Automatic hoist control system and method for intelligent water utilization

Abstract

The invention discloses a hoist automatic control method for intelligent water conservancy, and belongs to the technical field of intelligent water conservancy. The method specifically comprises the following steps: S1, constructing a multi-dimensional parameter monitoring system: deploying multiple types of sensors on a hoist and corresponding water conservancy facilities, collecting water conservancy environment parameters, equipment operation parameters and environment meteorological parameters in real time, performing cleaning, format conversion and space-time alignment on data, and synchronizing the processed data to an intelligent water conservancy cloud platform; s2, historical data preprocessing and depth feature extraction: obtaining historical data of the multi-dimensional data, performing multi-dimensional preprocessing, constructing depth feature extraction ore pulp based on the preprocessed historical data and the data acquired in real time, and extracting depth features capable of accurately reflecting a water conservancy scene rule and an equipment operation state. By constructing a multi-dimensional parameter monitoring system, three types of parameters of a water conservancy environment, equipment operation and environmental meteorology are collected in real time, and decision one-sidedness caused by only depending on water level data is avoided.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of smart water conservancy, in particular to a smart water conservancy use hoist automatic control system and method. BACKGROUND

[0002] In the smart water conservancy system, the hoist is the core equipment for controlling the operation of the gate, and its control precision and response efficiency directly affect the effect of water conservancy operations such as water resources scheduling, flood control and drainage, and irrigation water supply. Most of the existing hoist control methods only rely on water level data to operate the hoist, without considering multi-dimensional parameters such as water flow speed, water quality, gate operation state (such as wear degree, risk of jamming), etc., resulting in one-sided control decisions and prone to misoperation or lag problems. For example, during the flood season, when the water flow speed increases suddenly, only relying on water level data to control the gate opening may damage the gate or hoist components due to excessive water flow impact. SUMMARY

[0003] The purpose of the present application is to provide a smart water conservancy use hoist automatic control system and method that can solve the problem of single parameter monitoring in existing hoist control methods.

[0004] Technical solution: To solve the above technical problems, according to one aspect of the present application, more specifically, a smart water conservancy use hoist automatic control method, specifically including the following steps:

[0005] S1, Construct a multi-dimensional parameter monitoring system: deploy multiple types of sensors on the hoist and its corresponding water conservancy facilities, real-time collect water conservancy environmental parameters, equipment operation parameters, and environmental meteorological parameters, and perform data cleaning, format conversion, and spatio-temporal alignment, and synchronize the processed data to the smart water conservancy cloud platform;

[0006] S2, Historical data preprocessing and deep feature extraction: obtain historical data of multi-dimensional data, perform multi-dimensional preprocessing, and based on the preprocessed historical data and real-time collected data, construct a deep feature extraction mine pulp, and extract deep features that can accurately reflect water conservancy scene rules and equipment operation state;

[0007] S3, Intelligent decision-making model construction based on deep features: construct an intelligent decision-making model that integrates multiple algorithms with deep features as input, to realize accurate decision-making of scene recognition, parameter calculation, and safety warning;

[0008] S4, Dynamic adaptive control strategy execution: according to the scene type, target control parameter, and safety warning result output by the intelligent decision-making model, use the hoist control system of the PLC controller to execute the dynamic adaptive control strategy;

[0009] S5, energy saving optimization and data interaction: real-time monitoring of energy consumption, intelligent and graceful start-stop machine start-stop time, when the start-stop machine is in idle state, automatically enter low-power sleep mode, real-time data upload, management personnel issue remote control instructions through the intelligent water conservancy cloud platform, and establish data sharing protocol with other water conservancy systems in the basin;

[0010] S6, control effect feedback and model iteration: evaluate the scheduling accuracy, equipment operation and scene adaptation, periodically incrementally train the intelligent model, optimize the algorithm and upgrade, and verify the iteration effect.

[0011] Further, in step S1, the water environment parameters include water level, water flow speed, pH value and turbidity; the equipment operation parameters include operating voltage, start-stop machine output torque, gate lifting displacement, temperature of start-stop machine and key components; the environmental meteorological parameters include rainfall and wind speed; the data cleaning uses 3σ criterion to remove outliers, and the missing data is filled by linear interpolation method; format conversion: unify the heterogeneous data output by different sensors into JSON format; space-time alignment: take the system time of the edge computing node as the benchmark, synchronize the time stamp of parameters with different collection frequencies, and associate the corresponding spatial position information to form a space-time unified data set, and synchronize the processed data to the intelligent water conservancy cloud platform.

