to complex dynamic conditions such as changes in media and equipment wear, leading to a decline in control quality or even system instability, this invention adopts an adaptive control strategy based on virtual physical quantities and health status characteristics. This scheme can dynamically adjust control parameters according to media state factors and hydraulic efficiency, and make fine adjustments in conjunction with real-time performance evaluation. It can also achieve smooth degradation switching of control strategies according to health status, thereby significantly enhancing the robustness, adaptability and safety of the control system under different operating conditions, and realizing the transformation from passive response to active optimization. 4. Compared to the traditional centralized control architecture, which has the risk of single point of failure and whose load allocation strategies often fail to balance energy efficiency and equipment health, this invention constructs a distributed collaborative decision-making mechanism based on cooperative game theory. This scheme enables each pump controller to become an autonomous intelligent agent, which, through mutual negotiation, converges quickly to the Nash equilibrium point based on the local benefit function and the system hydraulic model. This decentralized architecture not only eliminates single point of failure and improves system resilience, but also achieves optimal load allocation in the cluster, while pursuing the goals of minimizing total system energy consumption and balancing the losses of each pump. 5、The prior art mainly relies on high-cost and easily damaged physical flow meters for single-point monitoring, which has the disadvantages of strong hardware dependence, high maintenance cost and inability to early warn equipment failure. The virtual sensor scheme based on multi-sensor signal fusion and lightweight machine learning model is adopted, the pump displacement is estimated in real time by fusing electrical, vibration and other multi-dimensional operation signals, and online fault diagnosis is performed; the software algorithm replaces or redundantly uses physical sensors, which significantly reduces the hardware cost, and improves the predictive maintenance capability and overall reliability of the system through early fault identification. BRIEF DESCRIPTION OF DRAWINGS

[0017] Fig. 1 A constructional schematic view of a pumping displacement control system based on a submersible pump is shown. Fig. 2 A constructional schematic view of a dynamic cluster management module in the pumping displacement control system based on the submersible pump is shown. DETAILED DESCRIPTION

[0018] The present application will be further described below in conjunction with the drawings and examples.

[0019] Please refer to Figs. 1-2 The present application provides an embodiment: a pumping displacement control system based on a submersible pump, comprising: I. A submersible pump unit, comprising at least two submersible pumps operating in parallel.

[0020] II. Sensor network Corresponding to each submersible pump, for real-time acquisition of multi-dimensional operation signals of the corresponding submersible pump, including electrical signals, rotational speed signals and vibration signals, the sensor network further comprises a flow sensor installed on the pump group output main pipeline for measuring the actual output flow of the system.

[0021] III. Embedded controller cluster Comprising a plurality of embedded controllers respectively corresponding to each submersible pump, the embedded controllers are interconnected through an industrial communication network, and each embedded controller is internally provided with the same software function module, which includes: 1. Signal processing and virtual sensor module For fusion analysis based on operation signals and actual output flow through a lightweight machine learning model, the real-time virtual displacement and health status characteristics of the corresponding submersible pump are calculated, and the workflow is: S11: Data acquisition and synchronization, acquiring multi-dimensional operation signals of the submersible pump and system total flow signals, and time synchronization based on time stamp, specifically: S111: Synchronously acquiring three-phase current, three-phase voltage, motor speed and multi-axis vibration acceleration signals installed at specific positions of the pump body at a fixed sampling frequency; S112: Acquire the instantaneous system total flow measured by the flow sensor installed on the pump group output header pipeline; S113: Time stamp all collected signals uniformly and store them in a fixed-length first-in-first-out data buffer for subsequent frame processing; S12: Signal preprocessing and feature engineering, clean and transform the operation signals, and extract time domain, frequency domain and time-frequency domain features to construct high-representative feature vectors, specifically: S121: Electrical quantity calculation, according to the collected three-phase electrical signals, real-time calculation of instantaneous total active power, apparent power, power factor and total harmonic distortion rate of current; S122: Vibration signal processing, after band-pass filtering of multi-axis vibration acceleration signals, calculate the effective value of the synthesized vibration acceleration, and perform fast Fourier transform on the synthesized signal to extract the following frequency band energy features: bearing characteristic frequency band energy, impeller passing frequency band energy and high-frequency wideband energy; S123: Speed fluctuation analysis, calculate the standard deviation of the speed signal in the current time window as a feature of load stability; S124: Feature vector construction, combine all the calculated and extracted features with the original speed to form a multi-dimensional feature vector: ; Wherein, is the instantaneous active power, is the instantaneous apparent power, is the instantaneous power factor, is the total harmonic distortion rate of current, is the effective value of the synthesized vibration acceleration, is the bearing characteristic frequency band energy, is the impeller passing frequency energy, is the high-frequency wideband energy, is the speed fluctuation standard deviation, is the real-time speed; S125: Feature standardization, online standardization processing of the feature vector to meet the scale requirements of the model input; S13: Virtual flow estimation, input the feature vector into the lightweight machine learning model, take the readings of the flow sensor under standard working conditions as the supervision benchmark, and calculate the real-time virtual displacement of the submersible pump; S14: Online fault diagnosis, based on the feature vector, real-time identification of specific fault modes and health degradation trends of the pump through the lightweight fault diagnosis model, specifically: S141: Fault feature extraction, extract key subsets from the feature vector for specific fault modes, including: For winding and deposition: mainly focus on the increase of total harmonic distortion of current, the relative change of high-frequency broadband energy; For gas pocket: mainly focus on the sudden peak of synthetic vibration acceleration effective value, the fluctuation of instantaneous power factor, and the abnormal change of bearing characteristic frequency band energy and impeller passing frequency energy; For surge: mainly focus on the periodic sharp fluctuation of standard deviation of rotating speed fluctuation, and the same frequency oscillation characteristics of pressure and flow signals; S142: Diagnose model inference, input the key subset into the lightweight fault classification model, and output the fault probability vector; S143: Normalize and weightedly fuse the fault probability vector, vibration energy trend index and current harmonic trend index to generate a comprehensive health state feature vector, which represents the overall health degree and fault evolution trend of the device; S15: Confidence fusion and output, evaluate the reliability of virtual displacement, and integrate virtual displacement and fault diagnosis results to generate the final virtual displacement output value and multi-dimensional health state feature vector, specifically: S151: Virtual flow confidence evaluation, design confidence evaluation function: ; Wherein, is the confidence evaluation function, is the absolute error of virtual flow and measured total flow, is the rated flow of the pump, is the information entropy of the fault probability vector, and are adjustment coefficients; S152: Final output decision, the specific rules are: When is higher than the preset high threshold, output as the final estimated displacement of the pump; When is between the high and low thresholds, the weighted fusion of and the estimated value based on the simplified hydraulic model is output; When is lower than the preset low threshold, trigger strong abnormal alarm, and switch to the conservative estimation mode based on the fixed hydraulic model; S153: Module output, finally output the calibrated virtual displacement value, health state feature vector and current confidence level of the pump.

