Intelligent water pump controller with built-in AI chip and control method

Abstract

The invention discloses an intelligent water pump controller with a built-in AI chip and a control method, and particularly relates to the technical field of intelligent control. On the basis of a prediction result, an optimal control action sequence and physical constraints in a physical entity relation graph, a pressure, a flow reference value and a frequency limit value of each timestamp in a future stage are generated, so that the service life of equipment is prolonged to a great extent, and the energy loss is reduced; pressure and flow reference values of each timestamp in a future stage in a pre-control strategy are taken as set standards, sensor data at the current moment in a clean data set are called to perform fuzzy PID control adjustment, then constraint correction is performed based on a physical entity relationship and a frequency limit value generated by the pre-control strategy, a control instruction is generated, and the control instruction is controlled to be in a fuzzy control state. The matching degree of the control instruction and the current dynamic regulation and control of the water pump is guaranteed, and the regulation and control response time delay is reduced.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and more specifically, to an intelligent water pump controller and control method with a built-in AI chip. Background Technology

[0002] Water pumps are widely used in modern industry, agriculture, and daily life. They are indispensable in water supply systems, irrigation systems, sewage treatment, and cleaning. Traditional water pump control methods are mainly based on simple manual operation or automatic control based on fixed thresholds. In manual control, operators need to manually start or stop the pump according to actual needs and site conditions. While this method is simple and direct, it requires real-time human monitoring and operation, which is not only inefficient but also prone to human error. This can lead to untimely or unreasonable pump start-up and shutdown, resulting in energy waste or failure to meet actual water demand. Therefore, it is necessary to upgrade water pump control methods to better meet water demand and reduce energy loss.

[0003] Existing water pump control methods utilize basic sensors such as current, voltage, and temperature, along with acoustic or vibration monitoring modules, to achieve multi-dimensional sensing capabilities of the water pump's operating status. Based on a preset speed control curve, a PID feedback regulation mechanism is employed to dynamically adjust the output frequency according to the measured values ​​from pressure or flow sensors, thereby achieving dynamic control of the water pump's operation. In terms of anomaly handling, a static threshold-triggered protection mechanism is set up. When the monitoring results reach the preset upper or lower threshold, fault protection is triggered, which largely ensures the stability of water supply, improves the safety of water pump control, and reduces energy consumption.

[0004] However, existing methods still have some problems: control relies on current state feedback and does not demonstrate demand prediction capabilities; when pipeline pressure changes abruptly, the control response delay is too long, which is detrimental to ensuring water supply stability; the monitored acoustic, vibration, and temperature data are processed independently, lacking a spatiotemporal alignment mechanism, making it impossible to identify early bearing fault characteristics of the "high vibration + specific frequency domain acoustic" combination, and making timely fault maintenance difficult. Therefore, it is necessary to further upgrade the predictive and fault diagnosis capabilities of pump control methods, reduce response delays, and improve the accuracy of fault diagnosis. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides an intelligent water pump controller and control method with built-in AI chip to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an intelligent water pump controller with a built-in AI chip, comprising: Pump data acquisition module: used to collect pump operating data, environmental data, and equipment status data; Multi-source data fusion module: The collected data is cleaned, spatiotemporally aligned, feature extracted, dimensionality reduced, and knowledge graph constructed to obtain a clean dataset, anomaly logs, spatiotemporally synchronized data matrix, standardized feature vector set, low-dimensional core feature vector set, physical entity relationship graph, and parameter association strength matrix. Pre-control strategy generation module: predicts future water demand and remaining lifespan of key components, generates optimal control action sequence, and generates pre-control strategy based on prediction results, optimal control action sequence, and physical constraints in physical entity relationship graph; Real-time control command generation module: It uses the pre-control strategy as the setting standard, retrieves the sensor data of the current moment in the clean dataset for fuzzy PID control adjustment, then performs constraint correction based on physical entity relationships, and generates control commands; Fault handling control module: It uses the isolated forest algorithm to detect anomalies in the real-time data stream, triggers fault diagnosis after confirming data anomalies, and performs fault classification and emergency handling classification response based on the fault diagnosis results; Control performance index collection module: used to collect predictive control performance indexes, real-time control performance indexes, and fault handling control performance indexes within a preset period; Control effect evaluation module: Evaluates whether the predictive control effect, real-time control effect, and fault handling control effect meet expectations in sequence. If they all meet expectations, the comprehensive control effect index is calculated and automatically stored; otherwise, the control effect type that does not meet expectations is output.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a smart water pump control method with a built-in AI chip, comprising the following steps: S1. Collect pump operation data, environmental data, and equipment status data; S2. The collected data is cleaned, spatiotemporally aligned, feature extracted, dimensionality reduced, and knowledge graph constructed to obtain a clean dataset, anomaly logs, spatiotemporally synchronized data matrix, standardized feature vector set, low-dimensional core feature vector set, physical entity relationship graph, and parameter association strength matrix. S3. Predict future water demand and remaining lifespan of key components, generate optimal control action sequence, and generate pre-control strategy based on prediction results, optimal control action sequence, and physical constraints in physical entity relationship graph. S4. Using the pre-control strategy as the set standard, retrieve the sensor data at the current moment in the clean dataset for fuzzy PID control adjustment, then perform constraint correction based on physical entity relationships, and generate control commands. S5. Use the isolated forest algorithm to detect anomalies in the real-time data stream. Once an anomaly is confirmed, trigger fault diagnosis and classify the fault and emergency response levels based on the fault diagnosis results. S6. Collect predictive control performance indicators, real-time control performance indicators, and fault handling control performance indicators within a preset period. S7. Evaluate whether the predictive control effect, real-time control effect, and fault handling control effect meet expectations in sequence. If they all meet expectations, calculate the comprehensive control effect index and store it automatically. Otherwise, output the control effect type that does not meet expectations.