[0012] Further, step S2 specifically comprises the following steps:

[0013] S21, retrieve historical data from the intelligent water conservancy cloud platform, remove redundant data collected repeatedly based on data time stamp and parameter identification, smooth the parameter data with large fluctuations by using sliding average algorithm, standardize the parameter data with different dimensions by using Z-Score standardization formula, and divide the processed data into training set, validation set and test set according to the ratio of 7:2:1;

[0014] S22, use one-dimensional convolutional neural network to extract local features of time series data, capture the change trend of parameters in time dimension, output time series primary feature vector with dimension 256, use multilayer perceptron to perform nonlinear mapping on numerical data, mine numerical correlation features of parameters, output numerical primary feature vector with dimension 64, use one-hot encoding to convert category information into binary feature vector, and output category primary feature vector with consistent dimension and category number;

[0015] S23, calculate the correlation weight between different types of primary feature vectors by using scaling dot product attention algorithm, weight sum the primary feature vectors according to the attention weight, and then fuse the time series, numerical and category features by splicing operation to form intermediate fusion feature vector with dimension 512, and normalize the intermediate feature vector by using layer normalization technology, with mean value 0 and variance 1.

[0016] S24, using a stacked encoder to compress and mine the normalized intermediate fusion feature vector, learning the internal structure of the feature on the unlabeled data through the encoding-decoding process of the autoencoder, fine-tuning the hidden layer parameters of the autoencoder combined with the label information in the historical data to further optimize the feature expression ability, extracting the output of the topmost hidden layer of the autoencoder to obtain a deep feature vector with a dimension of 128.

[0017] Further, the step S3 specifically comprises the following steps:

[0018] S31, using an improved convolutional neural network to construct a scene classification model, the input layer receives the deep feature vector, the convolutional layer extracts the local correlation information of the feature, the pooling layer reduces the feature dimension, the fully connected layer performs feature mapping, and the output layer outputs the probability value of each type of scene, the pupa cross-entropy function, and the Adam optimizer is used to train the model on the training set, and the number of convolutional kernels and the number of neurons in the fully connected layer are adjusted through the validation machine to make the scene recognition accuracy ≥95%, when the real-time feature is input into the deep feature input model, the scene type with the maximum output probability is output as the current water conservancy scene, and the scene confidence is also output, and when the confidence ≥0.8, it is determined that the recognition is effective, otherwise the artificial review process is started;

[0019] S32, for different scene types, a hybrid model combining long short-term memory network and reinforcement learning is used to calculate the optimal target operating parameters of the opening and closing machine, a bidirectional LSTM network is constructed, the time sequence correlation features in the deep feature vector are input, and the initial control parameters of the opening and closing machine are predicted, a Q-Learning reinforcement learning model is constructed with the highest water resource scheduling accuracy and the lowest energy consumption as the reward function, and the output parameters output by the LSTM are iteratively optimized: when the scheduling accuracy score ≥90 points and the energy saving score ≥85 points, the parameters remain unchanged, when the scheduling accuracy score <85 points, the gate opening amplitude is adjusted preferentially, the adjustment step is ±5%, and the scheduling accuracy is improved, when the energy saving score <80 points, the operation power adjustment step is optimized by ±100W, and the energy consumption is reduced, and the output optimized target control parameters include: gate lifting speed, opening / closing amplitude, and operation power.

[0020] S33, based on the device state features in the deep features, a fusion algorithm of support vector machine and gradient boosting is used to construct a fault prediction and risk assessment model, a SVM binary classification model is constructed with the device operating features in the deep feature vector as input to judge whether the device has a fault risk, and an XGBoost multi-classification model is input to the samples judged as abnormal by the SVM, combined with the fault verification degree label level in the historical fault data: low risk: fault probability <30%, medium risk: 30%≤fault probability≤70%, high risk: fault probability >70%.