[0022] In this embodiment, by constructing a signal processing and virtual sensing module integrating multi-sensor fusion, embedded AI and model self-checking, real-time and highly reliable sensing of the flow and health status of the submersible pump is achieved. Specifically, the module synchronously collects multi-dimensional operating signals such as current, voltage, speed and vibration, and combines with the flow meter reading to construct a representative vector through a series of feature engineering, and input the lightweight machine learning model to output real-time virtual displacement estimation; at the same time, the model can diagnose specific faults such as winding and air lock to generate a health status feature vector; finally, through a unique confidence evaluation function, the virtual displacement is self-checked and decided to ensure reliable output. The software algorithm replaces or redundancies the high-cost and easily-damaged physical flow meter, reducing hardware dependence and maintenance cost; through online fault diagnosis and confidence evaluation, the fault tolerance, state sensing ability and decision reliability of the system are significantly improved, laying a solid data foundation for predictive maintenance and adaptive optimization control of the pump, and realizing the intelligent leap from passive response to active sensing.

[0023] 2. Physical information constraint module A lightweight physical information neural network integrating pump hydraulic mechanism equation and motor mechanism equation is built in, which is used to calculate the virtual physical quantities of the submersible pump, including hydraulic efficiency, mechanical loss and medium state change quantity, by taking the easily measured electrical quantities in the operating signals as input. The specific working process is as follows: S21: receiving the standardized electrical feature vector from the signal processing module; S22: inputting the feature vector into the trained physical information neural network to calculate the virtual physical quantity output through forward propagation; S23: based on the medium state change factor, dynamically adjusting the PID parameters of the control system, the specific rules are as follows: when the medium state change factor continuously deviates from 1, adjust the proportional gain and integral time of the PID controller to adapt to the change of the system dynamic characteristics caused by the change of the medium; S24: outputting the hydraulic efficiency and total mechanical loss power to the health assessment module as key indicators for assessing the energy efficiency state and mechanical wear state of the pump.

[0024] Specifically, the physical information neural network in the physical information constraint module adopts an encoder-decoder structure, which includes: A11: input layer, used to receive the normalized electrical feature vector containing active power, apparent power, power factor and speed; A12: encoder network, composed of multiple fully connected layers, used to extract high-dimensional latent features from input features; A13: physical constraint fusion layer, hidden layer of the network, the design of the neuron activation function of which constrains the mechanism relationship between pump head, flow, shaft power and motor electromagnetic torque; A14: output layer, output virtual physical quantities, including real-time hydraulic efficiency, total mechanical loss power and dimensionless factor representing medium state change.

[0025] Specifically, in the training process of the physical constraint fusion layer, a composite physical loss function is applied, which is defined as the sum of mechanism equation residuals: ; wherein, is the composite physical loss function, and is a weighting coefficient, is a pump hydraulic mechanism loss term, defined as: ; wherein, is the pump shaft power derived by the network, is the standard medium density, is the gravitational acceleration, is the head derived by the network according to the similarity law and the current speed and medium factor, is the predicted flow of the pump under the current working condition, is the predicted pump hydraulic efficiency; is a motor mechanism loss term, defined as: ; wherein, is the electromagnetic torque calculated by the input electrical quantity according to the motor theory, is the angular velocity, is the total mechanical loss torque output by the network.

[0026] In the present embodiment, the physical definition and calculation of the medium state change factor are realized by the network through a specific physical constraint path, which couples the flow, head and speed through the similarity law of the pump and the modified pipeline characteristic curve, specifically: define the virtual head-flow relationship containing the influence of the medium state: ; wherein, is the head derived by the network according to the similarity law and the current speed and medium factor, is the real-time speed of the pump, is the rated speed, is the zero-flow head of the pump, is the comprehensive pipeline system resistance coefficient, is the medium state change factor, is the predicted flow of the pump under the current working condition; The network needs to make the output of the network satisfy the following relationship during training and inference , , At the same time, the above relationship and formula are satisfied, so that the medium state change factor becomes a learnable, normalized physical quantity that comprehensively reflects the influence of medium density and viscosity change on pump characteristics. When the medium state change factor is equal to 1, the medium is standard water. When the medium state change factor is less than 1, it may indicate that the medium contains gas or the density is reduced. When the medium state change factor is greater than 1, it may indicate that the medium viscosity or density is increased.

[0027] In this embodiment, the total mechanical loss power is decomposed into three sub-terms of bearing friction loss, seal friction loss and disc friction loss, and is constrained by the following empirical physical relationship during network training: ; wherein, is the total mechanical loss power, is the bearing friction loss, is the seal friction loss, is the disc friction loss, , and are loss coefficients learned by the network and related to specific pump types.