[0008] The technical effects and advantages of this invention are as follows: 1. A pre-control strategy generation module is set up to input the flow time series data in the spatiotemporal synchronized data matrix, the environmental features in the standardized feature vector set, and the calendar features in the physical entity relationship graph into the constructed LSTM network model for load prediction, and output the demand water flow value for each time point in the future stage. The low-dimensional core feature vector set, the vibration features in the standardized feature vector set, and the equipment operation history in the physical entity relationship graph are input into the constructed XGBoost regression model for equipment life prediction, and output the remaining life of key equipment. Based on the energy consumption optimization objective, the optimal control action sequence is generated. Based on the load prediction results, the optimal control action sequence, the equipment life prediction results, and the physical constraints in the physical entity relationship graph, the pressure, flow reference values, and frequency limit values ​​for each time point in the future stage are generated, which greatly reduces the control response delay, extends the equipment life, and reduces energy consumption.

[0009] 2. A multi-source data fusion module is set up to perform data cleaning, spatiotemporal alignment, feature extraction, data dimensionality reduction, and knowledge graph construction on the collected data to obtain a clean dataset, anomaly logs, spatiotemporally synchronized data matrix, standardized feature vector set, low-dimensional core feature vector set, physical entity relationship graph, and parameter association strength matrix. A fault handling control module is set up to detect anomalies in the real-time data stream using the isolated forest algorithm. After confirming data anomalies, fault diagnosis is triggered. Based on the fault diagnosis results, fault classification and emergency response are performed in a graded manner. It can identify hidden faults from the physical relationships of multi-source data, improve the accuracy of fault diagnosis, and initiate graded emergency response based on the fault level, avoiding the drawbacks of direct shutdown in traditional solutions. Attached Figure Description

[0010] Figure 1 This is a system structure block diagram of the present invention.

[0011] Figure 2 This is a diagram illustrating the method steps of the present invention. Detailed Implementation

[0012] 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.

[0013] like Figure 1 The embodiment shown provides an intelligent water pump controller with a built-in AI chip, including a water pump data acquisition module, a multi-source data fusion module, a pre-control strategy generation module, a real-time control command generation module, a fault handling control module, a control effect index collection module, a control effect evaluation module, and a database. The water pump data acquisition module is connected to the multi-source data fusion module and the fault handling control module. The multi-source data fusion module is connected to the pre-control strategy generation module, the real-time control command generation module, and the fault handling control module. The pre-control strategy generation module is connected to the real-time control command generation module. The pre-control strategy generation module, the real-time control command generation module, and the fault handling control module are all connected to the control effect index collection module. The control effect index collection module is connected to the control effect evaluation module. All modules in the system are connected to the database.

[0014] Pump data acquisition module: used to collect pump operating data, environmental data, and equipment status data; Furthermore, the pump's operating data includes pressure data, temperature data, flow data, vibration data, and electrical data; environmental data includes ambient temperature and humidity data and meteorological data; and equipment status data includes motor status data, valve status data, and energy consumption data.

[0015] Specifically, in this embodiment, the pressure data includes pump inlet and outlet pressures and pressures at key nodes in the pipeline network; the temperature data includes motor winding temperature, bearing temperature, conveying medium temperature, and coolant temperature; the flow data includes instantaneous flow rate, cumulative flow rate, and flow direction; the vibration data includes three-dimensional vibration acceleration and vibration spectrum; the electrical data includes the effective value of three-phase current, the effective value of voltage, active power, and power factor; the ambient temperature and humidity data includes ambient temperature and relative humidity; the meteorological data includes atmospheric pressure and rainfall; the motor status data includes speed and direction of operation; the valve status data includes opening percentage and response delay; and the energy consumption data refers to real-time electrical energy consumption.

[0016] Specifically, in this embodiment, pressure data is collected through a pressure sensor, temperature data through a temperature sensor, flow data through a flow sensor, vibration data through a vibration sensor, electrical data through a power quality analyzer, ambient temperature and humidity data through a digital temperature and humidity sensor, atmospheric pressure through an atmospheric pressure sensor, rainfall through a rain gauge, motor speed through a Hall effect speed sensor, motor running direction through a dual-channel quadrature encoder, valve opening percentage through an absolute encoder, and real-time power consumption through a smart meter.