[0021] Further, in step S4, according to the scene type, target control parameter and safety warning result output by the intelligent decision model, a dynamic adaptive control strategy is executed by the hoist control system of the PLC controller, specifically including:

[0022] Normal scene control: The hoist motor is started according to the gate lifting speed and opening amplitude, and frequency conversion technology is used to realize smooth operation. Real-time monitoring of water conservancy environmental parameters, when the water level fluctuation is within the threshold range and the water flow velocity change is less than the threshold value, automatically fine-tune the running power and gate lifting speed, when the output scene confidence is less than the threshold value, pause the normal control, start the scene review process, confirm the scene type and then execute the corresponding control;

[0023] Emergency scene control: When the scene is identified as an emergency drainage scene, and the water flow velocity exceeds the threshold or the rainfall reaches the threshold, the emergency control mode is immediately started, the gate lifting speed is increased, and the gate opening amplitude is increased. Real-time detection of torque, when the torque exceeds the threshold, automatically reduce the running power, and start the gate buffer device, establish a linkage mechanism with other hoist systems of the reservoir and river, share water level and flow data through the smart water conservancy cloud platform, and realize multi-gate collaborative scheduling;

[0024] Fault response control: Low risk level: The hoist is normally operated, the monitoring frequency is improved, and the parameter changes are tracked in real time; Medium risk level: automatically reduce the hoist operating load, send warning information to the management personnel through the smart water conservancy cloud platform, and start the backup sensor data verification; High risk level: immediately trigger the emergency stop protection, if a backup hoist device is configured, automatically switch to the backup device for operation, and synchronize the fault position, fault type and emergency treatment measures to the smart water conservancy cloud platform and the management personnel mobile terminal; If no backup hoist device is configured, start the manual emergency process, push the fault site disposal guide through the smart water conservancy cloud platform, and at the same time, link the surrounding water conservancy facilities to adjust the operating state to avoid imbalance of the water conservancy system operation.

[0025] Further, in step S5, based on the collected current and voltage data, the real-time running energy consumption is calculated in combination with the hoist running time, the consumption data is updated regularly and uploaded to the smart water conservancy cloud platform, and the particle swarm optimization algorithm is used to optimize the hoist start-stop time with the lowest energy consumption and scheduling as the objective function. When multiple hoists are operated cooperatively, the load of each device is automatically distributed, and when the hoist is in an idle state, a low-power sleep mode is automatically performed.

[0026] Furthermore, in step S6, the deviation between the daily statistical analysis of the timing water conservancy operation results and the target value is calculated, the causes of the deviation are analyzed, and a deviation analysis report is generated. Weekly statistics are compiled on the gate hoist failure rate, mean time between failures (MTBF), and energy consumption compliance rate. If the failure rate exceeds a threshold or the energy consumption compliance rate falls below a threshold, equipment maintenance reminders and model optimization processes are triggered. Monthly assessments are conducted on the adaptability of control strategies for different scenarios. The strategy adaptability level is determined based on user feedback and data indicators. Scenarios with the lowest adaptability level are re-optimized in terms of feature extraction weights and model parameters. Each quarter, newly added historical data is processed and included in the training set for intelligent... The decision model undergoes incremental training, employing a method of freezing the bottom layer and fine-tuning the top layer: freezing the parameters of the feature extraction layer and fine-tuning only the parameters of the top fully connected layer of the scene recognition, control parameter calculation, and safety early warning modules. When a new scene or new fault type appears, the deep feature extraction framework and intelligent decision model are updated. After model iteration, the model is validated on a test set and in actual water conservancy scenarios. If the scheduling accuracy deviation is not higher than the threshold, the fault early warning timeliness rate is not lower than the threshold, and the energy consumption reduction is not lower than the threshold during the validation period, the iteration takes effect; if the standards are not met, the model returns to readjust the feature extraction weights or algorithm parameters and is re-validated until the requirements are met.

[0027] According to another aspect of the present invention, an automatic control system for gate opening and closing in smart water utilization is provided. The system is used to implement the above-described automatic control method for gate opening and closing in smart water utilization, including: a multi-dimensional parameter monitoring module, a historical data preprocessing and deep feature extraction module, an intelligent decision-making module, a dynamic adaptive control strategy module, an energy-saving optimization and data interaction module, and a control effect feedback and model iteration module.