[0028] In this embodiment, a mechanism and data deep fusion physical information constraint module is constructed, which directly embeds the pump hydraulic equation and motor mechanism equation as a hard constraint into the model training and inference process through a carefully designed lightweight physical information neural network. The module takes easily measured electrical quantities as input, not only outputs key virtual physical quantities such as hydraulic efficiency and mechanical loss, but also innovatively calculates a dimensionless factor that can directly represent the change of medium state through a physical constraint path (fusing the similarity law and the modified pipeline characteristics). On the one hand, the physical loss function and the structure design ensure that the output quantities conform to the basic physical laws, significantly improving the generalization ability and extrapolation reliability of the model, and effectively overcoming the "black box" problem of traditional pure data models and the failure outside the working condition; on the other hand, it provides reliable input with clear physical meaning for the downstream PID parameter adaptive adjustment (based on the medium factor) and equipment health assessment (based on efficiency and loss), thereby realizing the paradigm upgrade from "empirical fitting" to "mechanism-guided intelligent perception" within the limited computing power of embedded devices, and laying a solid physical foundation for the robust and optimal operation of the control system in complex working conditions such as changing media and equipment wear.

[0029] 3. Adaptive control module The application is used for dynamically adjusting parameters of a PID controller according to real-time virtual displacement, virtual physical quantity and preset displacement target, and generating a rotating speed control instruction to drive a motor of a corresponding submersible pump.

[0030] The adaptive control module specifically comprises: A21: a parameter dynamic adjustment unit, configured to calculate a basic parameter group of the PID controller in real time according to a medium state change factor and hydraulic efficiency in the virtual physical quantity; A22: a control performance evaluation and fine adjustment unit, configured to perform online fine adjustment and optimization on the basic parameter group according to a tracking error of the virtual displacement and error change rate characteristics; A23: a fault state control strategy switching unit, configured to switch between normal adaptive PID control, degraded robust PID control and a safety shutdown protocol according to a health state feature vector; A24: a rotating speed instruction generation unit, configured to calculate and output a rotating speed control instruction for driving the motor of the submersible pump based on the finally determined PID parameters and the preset displacement target.

[0031] In the embodiment, the specific working process of the parameter dynamic adjustment unit is as follows: S31: receiving the medium state change factor and the hydraulic efficiency from the physical information constraint module; S32: calculating a base value of the proportional gain, and the calculation formula is as follows: ; Wherein, the base value of the proportional gain, the medium state change factor, the current real-time pump hydraulic efficiency, the nominal proportional gain of the controller, the adjustment sensitivity coefficient for the medium state factor, the rated efficiency of the pump running at the optimal operating point, the adjustment sensitivity coefficient for the efficiency change; S33: calculating a base value of the integral time, and the calculation formula is as follows:

[0032] Wherein, the base value of the integral time, the nominal integral time, the exponential coefficient greater than 0 determined according to the system inertia; S34: calculating a base value of the differential time, and the calculation formula is as follows:

[0033] Wherein, is the base value of the integral time, is the nominal differential time, is the compensation coefficient; S35: output the dynamic PID parameter base value group.

[0034] In the embodiment, the control performance evaluation and fine-tuning unit works specifically as follows: S41: receive the virtual displacement and the displacement target setting value, calculate the current control error and the error change rate; S42: define the transient performance index: ; wherein, is the transient performance index, is the current control error, is the error change rate, is the; S43: if the transient performance index continuously exceeds the threshold value within the set observation time window, start parameter fine-tuning, the fine-tuning amount is related to the integral of the error and the error change rate, and the mathematical expression is: ; wherein, is the fine-tuning amount, is the maximum error limit, is the sign function; S44: superimpose the fine-tuning amount to the base parameter, and simultaneously perform similar but smaller amplitude compensatory fine-tuning on the base value of the integral time and the base value of the differential time, to finally generate the PID parameter for the current control period.

[0035] In the embodiment, the fault state control strategy switching unit makes decisions according to the health state feature vector from the signal processing and virtual sensing module, and the rules include: when the health state feature vector indicates normal state or slight degradation, adopt the parameters output by the control performance evaluation and fine-tuning unit for adaptive PID control; when the health state feature vector indicates moderate fault risk, switch to the degraded robust control strategy, specifically: switch the PID controller to PI control only, and reduce the proportional gain to times of the original value, at the same time, increase the integral time to slow down the control action and avoid triggering the fault; when the health state feature vector indicates serious fault, switch to the safe shutdown protocol, including: generate a gentle speed reduction ramp instruction, cut off the power after the speed drops to the safety threshold, and trigger the advanced alarm.

[0036] In the embodiment, the speed instruction generation unit works specifically as follows: S51: using the final determined PID parameters, the position type PID algorithm is used to calculate the speed adjustment amount required in the current period, and the calculation formula is: ; wherein, is the control period, is the proportional gain of the superimposed fine adjustment amount, is the integral time of the superimposed fine adjustment amount, is the differential time of the superimposed fine adjustment amount; S52: superimpose the speed adjustment amount on the speed command of the last period to obtain the speed command of the current period; S53: apply physical limiting to the speed command to ensure that it is within the safe operating speed range of the pump, and apply a change rate limit to ensure smooth speed change; S54: output the final speed command after limiting to the motor driver.

[0037] The embodiment realizes precise, robust and high-reliability closed-loop control of the displacement of the submersible pump by constructing a multi-mode, adaptive and safe degradation intelligent control module. The core of the module is to deeply integrate the output of the upstream sensing module (virtual displacement, virtual physical quantity, health status), and to build a three-layer progressive control architecture based on this: first, the parameter dynamic adjustment unit reconstructs the PID parameter base value in real time according to the medium state change factor and the hydraulic efficiency through a formula with clear physical meaning, so that the dynamic characteristics of the controller automatically match the current working condition; second, the control performance evaluation and fine adjustment unit performs online "fine adjustment" of the parameters based on the transient performance index to optimize the dynamic response; finally, the fault state control strategy switching unit realizes gradient strategy degradation from "fine optimization" to "robust stability maintenance" to "safe shutdown" according to the health characteristics. Finally, the speed command generation unit outputs a smooth driving command under multiple safety limiting. The control system can actively adapt to medium changes and equipment performance drift, pursue optimal control quality under normal working conditions, and automatically implement preventive protection under abnormal working conditions, thereby systematically improving the adaptability, stability and equipment operation safety in complex and variable field environments, and realizing integrated autonomous control from "fixed parameter control" to "state sensing-intelligent decision-making-safe execution".