[0017] Multi-source data fusion module: The collected data is cleaned, spatiotemporally aligned, feature extracted, dimensionality reduced, and knowledge graph constructed to obtain a clean dataset, anomaly logs, spatiotemporally synchronized data matrix, standardized feature vector set, low-dimensional core feature vector set, physical entity relationship graph, and parameter association strength matrix. Furthermore, the multi-source data fusion module includes a data receiving unit, a data cleaning unit, a spatiotemporal alignment unit, a feature extraction unit, a data dimensionality reduction unit, a knowledge graph construction unit, and an automatic data storage unit. The data receiving unit receives the collected water pump data; the data cleaning unit removes outliers from the received data based on the 3σ principle to obtain a clean dataset and anomaly-marked logs; the spatiotemporal alignment unit performs time synchronization processing, spatial coordinate transformation, and data resampling on the clean dataset to obtain a spatiotemporal synchronized data matrix; the feature extraction unit performs temporal feature calculation, windowing preprocessing, and FFT on the spatiotemporal synchronized data matrix. The transformation, frequency domain feature extraction, and standardization processes yield a standardized feature vector set. The data dimensionality reduction unit performs PCA and t-SNE dimensionality reduction on the standardized feature vector set to obtain a low-dimensional core feature vector set. The knowledge graph construction unit defines physical entity relationships and calculates association strength on the spatiotemporal synchronized data matrix to obtain a physical entity relationship graph and a parameter association strength matrix. The automatic data storage unit automatically stores the clean dataset, anomaly labeling log, spatiotemporal synchronized data matrix, standardized feature vector set, low-dimensional core feature vector set, physical entity relationship graph, and parameter association strength matrix into the database and marks the update timestamp.

[0018] In this embodiment, the specific steps for data cleaning are as follows: A1. Data Grouping: The collected data is grouped according to the type of sensor used, and each group is processed independently to avoid threshold interference between different types of acquisition tools; A2. Time window slicing: The fixed time window is 60 seconds, and the window overlap rate is controlled at 50%; A3. Dynamic threshold calculation: Calculate the window mean μ a The specific formula is as follows: n1 is the number of sampling points within the time window, x i For the original measurement value of the i-th sampling point, calculate the window standard deviation σ. a The specific formula is as follows: The threshold range is [μ a -3σ a ,μ a +3σ a ]; A4. Outlier identification: Mark data points that exceed the threshold range; A5. Anomaly Handling: Three consecutive points exceeding 4σ a Data points are deleted directly within the specified range; a single point exceeding 3σ is excluded. a But less than 4σ a The replacement value x at time t is obtained by replacing the exponentially weighted moving average. t The specific calculation formula for ' is: a1 is the exponential weighting coefficient, and the specific calculation formula is as follows: N eff For the effective observation window length, x t x is the original measurement value at time t. t-1 'This is the replacement value of the exponentially weighted moving average at time t-1, which is numerically equivalent to the normal original measurement at time t-1; A6. Interpolation Filling: Locates the missing point. For single-point missing points, linear interpolation is used. The specific formula is as follows: x t * x t+1 x t-1 The values ​​are, in order, the missing point imputation value at time t, the valid point measurement value at time t+1, and the valid point measurement value at time t-1. For two consecutive missing points, quadratic polynomial interpolation is used, and the specific formula is as follows: (t0,x0)(t1,x1)(t2,x2) are the three nearest valid points before and after the missing point, and a4=x0. , For three to five consecutive missing points, cubic spline interpolation is used; for more than five missing points, historical pattern matching is used; and for boundary missing points, nearest neighbor copying is used. A7. Cross-sensor verification: Perform pressure-flow relationship verification, current-vibration relationship verification, and temperature-power relationship verification respectively. If the physical constraints are met, output a clean dataset; otherwise, mark it as invalid and generate an anomaly log.

[0019] In this embodiment, it is specifically noted that spatiotemporal alignment includes time synchronization processing, spatial coordinate transformation, and data resampling. The output is a spatiotemporal synchronization data matrix and a sensor position mapping table. The specific execution steps of the time synchronization processing are as follows: The master clock (edge ​​AI chip) broadcasts a Sync message every second, and the slave clocks (each sensor) record the reception time t2; the master clock sends a FollowUp message carrying the precise transmission time t1, the slave clock sends a DelayReq message and records the transmission time t3, and the master clock replies with a DelayResp message carrying the reception time t4; timestamp correction is performed, and the specific formula is: , Δt is the master / slave clock offset, d msr Main → From propagation delay, d msd For the propagation delay from master to slave, t raw For the original timestamp, t corrected For the corrected timestamp. The specific steps of spatial coordinate transformation are as follows: Mark the intersection of the pump axis and the outlet flange as the coordinate origin, mark the flow direction along the outlet pipe as the Z-axis, mark the vertical upward direction as the X-axis, and determine the Y-axis using the right-hand rule to establish a unified coordinate system; measure the installation position of each sensor and obtain the sensor coordinates [x...]. local ,y local ,z local ] T Calculate the homogeneous transformation matrix , where R 3×3 Let T be the rotation matrix. 3×1 Let x be the translation vector. local ,y local ,z local The following are the coordinate values ​​along the X, Y, and Z axes marked on the sensor's housing in its own coordinate system, where θ represents the sensor's mounting angle deflection, and t... x t y t z The translations are in the X, Y, and Z directions, respectively, with the last term 1 representing the homogeneous coordinate identifier; the coordinate mapping equation is as follows: x, y, z represent the offsets of the sensor measurement points in the X, Y, and Z directions within the sensor's own coordinate system, while x', y', and z' represent the coordinates in the X, Y, and Z directions within a unified coordinate system. Data resampling includes target frequency determination and cubic spline interpolation processing.