[0028] Multi-dimensional parameter monitoring module: It is used to deploy various types of sensors on the gate hoist and its corresponding water conservancy facilities to collect water conservancy environmental parameters, equipment operating parameters, and environmental meteorological parameters in real time, and to clean, convert, and align the data in time and space, and then synchronize the processed data to the smart water conservancy cloud platform.

[0029] Historical data preprocessing and deep feature extraction module: used to acquire historical data of multidimensional data, perform multidimensional preprocessing, and construct a deep feature extraction slurry based on the preprocessed historical data and real-time acquired data to extract deep features that can accurately reflect the laws of water conservancy scenarios and the operating status of equipment;

[0030] Intelligent Decision Module: Used to build an intelligent decision model that integrates multiple algorithms with deep features as input, so as to achieve accurate decision-making in scene recognition, parameter calculation and safety warning;

[0031] Dynamic Adaptive Control Strategy Module: This module is used to execute dynamic adaptive control strategies in the gate control system of the PLC controller based on the scenario type, target control parameters, and safety warning results output by the intelligent decision model.

[0032] Energy-saving optimization and data interaction module: for real-time monitoring of energy consumption, intelligent and graceful start-stop of the machine, when the machine is idle, automatically enter low-power sleep mode, real-time data upload, management personnel through the wisdom of water conservancy cloud platform to issue remote control instructions, and establish data sharing agreement with other water conservancy systems in the basin;

[0033] Control effect feedback and model iteration module: for evaluating the scheduling accuracy, device operation and scene adaptation, periodically incrementally training the intelligent model, optimizing the algorithm and upgrading, and verifying the iteration effect.

[0034] Beneficial effects:

[0035] 1. By constructing a multi-dimensional parameter monitoring system, real-time collection of water conservancy environment, device operation, environmental meteorological three types of parameters, avoiding the decision-making one-sidedness caused by relying only on water level data; at the same time, through data cleaning, time and space alignment and deep feature extraction, combined with the intelligent decision-making model of multi-algorithm fusion, the control decision is more in line with the water conservancy scene law and device operation state, reducing the misoperation and lag problem, for example, in flood season, the gate can be adjusted by comprehensively considering the water flow velocity, torque and other parameters to avoid the damage of water flow impact force to the equipment.

[0036] 2. Dynamic adaptive control strategy executes differentiated schemes for different scenes of conventional, emergency and fault: in the conventional scene, through variable frequency speed regulation and real-time fine adjustment, the deviation of water supply / irrigation amount is ensured to be ≤3%; in the emergency scene, the gate speed can be quickly increased, the opening degree can be increased, and multiple gates can be linked to shorten the drainage time by ≥20%; in the fault scene, the operation load is dynamically adjusted or the emergency stop is triggered according to the risk level, which adapts to the dynamic changes such as seasonal alternation and sudden rainstorm, and meets the precise control demand in different scenes.

[0037] 3. In terms of energy saving, the particle swarm optimization algorithm is used to optimize the start-stop time of the machine and the load distribution of multiple devices, combined with the low-power sleep mode in idle state, the actual energy consumption is reduced by 12%-18%; in terms of safety, the SVM and XGBoost fusion algorithm is used to realize fault prediction in advance, the monitoring frequency is increased for low risk, the load is reduced for early warning of medium risk, and the standby equipment is switched for high risk, reducing more than 70% of sudden failures, balancing operation efficiency, energy saving and safety, and reducing equipment maintenance and energy consumption cost.

[0038] 4. Management personnel can issue instructions through the cloud platform; at the same time, data sharing agreement is established with reservoir dispatching, irrigation management and other systems in the basin to realize cross-system collaborative scheduling without manual intervention, solve the problems of data interaction lag and weak remote control in traditional methods, and help the efficient collaborative operation of the wisdom water conservancy system

[0039] 5. Control effect feedback and model iteration mechanism, through daily scheduling accuracy evaluation, weekly equipment operation statistics, monthly scene adaptation evaluation, new data is regularly included in the training set for incremental training, and can adapt to new scenes and new fault types, ensure that the model maintains high accuracy and adaptability for a long time, and ensures the stable operation of the hoist and water conservancy system. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 is a method flow diagram. DETAILED DESCRIPTION

[0041] In order to make the technical scheme of the present application clearer, the present application will be further described in detail below in combination with the drawings and specific embodiments.