[0038] 4. Local health assessment module used for calculating the local health score of the corresponding controller based on the health status characteristics, dynamic load parameters and equipment historical data.

[0039] The local health assessment module specifically includes: A31: Multidimensional health index calculation unit, used for processing health state feature vector, virtual physical quantity and real-time dynamic load parameter, and calculating standardized basic health index; A32: Historical data analysis and trend extraction unit, used for accessing equipment historical operation data, and calculating health degradation trend index based on time series analysis; A33: Comprehensive health degree fusion calculation unit, used for fusing and calculating final health degree score through weighted evaluation model based on basic health index and trend index.

[0040] In the embodiment, the basic health index calculated by the multidimensional health index calculation unit includes fault risk index, performance degradation index and load severity index, and the calculation formulas are as follows: Fault risk index: Wherein, is the fault risk index, is the probability of winding and deposition fault output by the fault diagnosis model in real time, is the fault probability of air block and cavitation output by the fault diagnosis model in real time, is the fault probability of surge output by the fault diagnosis model in real time, , and are weight coefficients corresponding to the three faults; Performance degradation index: Wherein, is the performance degradation index, is the current hydraulic efficiency, is the benchmark efficiency of the equipment; Load severity index: Wherein, is the load severity index, is the standard deviation of speed fluctuation, is the average speed, is the current peak value in the statistical period, is the rated current of the motor.

[0041] In the embodiment, the historical data analysis and trend extraction unit specifically includes the following steps when working: S61: Circularly storing time series of key performance parameters in the local memory, including daily average efficiency, average fault risk index and peak load severity; S62: Performing sliding window linear regression analysis on the key performance parameter sequence to calculate the recent degradation slope; ​​​S63: Calculate the historical trend indicator, the calculation formula is: ; wherein, is the historical trend indicator, is the weight coefficient of the efficiency degradation trend, is the recent degradation slope of the daily average efficiency, is the weight coefficient of the failure risk growth trend, is the recent degradation slope of the average failure risk indicator, is the weight coefficient of the load severity growth trend, is the recent degradation slope of the peak load severity; S64: Based on the cumulative running time and the historical load severity, the mechanical life consumption score is estimated using a linear damage accumulation model.

[0042] In the present embodiment, the formula for calculating the final health degree score by the comprehensive health degree fusion calculation unit is: ; wherein, is the local health degree comprehensive score, is the penalty mapping function of the failure risk indicator, is the penalty function of the performance degradation indicator, is the penalty function of the load severity indicator, is the historical trend indicator, is the mechanical life consumption score, specifically: , is the adjustment parameter; ; ; is the load severity alarm threshold.

[0043] The embodiment realizes accurate and dynamic scoring of the health state of the submersible pump by constructing a comprehensive quantitative evaluation model integrating real-time state, historical trend and load characteristics. The module is based on real-time indexes such as fault risk, efficiency degradation and load severity, combines the performance degradation trend extracted from historical data, and finally outputs a comprehensive health degree score ranging from 0 to 100 through a scientific weighted fusion formula. The core benefit is that the scattered and qualitative device state information (such as vibration, efficiency, fault probability) is converted into a unified, quantifiable and comparable decision basis, which not only provides an accurate and timely trigger signal for predictive maintenance, effectively avoiding sudden shutdown, but more importantly, the score is used as a key input by the dynamic cluster management module for master controller election, thereby realizing intelligent inclination of control right to the 'healthiest' node at the system level, significantly improving the operation reliability, resource utilization efficiency and life cycle management level of the entire pump cluster system.

[0044] 5. Distributed cooperative decision module For generating a globally optimal load distribution scheme in the embedded controller cluster based on the cooperative game algorithm, the real-time state information of each submersible pump and a simplified system hydraulic model through distributed negotiation.

[0045] The distributed cooperative decision module specifically includes: A41: Local benefit calculation unit, for constructing and calculating the local benefit function based on the real-time state, health degree and energy consumption characteristics of the pump; A42: Hydraulic model calculation unit, which internally has a simplified system hydraulic model for evaluating the change of the pump's operating point and its influence on the head-flow of the system pipeline under different load distribution schemes; A43: Neighbor information processing unit, for periodically receiving and analyzing the state and decision proposal information of adjacent controllers through the communication network; A44: Distributed negotiation engine, based on the cooperative game algorithm, using local information, neighbor information and hydraulic model calculation results, performing iterative calculation to seek the Nash equilibrium point of the cluster; A45: Decision output and consistency verification unit, for outputting the final speed set point of the pump after negotiation convergence and performing final consistency verification with other nodes in the cluster.

[0046] In the embodiment, the local benefit function constructed by the local benefit calculation unit integrates operating cost, health degree cost and state switching penalty, and the calculation formula is: ; Where, represents the local benefit function value of the i-th submersible pump, represents the allocated power of the i-th submersible pump, represents the estimated working head of the i-th submersible pump under a specific working condition, represents the operating energy consumption cost function of the i-th submersible pump, is a health cost weight coefficient, is a health cost term representing the i-th submersible pump, is a load variation penalty weight coefficient, is a load variation penalty term representing the i-th submersible pump, represents the power variation of the i-th submersible pump.

[0047] In this embodiment, the simplified system hydraulic model core built-in the hydraulic model calculation unit is a head balance equation and a flow superposition equation based on the basic law of parallel pump system, which is specifically represented as: for all running pumps i; ; wherein, is the total pipeline head of the system, is the head-flow characteristic curve corresponding to the flow of of pump i at the current speed, is the pipeline characteristic curve of the total pipeline.