[0020] In this embodiment, the specific steps for feature extraction are as follows: input the spatiotemporal synchronization data matrix, calculate the temporal features for each time window, and output the temporal feature vector set V. time The Hanning window function w(n) is applied for windowing preprocessing, specifically as follows: If n2 is the number of sampling points within the time window, then the windowed signal x w (n) = x(n) × w(n), where n is the time-domain sampling point index; The transformed value X is obtained by performing a Fast Fourier Transform (FFT) on the windowed signal. k The specific formula is as follows: k is the frequency index, i is the imaginary unit, and the amplitude spectrum A is calculated. k The specific formula is as follows: k=1,2,...,n² / 2-1; calculate frequency domain features and output a set of frequency domain feature vectors; extract key time domain features and key frequency domain features based on pump fault types to obtain the target feature vector combination V. feature (12-dimensional time-domain features + 8-dimensional frequency-domain features. The 12-dimensional time-domain features include peak value, peak-to-peak value, mean, standard deviation, root mean square value, skewness, kurtosis, waveform factor, impulse factor, margin factor, zero-crossing rate, and envelope energy. The 8-dimensional frequency-domain features include the main frequency amplitude ratio, spectral centroid, spectral variance, 1-3 times blade frequency energy, bearing characteristic band ratio, 50 / 100Hz harmonic ratio, 0.5-2kHz band entropy, and fundamental frequency sideband energy difference); After standardizing the target feature vector, we have v norm =(v-μ v ) / σ v ,μ v σ is the historical mean of this feature. v The historical standard deviation of this feature is given; the standardized feature vector set is output.

[0021] In this embodiment, it is specifically necessary to explain that data dimensionality reduction includes PCA dimensionality reduction and t-SNE dimensionality reduction, ultimately outputting a low-dimensional core feature vector (8-dimensional principal components). PCA dimensionality reduction reduces the original high-dimensional data to an intermediate dimension, aiming to remove linear correlations, reduce dimensionality, and retain most of the information. The execution process is as follows: calculate the covariance matrix; solve for the eigenvalues ​​and eigenvectors of the covariance matrix; select the top k principal components based on the cumulative contribution rate; project the standardized data onto the selected principal components to obtain the PCA-reduced data. t-SNE dimensionality reduction is performed based on the PCA dimensionality reduction result (non-linear dimensionality reduction) to further reduce to a lower dimension, so as to maintain the local structure of the high-dimensional data (especially the similarity of neighboring points) in the low-dimensional space. The execution process is as follows: set the perplexity, calculate the Gaussian kernel width for each point; calculate the high-dimensional similarity matrix; randomly initialize the low-dimensional representation; calculate the low-dimensional similarity, calculate the gradient, and update the low-dimensional representation.

[0022] In this embodiment, it is specifically noted that the final output of the knowledge graph construction is a physical entity relationship graph, a parameter association strength matrix, and a dynamic update log. The specific execution flow of knowledge graph construction is as follows: Input cleaned and aligned feature data, identify physical entities (such as water pumps, motors, valves, and pipes), extract parameter attributes (such as water pump entity attributes: outlet pressure, flow rate, efficiency, vibration amplitude; such as motor entity attributes: current, speed, winding temperature, power factor), and output an entity-attribute mapping table; after determining the physical law constraint equations, analyze the data driving relationships and output an entity relationship topology graph (RDF triples). The physical law constraint equations include, but are not limited to, Bernoulli's equation, the law of conservation of energy, the continuity equation, Joule's law, the water pump characteristic equation, the electromechanical coupling vibration equation, and the thermodynamic equilibrium equation. The relationship types include driving relationships, generation relationships, constraint relationships, and causal relationships; calculate the association strength and output the association strength matrix, such as the pressure-flow association strength R. PQ The specific calculation formula is as follows: cov(P,Q) is the covariance between pressure and flow rate, σ P σ represents the standard deviation of pressure. Q For the standard deviation of the flow rate, Let a5 be the partial derivative of pressure with respect to flow rate, and let a5 be the physical weight coefficient, which is a value taken from industrial experience; determine the knowledge graph storage structure; and dynamically update the knowledge graph based on the incremental Bayesian formula.

[0023] Pre-control strategy generation module: predicts future water demand and remaining lifespan of key components, generates optimal control action sequence, and generates pre-control strategy based on prediction results, optimal control action sequence, and physical constraints in physical entity relationship graph; Furthermore, the pre-control strategy generation module includes a data retrieval unit, a load forecasting unit, an equipment lifespan forecasting unit, an energy consumption optimization unit, a pre-control strategy generation unit, and an automatic data storage unit. The data retrieval unit retrieves the clean dataset, the spatiotemporal synchronized data matrix, the standardized feature vector set, the low-dimensional core feature vector set, the physical entity relationship graph, and the parameter correlation strength matrix. The load forecasting unit inputs the flow time series data from the spatiotemporal synchronized data matrix, the environmental features from the standardized feature vector set, and the calendar features from the physical entity relationship graph into the constructed LSTM network model for load forecasting, outputting the required water flow value for each time stamp in the future stage. The equipment lifespan forecasting unit inputs the low-dimensional core... Vibration features from the core feature vector set and the standardized feature vector set, as well as the equipment operation history from the physical entity relationship graph, are input into the constructed XGBoost regression model for equipment life prediction, outputting the remaining life of key equipment; the energy consumption optimization unit generates the optimal control action sequence based on the energy consumption optimization objective; the pre-control strategy generation unit generates pressure, flow reference values, and frequency limit values ​​for each time stamp in the future stage based on the load prediction results, the optimal control action sequence, the equipment life prediction results, and the physical constraints in the physical entity relationship graph; the automatic data storage unit automatically stores the generated pressure, flow reference values, and frequency limit values ​​for each time stamp in the future stage into the database and marks the update timestamp.