[0042] Example 1

[0043] Implementation of intelligent water utilization automatic control of hoist in irrigation scene

[0044] First, build a multi-dimensional parameter monitoring system:

[0045] Sensor deployment: deploy sensors on the hoist, irrigation channel and surrounding area:

[0046] Water conservancy environmental parameters: water level sensor (channel center, sampling frequency 1 time / minute, accuracy ±0.01m), water flow velocity sensor (channel section, 1 time / 30 seconds, ±0.05m / s), pH value sensor (channel inlet, 1 time / hour, ±0.1pH), turbidity sensor (same as pH value sensor position, 1 time / hour, ±1NTU);

[0047] Device operating parameters: voltage sensor (hoist motor input end, 1 time / 10 seconds, ±0.5V), torque sensor (hoist transmission shaft, 1 time / 15 seconds, ±1N·m), displacement sensor (gate side, 1 time / 5 seconds, ±0.005m), temperature sensor (motor housing and gear box, 1 time / 20 seconds, ±0.5℃);

[0048] Environmental meteorological parameters: rainfall sensor (weather station beside the channel, 1 time / 10 minutes, ±0.1mm), wind speed sensor (same as weather station position, 1 time / 15 minutes, ±0.1m / s).

[0049] Data processing and transmission:

[0050] Data cleaning: use 3σ criterion to remove outliers of water level, torque and other parameters (such as a water level data deviating from the mean value by 4 times the standard deviation, which is judged as an outlier and removed), and fill in 1 time missing data of water flow velocity sensor (continuous missing 2 points) by linear interpolation method;

[0051] Format conversion: unify the analog data of the voltage sensor and the digital data of the displacement sensor into JSON format;

[0052] Space-time alignment: take the system time of the edge computing node (industrial-grade embedded processor, 1.2 GHz main frequency, 2 GB memory) as the reference, synchronize all parameter timestamps (time error ≤80 ms), and associate sensor installation coordinates (such as water level sensor coordinates X=120.5°, Y=30.2°) to form a unified space-time data set, which is synchronized to the smart water conservancy cloud platform (distributed storage, capacity 15 TB, read / write speed 120 MB / s).

[0053] Second step, historical data preprocessing and deep feature extraction:

[0054] 1. Historical data preprocessing: retrieve the historical data of the irrigation area for the past 5 irrigation seasons (120 GB) from the cloud platform, remove duplicate data based on timestamps and parameter identifiers (duplication rate 0.3%); use the sliding average algorithm (window 10 data points) to smooth the water flow velocity and torque data; use the Z-Score standardization formula Map parameters such as water level (unit m) and voltage (unit V) to the [-1, 1] interval; divide into training set (84 GB), validation set (24 GB), and test set (12 GB) in the ratio of 7:2:1.

[0055] 2. Primary feature extraction: use 1D-CNN (3 layers of convolution, 3x3 kernel, ReLU activation) to process water level, voltage, and other time series data, output 256-dimensional time series primary feature vectors; use MLP (2 layers of hidden layers, 128 / 64 neurons, Sigmoid activation) to process pH value and turbidity numerical data, output 64-dimensional numerical primary feature vectors; use one-hot encoding to convert "irrigation scenario" and "normal operation" category data into binary vectors (e.g. "irrigation scenario" corresponds to a binary vector with dimension 1).

[0056] 3. Intermediate feature fusion: use the scaled dot product attention algorithm to calculate weights (under irrigation scenario, water level time series feature weight 0.7, turbidity numerical feature weight 0.2), concatenate the three types of primary features after weighted summation, form a 512-dimensional intermediate fusion feature vector; process through layer normalization (mean 0, variance 1).

[0057] 4. Deep feature output: use a stacked autoencoder (3 layers of hidden layers, 512 / 256 / 128 neurons, LeakyReLU activation), unsupervised pre-training (reconstruction error 0.04), and fine-tune parameters combined with "irrigation scenario labels" to extract the topmost output and obtain a 128-dimensional deep feature vector.