[0048] In this embodiment, the cooperative game algorithm executed by the distributed negotiation engine adopts a distributed optimization method based on marginal benefit consistency, and the iteration steps are: S71: initialization, each node i proposes an initial flow proposal based on the current demand according to a simple load distribution rule, and calculates the corresponding initial marginal benefit; S72: information exchange, each node i broadcasts its current proposal and marginal benefit to all other nodes; S73: local update, after each node i receives the information of other nodes, the system average marginal benefit is calculated, and the self-proposal is updated according to the following formula: ; wherein, represents the updated flow proposal value allocated to the i-th submersible pump at the next iteration round k+1, represents the flow proposal value allocated to the i-th submersible pump at the iteration round k, is a convergence step size, represents the average marginal benefit of all running pumps in the entire submersible pump cluster at the kth iteration round, is a local marginal benefit representing the i-th pump based on its current proposal and the corresponding head calculated by the hydraulic model at the kth iteration round; S74: Convergence judgment, when the marginal benefit of all pumps meets , and the total flow meets the accuracy requirement of demand, the iteration ends, and Nash equilibrium is reached, is a preset small threshold.

[0049] The embodiment realizes self-organizing optimization distribution of pump group load by constructing a decentralized intelligent game system. Each pump controller serves as an autonomous intelligent agent, based on a local benefit function which integrates energy consumption, health degree and adjustment penalty, combined with a simplified system hydraulic model, through marginal benefit exchange and iteration, the whole cluster is driven to quickly converge to the Nash equilibrium state where the marginal benefits of each pump are equal, so that in the absence of central node coordination, a globally optimal load distribution scheme with the lowest total energy consumption and the most balanced equipment wear is generated in a distributed manner. The completely distributed architecture eliminates the risk of single point failure, enhances the system robustness; by directly incorporating the device health status into the game objective, the synergy of energy efficiency optimization and device life extension is realized; at the same time, based on the physical model and the mechanism of real-time negotiation, the system can dynamically adapt to changes in working conditions and network fluctuations, ultimately realizing significant improvement in overall operation efficiency, reliability and economy of the pump group.

[0050] IV. Dynamic cluster management module for assigning dynamic weights to each controller according to the local health degree score, network state and topological position of each embedded controller, and dynamically electing a master controller in a decentralized manner, realizing dynamic reconstruction of the control cluster structure.

[0051] The dynamic cluster management module specifically includes: A51: Dynamic weight calculation unit, for periodically calculating the dynamic weight value of the node, specifically, the dynamic weight value integrates local node health degree, calculation resource state, network connection quality and topological position information; A52: Cluster state perception and synchronization unit, for maintaining the dynamic weight list and role information of all online nodes in the cluster through periodic heartbeat messages and state broadcasts; A53: Distributed consensus election unit, using a competition-response mechanism based on dynamic weights, to reach an agreement on the identity of the master controller in the cluster without a central coordinator; A54: Smooth control right transfer unit, for managing and controlling the transfer and synchronization process related to control strategy and historical state data when the master controller changes; A55: Abnormality handling and cluster reconstruction unit, for detecting node failure and network partition abnormality, and triggering topology reconstruction and weight recalculation of the cluster.

[0052] In the embodiment, the formula for calculating the dynamic weight value of node i by the dynamic weight calculation unit is: ; wherein, is the normalized health score, which is normalized from the health score provided by the local health assessment module, is the normalized computing resource surplus, defined as , is the comprehensive score of network status, calculated by the formula: , is the average communication delay from the node to all other nodes, is the maximum allowed delay, is the recent network packet loss rate, is the node betweenness centrality based on network topology, , and are the weight coefficients of each item; is the control load factor, defined as , , , and are the global weighting coefficients corresponding to the respective items.

[0053] As a preferred, the workflow of the distributed consensus election unit is a priority-based broadcast-response mechanism, which specifically includes the following steps: S81: Each node periodically broadcasts its heartbeat message, which at least contains its node, current dynamic weight, current role, and a logical timestamp; S82: If a node detects that its weight is higher than the weight of the current known master controller and the difference exceeds the threshold value, it broadcasts a master controller declaration message to the cluster; S83: Within the preset competition window period, if a node receives a master controller declaration from other nodes with higher weight, the node withdraws its own declaration and becomes a slave controller, and if it does not receive a higher weight declaration, the node becomes the master controller after the window period ends; S84: The new master controller broadcasts a role confirmation message, and all slave controllers update the master controller information they maintain and send a confirmation response to the master controller.

[0054] In this embodiment, the smooth control right transfer unit includes the following when working: State synchronization request: The new master controller requests a key cluster state snapshot from the original master controller or directly from each pump controller, including the current global load distribution scheme, the health history of each pump, and the parameters of the cooperative control strategy being executed; Control loop seamless switching: During state synchronization, the local adaptive control module of each pump continues to operate based on the set point of the last cycle. After the new master controller completes state synchronization, it issues a synchronized and unified control cycle start signal in a broadcast manner. The cluster operates under the coordination of the new master controller from this point on. Historical log inheritance: The original master controller forwards its historical running logs that have not been uploaded to the cloud to the new master controller, ensuring the completeness of data records.

[0055] In this embodiment, the response mechanisms of the abnormality handling and cluster reconstruction unit for different abnormal scenarios include: Node failure detection: If any node is not perceived by other nodes for consecutive M heartbeat cycles, it is marked as suspected failure. The master controller sends a probe packet to it. If there is no response for consecutive N times, it is confirmed as failed, removed from the dynamic weight list, and the control task is redistributed. Network partition handling: When network partition occurs, causing a single cluster to split into multiple independent sub-clusters, each sub-cluster independently runs a distributed consensus election process, electing a master controller within each partition, and continues to operate. After the network is restored, the master controller with the highest weight becomes the master controller of the merged cluster and performs state merging. Dual master controller prevention: Embed logical clock and cluster generation number in heartbeat packet. When a node receives a declaration from a master controller with higher generation or later logical time but the same weight, it prioritizes the generation number and then the logical time to make a decision, avoiding dual master controllers.