[0024] In this embodiment, the specific steps for the energy consumption optimization unit to generate the optimal control action sequence are as follows: Retrieve the current operating state, flow prediction sequence, remaining equipment lifespan, low-dimensional core features, physical constraint model, and energy consumption calculation parameters; Define the state space based on the current operating state, flow prediction sequence, and low-dimensional core features, and construct a state transition model; Define action constraints based on equipment lifespan and physical constraints, and define the action space; Design a reward function based on the state transition model and action space; Perform value iteration solution based on the Bellman optimal equation, and record the optimal action sequence.

[0025] Real-time control command generation module: It uses the pre-control strategy as the setting standard, retrieves the sensor data of the current moment in the clean dataset for fuzzy PID control adjustment, then performs constraint correction based on physical entity relationships, and generates control commands; Furthermore, the real-time control command generation module includes a data retrieval unit, a feedforward setting unit, a PID control trigger determination unit, a PID control adjustment calculation unit, an adjustment correction unit, and a control command generation unit. The data retrieval unit retrieves real-time pressure and real-time flow from the pre-control strategy and the cleaning dataset. The feedforward setting unit marks the pressure and flow reference values ​​for each time point in the future stage in the pre-control strategy as the pressure and flow setpoints for each time point in the future stage. The PID control trigger determination unit triggers pressure PID control when the real-time pressure is inconsistent with the pressure setpoint, and triggers flow PID control when the real-time flow is inconsistent with the flow setpoint. The PID control adjustment calculation unit calculates the inverter frequency adjustment using a fuzzy PID algorithm when pressure PID control is triggered, and calculates the valve opening adjustment using a fuzzy PID algorithm when flow PID control is triggered. The adjustment correction unit constrains and corrects the calculated inverter frequency adjustment or valve opening adjustment based on the physical entity relationship and the frequency limit value generated by the pre-control strategy. The control command generation unit generates actuator control commands based on the corrected inverter frequency adjustment or valve opening adjustment.

[0026] In this embodiment, the specific steps for calculating the inverter frequency adjustment using the fuzzy PID algorithm are as follows: Calculate the pressure setpoint P at time t. t,set With real-time value P t,c error e t,P and error change rate de t,P / dt; Calculate the frequency adjustment Δf of the inverter at time t. t The specific formula is as follows: ,K p1 K i1 K d1 Based on the calculated pressure error e t,P and particle size error change rate de t,P / dt performs fuzzy inference and defuzzification, then updates the PID parameters in real time. The specific steps for calculating the valve opening adjustment using the fuzzy PID algorithm are as follows: Calculate the flow setpoint Q at time t. t,set With real-time value Q t,c error e t,Q and error change rate de t,Q / dt; Calculate the valve opening adjustment Δv at time t. t The specific formula is as follows: ,K p2 K i2 K d2 To calculate the flow error e t,Q and flow error change rate de t,Q / dt represents the PID parameters updated in real time after fuzzy inference and defuzzification.

[0027] Fault handling control module: It uses the isolated forest algorithm to detect anomalies in the real-time data stream, triggers fault diagnosis after confirming data anomalies, and performs fault classification and emergency handling classification response based on the fault diagnosis results; Furthermore, the fault handling control module includes a real-time data stream feature extraction unit, an anomaly detection unit, an anomaly determination unit, a fault diagnosis unit, a fault classification unit, and an emergency handling unit. The real-time data stream feature extraction unit is used to extract features from the real-time data stream. The anomaly detection unit uses the isolated forest algorithm to detect anomalies in the extracted real-time data stream features and outputs an anomaly score. The anomaly determination unit compares the anomaly score with a dynamic threshold; if the anomaly score is greater than the dynamic threshold, the feature is determined to be abnormal, triggering fault diagnosis. The fault diagnosis unit includes fault type identification based on convolutional attention networks and fault location localization based on Bayesian networks, generating a fault diagnosis report based on the diagnosis results. The fault classification unit divides the fault level into three levels: Level 1, Level 2, and Level 3. The emergency handling unit switches to energy-saving mode with optimized control when a Level 1 fault is triggered, switches to conservative mode with PID control when a Level 2 fault is triggered, and switches to safety mode with constant voltage control when a Level 3 fault is triggered.

[0028] Control performance index collection module: used to collect predictive control performance indexes, real-time control performance indexes, and fault handling control performance indexes within a preset period; Furthermore, predictive control performance indicators include average flow prediction error, average pressure prediction error, average RUL prediction error rate, energy consumption optimization coefficient per unit flow, and reference curve tracking compliance rate; real-time control performance indicators include pressure control deviation, flow control deviation, energy efficiency ratio compliance rate, and pressure fluctuation coefficient; fault handling control performance indicators include anomaly detection rate, average mode switching efficiency, and fault downtime percentage.