[0056] In this embodiment, the dynamic cluster management module realizes high-reliability autonomous management of submersible pump controller clusters by building a decentralized, multi-factor driven intelligent election and reconstruction mechanism. The core of this module is that each controller periodically calculates a dynamic weight that integrates its own health, computing resources, network quality (including delay, packet loss, and topology centrality), and current load. Based on this weight, a distributed competition-response and consensus protocol is used to dynamically elect the globally optimal master controller without a central coordinator. When the master controller changes, the module ensures smooth and seamless transfer of control and state data. At the same time, its embedded abnormality handling mechanism can automatically detect node failure and network partition and trigger self-repair and reconstruction of the cluster. This completely eliminates the single point of failure risk of traditional central control architecture. Through dynamic election of "the fittest", the cluster leadership is always held by the most healthy and reliable node, achieving optimal allocation of control resources. Its strong fault tolerance and self-healing ability ensures the system continues to operate stably when facing device failures or network fluctuations, greatly improving the overall resilience, adaptability, and long-term reliability of the entire pump group control system.

[0057] In this embodiment, the lightweight machine learning model used in the virtual flow estimation step is constructed and run following the following method: S131: Model selection and structure, a fully connected neural network with two hidden layers is adopted as the base model, and model pruning and quantization techniques are used to ensure that it meets the computational and storage limitations of embedded controllers; S132: Introduce a loss function with physical consistency constraints, define a composite loss function during model training: ; Where, is the mean square error loss of virtual flow and actual flow, is the virtual flow, is the actual flow, is the physical consistency loss, defined as: ; Where, is the estimated motor input active power, is the medium density, is the acceleration of gravity, is the current lift, is the efficiency estimated by the pump similarity law from the speed and virtual flow, is the estimated fixed loss power of the motor and pump, and are the weighting coefficients; S133: Model training, use the historical data set containing multiple working conditions to train the model to minimize the composite loss function; S134: Online inference, during runtime, input the normalized feature vector into the trained lightweight model, and directly output the virtual displacement of the corresponding pump.

[0058] In this embodiment, the lightweight fault classification model is a decision-level fusion diagnostic designed specifically for embedded pump control scenarios. Its core is to combine deeply with the physical fault mechanism of the pump, and the input is not the original signal, but a subset of the feature vector selected for specific purposes. For example, the model will allocate a subset dominated by current harmonics and high-frequency vibration energy for winding / accumulation faults, and a subset dominated by vibration burst features and electrical stability for airlock faults. This approach significantly reduces model complexity. The model output is a normalized fault probability vector, which directly quantifies the likelihood of multiple concurrent faults and seamlessly integrates with the subsequent health state vector, thereby achieving real-time, interpretable diagnosis of complex hydraulic machinery composite faults under extremely limited edge computing power.

[0059] The embodiment realizes high-precision virtual flow perception and high-reliability real-time fault diagnosis on the embedded end by designing two lightweight AI models deeply integrated with mechanism and data. For virtual flow estimation, the model innovatively introduces a physically consistent constraint loss function in training, taking the mechanism relationship between motor power, lift, efficiency, etc. as a regularization term, so that the data-driven model output conforms to the physical law, significantly improving the generalization ability and reliability of out-of-condition extrapolation. For fault diagnosis, the model uses a feature subset selection strategy based on physical mechanism, focusing on key signal features for different fault modes, and outputs a normalized fault probability vector to realize concurrent and quantitative diagnosis of complex faults such as winding, air lock, and surge. Under the condition of strict limitation of embedded device computing power and storage resources, not only the flow estimation accuracy comparable to physical sensors and early fault diagnosis capability are realized, but also the reliability, explainability and low-consumption operation of the model are ensured through physical constraints and feature focusing, laying an important intelligent sensing foundation for autonomous optimization and predictive maintenance of the entire pump control system.

[0060] Embodiment one: normal operation and optimized scheduling scenario In a large municipal rainwater lifting pump station, the control system of the application is deployed to manage three submersible pumps with a rated flow of 500 cubic meters per hour. The pump station is operating at a total target flow of 1200 cubic meters per hour according to the instructions of the dispatch center.

[0061] After the system starts, the embedded controllers corresponding to the three pumps begin to work. The signal processing and virtual sensor module synchronously acquires the three-phase current, voltage, speed of each pump and the vibration signal of the key position of the pump body, while the electromagnetic flowmeter on the main pipeline measures the actual total flow as 1202 cubic meters per hour. The module calculates real-time features such as the instantaneous active power of No. 1 pump as 85 kilowatts, the synthesized vibration effective value as 2.5 meters per square second, and the current total harmonic distortion rate as 5%, and through its built-in lightweight AI model, estimates the virtual displacement of No. 1 pump as 403 cubic meters per hour, while diagnosing its fault probability as extremely low, finally outputs the calibrated virtual displacement value as 402 cubic meters per hour, the health status vector is good, and the confidence is 95%. No. 2 pump and No. 3 pump also complete similar calculations.

[0062] Meanwhile, the physical information constraint module receives the electrical feature vector of No. 1 pump, and its built-in physical information neural network calculates the current hydraulic efficiency as 82%, the total mechanical loss power as 8 kW, and a medium state change factor as 1.05 according to the data of active power, rotating speed, etc., indicating that the current medium density is slightly higher than that of clean water. The adaptive control module immediately adjusts the parameters of the PID controller according to the medium factor of 1.05 and the efficiency of 82%, and reduces the proportional gain base value by about 5% to respond more smoothly to the changes in the flow set value. The local health assessment module then calculates the current health score of No. 1 pump as 92 points by comprehensively considering its low fault risk, current efficiency and smooth load fluctuation.

[0063] At the cluster level, the dynamic cluster management module starts to work. The health scores of the three controllers are 92, 88 and 90 respectively, and the network delay is very low. They periodically calculate their own dynamic weights, among which the weight value of No. 1 pump controller is the highest due to its highest health score and good calculation resource margin. Through distributed consensus election, No. 1 pump controller is successfully elected as the master controller of this round. Then, the distributed collaborative decision-making module starts to work. The master controller (No. 1 pump) initiates the collaborative decision-making process, and each pump exchanges marginal benefit information based on its own energy consumption characteristics (such as the efficiency of No. 2 pump being slightly lower, its unit flow power consumption being slightly higher) and health status, and after several rounds of iterative negotiation, a Nash equilibrium is finally reached to form a load distribution scheme: No. 1 pump undertakes 410 cubic meters per hour, No. 2 pump undertakes 395 cubic meters per hour, and No. 3 pump undertakes 395 cubic meters per hour. This scheme meets the total flow requirement while minimizing the overall energy consumption and balancing the equipment wear. The master controller issues the scheme to each pump, and the adaptive control module of each pump generates a rotating speed instruction accordingly to drive the frequency converter to adjust the motor operation. The whole system realizes optimal operation with high efficiency, stability and consideration of equipment life without human intervention.

[0064] Embodiment Two: Fault Tolerance and Adaptive Reconstruction Scenario In another mine drainage application scenario, four submersible pumps are operated in parallel. During the system operation, the impeller of No. 2 pump is suddenly partially wrapped by foreign matter, and at the same time, the controller of No. 3 pump, which currently serves as the master controller, causes the network to disconnect due to a sudden fault of its internal communication subboard.

[0065] After the failure occurs, the signal processing and virtual sensing module first captures the anomaly. The sensor network of pump 2 detects that its current harmonic distortion rate jumps from the normal 4% to 15%, and the high-frequency vibration energy also increases significantly, while the virtual displacement estimation value drops by 15% from the rated value. The lightweight fault diagnosis model calculates the fault probability of "impeller winding and deposition" to rise to 75% according to these feature changes. The module then reduces the confidence of its virtual displacement output to 65% and generates a health state feature vector containing high fault risk. After receiving this information, the local health assessment module, combined with its performance degradation indicators, quickly reduces the health score of pump 2 from 85 to 60.

[0066] Almost at the same time, the dynamic cluster management module detects that the heartbeat packet of pump 3 controller is lost. After not receiving its state for three consecutive periods, the remaining controllers in the cluster mark it as suspected failure. The original master controller is lost, triggering a new election process. The remaining 1, 2, and 4 controllers rebroadcast their dynamic weights. Among them, the 1 controller has the highest health score of 90 and has a good network state, and is elected as the new master controller. The smooth control right transfer unit starts immediately, and the 1 controller obtains the current load distribution state snapshot from each pump to ensure the continuity of control commands, completing the seamless handover of master control right.

[0067] The new master controller (pump 1) reinitiates collaborative decision-making according to the updated cluster state (pump 2 health score drops sharply). In the calculation of the distributed collaborative decision-making module, pump 2 has a high fault risk and low health score, which produces a significant "health cost" in the local benefit function, causing it to automatically tend to reduce the load in negotiation. The new allocation scheme is quickly reached: the load of pump 2 is significantly reduced to its safe operation lower limit, only bearing the minimum flow required to maintain cooling, while pumps 1 and 4 increase their loads accordingly to compensate for the total flow. At the same time, the system as a whole sends warning information about the winding failure of pump 2 and the controller switch to the cloud monitoring center. At the same time, the adaptive control module of pump 2 triggers a strategy switch according to its health state feature vector, from "adaptive PID control" to "robust PI control", reducing the control gain to avoid exacerbating the failure caused by rapid speed fluctuations. When facing double anomalies of equipment failure and controller failure, the whole system realizes stable degradation operation through intelligent sensing, active decision-making and architecture reconstruction, avoiding unplanned downtime and winning valuable response time for the maintenance personnel.

[0068] The embodiments of the application are described in detail above with reference to the drawings, but the application is not limited to the above-described embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the purpose of the application.

Claims

1. A pumping capacity control system based on a submersible pump, characterized in that: include: A submersible pump unit, comprising at least two submersible pumps operating in parallel; The sensor network is installed on each submersible pump to collect multi-dimensional operating signals of the corresponding submersible pump in real time, including electrical signals, speed signals and vibration signals. The sensor network also includes a flow sensor, which is installed on the pump set output main pipeline to measure the actual output flow of the system. An embedded controller cluster consists of multiple embedded controllers, each corresponding to a submersible pump, and the embedded controllers are interconnected through an industrial communication network. The dynamic cluster management module is used to assign dynamic weights to each embedded controller based on its local health score, network status, and topology location, and to dynamically elect a master controller in a decentralized manner, thereby realizing the dynamic reconstruction of the control cluster structure.

2. The pumping capacity control system based on a submersible pump according to claim 1, characterized in that: All embedded controllers have the same built-in software function modules, which include: The signal processing and virtual sensing module is used to perform fusion analysis based on the operating signal and actual output flow rate through a lightweight machine learning model to calculate the real-time virtual displacement and health status characteristics of the corresponding submersible pump. The physical information constraint module has a built-in lightweight physical information neural network that integrates the pump hydraulic mechanism equation and the motor mechanism equation. It is used to calculate the virtual physical quantities of the corresponding submersible pump by taking the easily measurable electrical quantities in the operating signal as input. The adaptive control module is used to dynamically adjust the parameters of the PID controller based on real-time virtual displacement, virtual physical quantities and preset displacement targets, and generate speed control commands to drive the motor of the corresponding submersible pump. The local health assessment module is used to calculate the local health score of the corresponding controller based on health status characteristics, dynamic load parameters and historical data of the equipment. The distributed collaborative decision-making module is used to generate a globally optimal load allocation scheme through distributed negotiation within an embedded controller cluster, based on a cooperative game algorithm, real-time status information of each submersible pump, and a simplified system hydraulic model.