[0029] In this embodiment, it is specifically necessary to explain that the average error A in traffic prediction... YQ It is expressed as the average absolute percentage error between the predicted flow and the actual flow, and the specific calculation formula is as follows: In the formula N a Q ci Q si The parameters are, in order: the number of time sampling points for flow rate within a preset period, the actual flow rate value at the i-th time sampling point, and the predicted flow rate value at the i-th time sampling point; the average error A of pressure prediction. YP Expressed as the average absolute percentage error between predicted and actual pressure, the specific calculation formula is as follows: In the formula N b P ci P siThe parameters are, in order: the number of time sampling points for pressure within the preset period, the actual pressure value at the i-th time sampling point, and the predicted pressure value at the i-th time sampling point; the average RUL prediction error rate A. YS The specific calculation formula is as follows: N c RUL ci RUL si UL di The parameters are, in order: the number of critical equipment, the actual remaining service life of the i-th critical equipment, the model-predicted remaining service life, and the rated service life; the specific calculation formula for the energy consumption optimization coefficient AE per unit flow rate is as follows: E b E a The energy consumption per unit flow rate within the preset period and the average energy consumption per unit flow rate under existing pump control are, in order. The specific calculation formula for the reference curve tracking compliance rate AC is as follows: , t a t zy The parameters are, in order, the reference curve tracking time within the preset period and the total running time; the actual value is considered to be within ±5% of the reference value when it meets the standard; pressure control deviation B. P The specific calculation formula is as follows: P ei Set the pressure value P for the i-th time sampling point. ci Let i be the actual pressure value at the i-th time sampling point, i=1,2,...,N b Flow control deviation B Q The specific calculation formula is as follows: Q ei Set the flow rate value Q for the i-th time sampling point. ci Let i be the actual flow rate at the i-th time sampling point, i=1,2,...,N a The specific formula for calculating the energy efficiency ratio (BR) compliance rate is as follows: , t b The time required for the energy efficiency ratio to meet the target within a preset cycle, t zy The energy efficiency ratio (EER) is ≥3.5 for the total operating time within the preset cycle, which is considered satisfactory; pressure fluctuation coefficient B U The specific calculation formula is as follows: μ P For N b The average actual pressure at each time sampling point, P ci The actual pressure value at the i-th time sampling point; the anomaly detection rate C W The specific calculation formula is as follows: m a m b The numbers represent, in order, the actual number of anomalies detected within the preset period, and the actual number of anomalies not detected; average mode switching efficiency C. H The specific calculation formula is as follows: N d t y t qi The parameters are, in order: number of mode switching times within a preset period, theoretical minimum mode switching time, and actual mode switching time. The theoretical minimum mode switching time is the maximum of the minimum response time of the mechanical actuator and the minimum response time of the electrical system. The percentage of downtime due to faults is C. G The specific calculation formula is as follows: , t g t is the downtime due to fault within a preset period. zy This represents the total runtime within the preset period.

[0030] Control effect evaluation module: Evaluates whether the predictive control effect, real-time control effect, and fault handling control effect meet expectations in sequence. If they all meet expectations, the comprehensive control effect index is calculated and automatically stored. Otherwise, the control effect type that does not meet expectations is output. Furthermore, the control effect evaluation module includes a data receiving unit, a predictive control effect coefficient calculation unit, a real-time control effect coefficient calculation unit, a fault handling control effect coefficient calculation unit, a control effect judgment unit, a comprehensive control effect index calculation unit, and a data output unit. The data receiving unit is used to receive predictive control effect indicators, real-time control effect indicators, and fault handling control effect indicators within a preset period; the predictive control effect coefficient calculation unit is used to calculate the predictive control effect coefficient X. A The specific formula is as follows: A E A C A YQ A YP A YS The following are the unit flow energy consumption optimization coefficient, reference curve tracking compliance rate, average flow prediction error, average pressure prediction error, and average RUL prediction error rate, respectively. To avoid a denominator of 0, a compensation of 1 is added. The real-time control effect coefficient calculation unit is used to calculate the real-time control effect coefficient X. B The specific calculation formula is as follows: B R B P B Q B U The parameters are, in order: energy efficiency ratio compliance rate, pressure control deviation, flow control deviation, and pressure fluctuation coefficient. To avoid a denominator of 0, a compensation of 1 is added. The fault handling control effect coefficient calculation unit is used to calculate the fault handling control effect coefficient X. C The specific formula is as follows: C W C H C GThe parameters are, in order: anomaly detection rate, average mode switching efficiency, and fault downtime percentage. To avoid a denominator of 0, a compensation of 1 is added. The predictive control effect coefficient, real-time control effect coefficient, and fault handling control effect coefficient calculated by the control effect judgment unit are compared with their corresponding preset values. If the calculated values ​​are all greater than or equal to the preset values, the control effect is considered to meet expectations; otherwise, it is considered not to meet expectations. The comprehensive control effect index calculation unit calculates the comprehensive control effect index and automatically stores it in the database. The comprehensive control effect index Y... U The specific formula is: The data output unit is used to transmit control effects that do not meet expectations to the pump control center.

[0031] Database: Used to store data information for all modules in the system.

[0032] In this embodiment, it should be noted that the preset values ​​and set values ​​used are selected based on actual needs, and no specific value limitation is imposed here.

[0033] like Figure 2 The embodiment shown provides a smart water pump control method with a built-in AI chip, including the following steps: S1. Collect pump operation data, environmental data, and equipment status data; S2. The collected data is cleaned, spatiotemporally aligned, feature extracted, dimensionality reduced, and knowledge graph constructed to obtain a clean dataset, anomaly logs, spatiotemporally synchronized data matrix, standardized feature vector set, low-dimensional core feature vector set, physical entity relationship graph, and parameter association strength matrix. S3. Predict future water demand and remaining lifespan of key components, generate optimal control action sequence, and generate pre-control strategy based on prediction results, optimal control action sequence, and physical constraints in physical entity relationship graph. S4. Using the pre-control strategy as the set standard, retrieve the sensor data at the current moment in the clean dataset for fuzzy PID control adjustment, then perform constraint correction based on physical entity relationships, and generate control commands. S5. Use the isolated forest algorithm to detect anomalies in the real-time data stream. Once an anomaly is confirmed, trigger fault diagnosis and classify the fault and emergency response levels based on the fault diagnosis results. S6. Collect predictive control performance indicators, real-time control performance indicators, and fault handling control performance indicators within a preset period. S7. Evaluate whether the predictive control effect, real-time control effect, and fault handling control effect meet expectations in sequence. If they all meet expectations, calculate the comprehensive control effect index and store it automatically. Otherwise, output the control effect type that does not meet expectations.