3. The pumping capacity control system based on a submersible pump according to claim 2, characterized in that: The signal processing and virtual sensing module, when in operation, specifically includes: S11: Data acquisition and synchronization, acquiring multi-dimensional operating signals of the submersible pump and total system flow signal, and performing time synchronization based on timestamps; S12: Signal preprocessing and feature engineering, which cleans and transforms the running signal, extracts time-domain, frequency-domain and time-frequency-domain features, and constructs highly representative feature vectors; S13: Virtual flow estimation. The feature vector is input into a lightweight machine learning model. The reading of the flow sensor under standard operating conditions is used as the supervision benchmark to calculate the real-time virtual discharge of the corresponding submersible pump. S14: Online fault diagnosis, based on feature vectors, uses a lightweight fault diagnosis model to identify specific fault modes and health degradation trends of the pump in real time. S15: Confidence fusion and output, assess the reliability of virtual displacement, and combine virtual displacement with fault diagnosis results to generate the final virtual displacement output value and multi-dimensional health status feature vector.

4. The pumping capacity control system based on a submersible pump according to claim 3, characterized in that: The physical information neural network in the physical information constraint module adopts an encoder-decoder structure, including: A11: Input layer, used to receive a normalized electrical feature vector containing active power, apparent power, power factor, and speed; A12: Encoder network, consisting of multiple fully connected layers, used to extract high-dimensional latent features from input features; A13: Physical constraint fusion layer, which is a hidden layer of the network. The design constraints of its neuron activation function integrate the mechanistic relationship between pump head, flow rate, shaft power and motor electromagnetic torque. A14: Output layer, outputs virtual physical quantities, including real-time hydraulic efficiency, total mechanical loss power, and dimensionless factors characterizing changes in the state of the medium.

5. A pumping capacity control system based on a submersible pump according to claim 4, characterized in that: During the training of the physical constraint fusion layer, the composite physical loss function applied is defined as the sum of the residuals of the mechanistic equations: ； in, For composite physical loss function, and These are weighting coefficients. The pump hydraulic mechanism loss term is defined as: ； in, For the pump shaft power implicitly derived by the network, For standard medium density, It is the acceleration due to gravity. The head is derived by the network based on the similarity law, the current rotational speed, and the medium factor. To predict the pump's flow rate under current operating conditions, The pumping efficiency predicted by the network; The motor mechanism loss term is defined as: ； in, The electromagnetic torque is calculated from the input electrical quantity based on motor theory. Angular velocity, This represents the total mechanical loss torque output by the network.

6. A pumping capacity control system based on a submersible pump according to claim 5, characterized in that: The adaptive control module specifically includes: A21: Parameter dynamic adjustment unit, used to calculate the basic parameter set of the PID controller in real time based on the medium state change factor and hydraulic efficiency in the virtual physical quantity; A22: Control performance evaluation and fine-tuning unit, used to fine-tune and optimize the basic parameter set online based on the tracking error and error change rate characteristics of the virtual displacement; A23: Fault state control strategy switching unit, used to switch between normal adaptive PID control, degraded robust PID control and safe shutdown protocol based on health state feature vector; A24: Speed ​​command generation unit, used to calculate and output speed control commands for driving the submersible pump motor based on the final determined PID parameters and preset displacement target.

7. A pumping capacity control system based on a submersible pump according to claim 6, characterized in that: The local health assessment module specifically includes: A31: Multidimensional health indicator calculation unit, used to process health status feature vectors, virtual physical quantities and real-time dynamic load parameters, and calculate standardized basic health indicators; A32: Historical data analysis and trend extraction unit, used to access historical operating data of equipment and calculate health degradation trend indicators based on time series analysis; A33: Comprehensive Health Calculation Unit, used to calculate the final health score by integrating basic health indicators and trend indicators through a weighted evaluation model.

8. A pumping capacity control system based on a submersible pump according to claim 7, characterized in that: The distributed collaborative decision-making module specifically includes: A41: Local benefit calculation unit, used to construct and calculate the local benefit function based on the real-time status, health and energy consumption characteristics of this pump; A42: Hydraulic model calculation unit, which contains a simplified system hydraulic model to evaluate the changes in the operating point of this pump and its impact on the head-flow rate of the main pipeline under different load distribution schemes; A43: Neighbor information processing unit, used to periodically receive and parse the status and decision proposal information of neighbor controllers through the communication network; A44: Distributed negotiation engine, based on cooperative game theory algorithm, uses local information, neighbor information and hydraulic model calculation results to perform iterative calculation to seek the Nash equilibrium point of the cluster; A45: Decision Output and Consistency Verification Unit, used to output the final speed setpoint of this pump after negotiation convergence, and to perform final consistency verification with other nodes in the cluster.

9. A pumping capacity control system based on a submersible pump according to claim 8, characterized in that: The dynamic cluster management module specifically includes: A51: Dynamic weight calculation unit, used to periodically calculate the dynamic weight value of this node. Specifically, the dynamic weight value integrates the local node's health, computing resource status, network connection quality, and topology location information. A52: Cluster Status Awareness and Synchronization Unit, used to maintain a dynamic weight list and role information of all online nodes in the cluster through periodic heartbeat messages and status broadcasts; A53: Distributed consensus election unit, which adopts a competition-response mechanism based on dynamic weights to enable the cluster to reach a consensus on the identity of the master controller in the absence of a central coordinator; A54: Smooth Control Transfer Unit, used to manage and synchronize the transfer and control process related to control strategies and historical status data when the main controller changes; A55: Anomaly Handling and Cluster Reconstruction Unit, used to detect node failures and network partition anomalies, and trigger cluster topology reconstruction and weight recalculation.

10. A pumping capacity control system based on a submersible pump according to claim 9, characterized in that: The workflow of the distributed consensus election unit is a priority-based broadcast-response mechanism, specifically including the following steps: S81: Each node periodically broadcasts its heartbeat message; S82: If a node detects that its own weight is higher than the weight of the currently known master controller and the difference exceeds a threshold, it broadcasts a master controller declaration message to the cluster. S83: During the preset competition window, if a node receives a master controller declaration from another node with higher weight, the node withdraws its own declaration and becomes a slave controller. If it does not receive a declaration with higher weight, the node officially becomes the master controller after the window ends. S84: The new master controller broadcasts a role confirmation message, and all slave controllers update the master controller information they maintain and send an confirmation response to the master controller.

Images