[0034] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A smart water pump controller with a built-in AI chip, characterized in that: include: Pump data acquisition module: used to collect pump operating data, environmental data, and equipment status data; Multi-source data fusion module: The collected data is cleaned, spatiotemporally aligned, feature extracted, dimensionality reduced, and knowledge graph constructed to obtain a clean dataset, anomaly logs, spatiotemporally synchronized data matrix, standardized feature vector set, low-dimensional core feature vector set, physical entity relationship graph, and parameter association strength matrix. Pre-control strategy generation module: predicts future water demand and remaining lifespan of key components, generates optimal control action sequence, and generates pre-control strategy based on prediction results, optimal control action sequence, and physical constraints in physical entity relationship graph; Real-time control command generation module: It uses the pre-control strategy as the setting standard, retrieves the sensor data of the current moment in the clean dataset for fuzzy PID control adjustment, then performs constraint correction based on physical entity relationships, and generates control commands; Fault handling control module: It uses the isolated forest algorithm to detect anomalies in the real-time data stream, triggers fault diagnosis after confirming data anomalies, and performs fault classification and emergency handling classification response based on the fault diagnosis results; Control performance index collection module: used to collect predictive control performance indexes, real-time control performance indexes, and fault handling control performance indexes within a preset period; Control effect evaluation module: Evaluates whether the predictive control effect, real-time control effect, and fault handling control effect meet expectations in sequence. If they all meet expectations, the comprehensive control effect index is calculated and automatically stored; otherwise, the control effect type that does not meet expectations is output.

2. The intelligent water pump controller with a built-in AI chip according to claim 1, characterized in that: The pump data acquisition module collects pump operation data including pressure data, temperature data, flow rate data, vibration data, and electrical data; environmental data including ambient temperature and humidity data and meteorological data; and equipment status data including motor status data, valve status data, and energy consumption data.

3. The intelligent water pump controller with a built-in AI chip according to claim 1, characterized in that: The multi-source data fusion module includes a data receiving unit, a data cleaning unit, a spatiotemporal alignment unit, a feature extraction unit, a data dimensionality reduction unit, a knowledge graph construction unit, and an automatic data storage unit. The data receiving unit receives collected water pump data; the data cleaning unit removes outliers from the received data based on the 3σ principle to obtain a clean dataset and anomaly-marked logs; the spatiotemporal alignment unit performs time synchronization processing, spatial coordinate transformation, and data resampling on the clean dataset to obtain a spatiotemporal synchronized data matrix; the feature extraction unit performs time-domain feature calculation, windowing preprocessing, FFT transformation, frequency-domain feature extraction, and standardization on the spatiotemporal synchronized data matrix to obtain a standardized feature vector set; the data dimensionality reduction unit performs PCA dimensionality reduction and t-SNE dimensionality reduction on the standardized feature vector set to obtain a low-dimensional core feature vector set. The knowledge graph construction unit defines physical entity relationships and calculates association strengths from the spatiotemporal synchronized data matrix to obtain a physical entity relationship graph and a parameter association strength matrix. The automatic data storage unit automatically stores the clean dataset, anomaly labeling log, spatiotemporal synchronized data matrix, standardized feature vector set, low-dimensional core feature vector set, physical entity relationship graph, and parameter association strength matrix into the database and marks the update timestamp.

4. The intelligent water pump controller with built-in AI chip according to claim 1, characterized in that: The pre-control strategy generation module includes a data retrieval unit, a load forecasting unit, an equipment lifespan forecasting unit, an energy consumption optimization unit, a pre-control strategy generation unit, and an automatic data storage unit. The data retrieval unit retrieves clean datasets, spatiotemporal synchronized data matrices, standardized feature vector sets, low-dimensional core feature vector sets, physical entity relationship graphs, and parameter correlation strength matrices. The load forecasting unit inputs the flow time-series data from the spatiotemporal synchronized data matrix, environmental features from the standardized feature vector set, and calendar features from the physical entity relationship graph into the constructed LSTM network model for load forecasting, outputting the required water flow value for each time stamp in the future stage. The equipment lifespan forecasting unit inputs the low-dimensional core data into the model. The feature vector set, vibration features in the standardized feature vector set, and equipment operation history in the physical entity relationship graph are input into the constructed XGBoost regression model for equipment life prediction, and the remaining life of key equipment is output. The energy consumption optimization unit generates the optimal control action sequence based on the energy consumption optimization objective. The pre-control strategy generation unit generates pressure, flow reference values, and frequency limit values ​​for each time stamp in the future stage based on the load prediction results, the optimal control action sequence, equipment life prediction results, and physical constraints in the physical entity relationship graph. The automatic data storage unit automatically stores the generated pressure, flow reference values, and frequency limit values ​​for each time stamp in the future stage into the database and marks the update timestamp.

5. The intelligent water pump controller with a built-in AI chip according to claim 1, characterized in that: The real-time control command generation module includes a data retrieval unit, a feedforward setting unit, a PID control trigger determination unit, a PID control adjustment calculation unit, an adjustment correction unit, and a control command generation unit. The data retrieval unit is used to retrieve real-time pressure and real-time flow from the pre-control strategy and the cleaning data set. The feedforward setting unit marks the pressure and flow reference values ​​for each time point in the future stage in the pre-control strategy as the pressure and flow set values ​​for each time point in the future stage; the PID control triggering judgment unit triggers pressure PID control when the real-time pressure is inconsistent with the pressure set value, and triggers flow PID control when the real-time flow is inconsistent with the flow set value. The PID control adjustment calculation unit calculates the inverter frequency adjustment using a fuzzy PID algorithm when pressure PID control is triggered, and calculates the valve opening adjustment using a fuzzy PID algorithm when flow PID control is triggered. The adjustment correction unit constrains and corrects the calculated inverter frequency adjustment or valve opening adjustment based on the frequency limit value generated by the physical entity relationship and the pre-control strategy. The control command generation unit generates actuator control commands based on the corrected inverter frequency adjustment or valve opening adjustment.

6. The intelligent water pump controller with built-in AI chip according to claim 1, characterized in that: The fault handling control module includes a real-time data stream feature extraction unit, an anomaly detection unit, an anomaly determination unit, a fault diagnosis unit, a fault classification unit, and an emergency handling unit. The real-time data stream feature extraction unit is used to extract features from the real-time data stream. The anomaly detection unit uses the isolated forest algorithm to detect anomalies in the extracted real-time data stream features and outputs an anomaly score. The anomaly detection unit compares the anomaly score with a dynamic threshold. If the anomaly score is greater than the dynamic threshold, the feature is determined to be abnormal, triggering fault diagnosis. The fault diagnosis unit includes fault type identification based on convolutional attention networks and fault location localization based on Bayesian networks, generating a fault diagnosis report based on the diagnosis results. The fault classification unit divides the fault level into three levels: Level 1, Level 2, and Level 3. The emergency handling unit switches to energy-saving mode with optimized control when a Level 1 fault is triggered, switches to conservative mode with PID control when a Level 2 fault is triggered, and switches to safety mode with constant pressure control when a Level 3 fault is triggered.

7. The intelligent water pump controller with built-in AI chip according to claim 1, characterized in that: The predictive control performance indicators collected by the control performance indicator collection module within a preset period include average flow prediction error, average pressure prediction error, average RUL prediction error rate, energy consumption optimization coefficient per unit flow, and reference curve tracking compliance rate; the real-time control performance indicators collected within a preset period include pressure control deviation, flow control deviation, energy efficiency ratio compliance rate, and pressure fluctuation coefficient; and the fault handling control performance indicators collected within a preset period include anomaly detection rate, average mode switching efficiency, and fault downtime percentage.

8. The intelligent water pump controller with built-in AI chip according to claim 1, characterized in that: The control effect evaluation module includes a data receiving unit, a predictive control effect coefficient calculation unit, a real-time control effect coefficient calculation unit, a fault handling control effect coefficient calculation unit, a control effect judgment unit, a comprehensive control effect index calculation unit, and a data output unit. The data receiving unit receives predictive control effect indicators, real-time control effect indicators, and fault handling control effect indicators within a preset period. The predictive control effect coefficient calculation unit calculates the predictive control effect coefficient; the real-time control effect coefficient calculation unit calculates the real-time control effect coefficient; and the fault handling control effect coefficient calculation unit calculates the fault handling control effect coefficient. The predictive control effect coefficient, real-time control effect coefficient, and fault handling control effect coefficient calculated by the control effect judgment unit are compared with their corresponding preset values. If all calculated values ​​are greater than or equal to the preset values, the control effect is deemed to meet expectations; otherwise, it is deemed not to meet expectations. The comprehensive control effect index calculation unit calculates the comprehensive control effect index and automatically stores it in the database. The data output unit transmits control effect types that do not meet expectations to the water pump control center.

9. A method for controlling an intelligent water pump with a built-in AI chip, using an intelligent water pump controller with a built-in AI chip as described in any one of claims 1-8, characterized in that: Includes the following steps: S1. Collect pump operation data, environmental data, and equipment status data; S2. The collected data is cleaned, spatiotemporally aligned, feature extracted, dimensionality reduced, and knowledge graph constructed to obtain a clean dataset, anomaly logs, spatiotemporally synchronized data matrix, standardized feature vector set, low-dimensional core feature vector set, physical entity relationship graph, and parameter association strength matrix. S3. Predict future water demand and remaining lifespan of key components, generate optimal control action sequence, and generate pre-control strategy based on prediction results, optimal control action sequence, and physical constraints in physical entity relationship graph. S4. Using the pre-control strategy as the set standard, retrieve the sensor data at the current moment in the clean dataset for fuzzy PID control adjustment, then perform constraint correction based on physical entity relationships, and generate control commands. S5. Use the isolated forest algorithm to detect anomalies in the real-time data stream. Once an anomaly is confirmed, trigger fault diagnosis and classify the fault and emergency response levels based on the fault diagnosis results. S6. Collect predictive control performance indicators, real-time control performance indicators, and fault handling control performance indicators within a preset period. S7. Evaluate whether the predictive control effect, real-time control effect, and fault handling control effect meet expectations in sequence. If they all meet expectations, calculate the comprehensive control effect index and store it automatically. Otherwise, output the control effect type that does not meet expectations.

Images