[0058] Third step, intelligent decision-making model construction based on deep features:

[0059] 1. Scene recognition: An improved CNN is used to build a scene classification model. The input is a 128-dimensional depth feature vector. Two layers of convolution (5x1 kernel) are used to extract local correlations. One layer of max pooling (2x1 kernel) is used for dimension reduction. Two layers of full connection (64 / 16 neurons) are used for mapping. Softmax outputs the scene probability. The cross-entropy is used as the loss function. The Adam optimizer (learning rate 0.001) is used for training on the training set. The validation set is used to adjust the parameters (8 convolution kernels, 64 / 16 neurons in the full connection layer). The scene recognition accuracy is 96%. After real-time input of the depth features, the output is "irrigation scene" (probability 0.92) and confidence 0.92 (≥0.8, valid).

[0060] 2. Control parameter calculation: A bidirectional LSTM (2 layers, 128 neurons, dropout 0.2) is constructed. The water level change trend and historical power time series features in the depth features are input. The initial parameters are predicted: gate lifting speed 0.2 m / min, opening amplitude 50%, running power 1.2 kW (prediction error 2.5%). The Q-Learning model is used for optimization. The reward function calculates the scheduling accuracy score as 92 points and the energy saving score as 86 points (both meet the standard). The parameters remain unchanged. The output target control parameters are: speed 0.2 m / min, opening 50%, power 1.2 kW.

[0061] 3. Safety warning: The torque and temperature features in the depth features are input to construct an SVM binary classification model (RBF kernel, C=10, gamma=0.1) to determine the "normal" state of the equipment. The XGBoost model does not need to be started, and there is no warning information.

[0062] Fourth step, dynamic adaptive control strategy execution:

[0063] Execute regular scene control: Start the gate motor according to the target parameters. Use variable frequency speed regulation technology (frequency 30 Hz) to achieve smooth operation. The edge computing node monitors in real time. The water level fluctuation is ±3% (≤ threshold ±5%). The water flow velocity change is 0.1 m / s (≤ threshold 0.2 m / s). The running power is automatically adjusted to 1.18 kW (±50 W). The gate speed remains 0.2 m / min. The scene recognition confidence is always ≥0.8, no need to review, continuously guarantee irrigation water supply, the actual water supply amount and demand deviation is 2.2% (≤3%).

[0064] Fifth step, energy saving optimization and data interaction:

[0065] Energy saving optimization: Based on current (real-time 1.8 A) and voltage (220 V) data, the formula Real-time energy consumption is calculated and updated every 1 minute and uploaded to the cloud platform; the particle swarm optimization algorithm optimizes the start-stop timing, and the power is reduced to 0.8 kW in advance before the end of irrigation (water level close to the target value, deviation 1.5%), achieving "soft landing" shutdown; the hoist is not idle and does not need to hibernate.

[0066] Data interaction: Through the 5G module (uplink speed 150 Mbps, time delay 15 ms), the running parameters and water supply data are uploaded every 10 seconds; the management personnel issue "maintain current parameters" instructions through the cloud platform, which are encrypted by AES-256 and checked by CRC32, and the execution results are returned within 1 second; the water level and water supply data are shared with the irrigation area management system without manual intervention.

[0067] Step 6, control effect feedback and model iteration:

[0068] Effect evaluation: Daily statistics of water supply deviation 2.2% (≤3%), generate unbiased analysis report; weekly statistics of failure rate 0 times / 100 hours, energy consumption compliance rate 90% (≥85%), no maintenance reminder; monthly evaluation of scene adaptability, user feedback on stable water supply, data indicators show that scene recognition accuracy is 96%, and the adaptation level is "excellent".

[0069] Model iteration: No new data in this quarter, no incremental training for the time being; no new scene or new failure, no need to upgrade the model; verification period 15 days, scheduling accuracy deviation 2.2%, no failure, energy consumption reduced by 13% (≥10%), no need to start iteration.

[0070] Example 2

[0071] Implementation of intelligent water utilization automatic control of hoist in flood season emergency drainage scene

[0072] Step 1, build a multi-dimensional parameter monitoring system:

[0073] Sensor deployment:

[0074] Water environment parameters: water level sensor (river section, 1 time / minute, ±0.01 m, monitor river water level), water flow velocity sensor (downstream of the gate, 1 time / 30 seconds, ±0.05 m / s), pH value sensor (river bank, 1 time / hour, ±0.1 pH), turbidity sensor (same as pH value location, 1 time / hour, ±1 NTU);

[0075] Device operating parameters: voltage sensor (hoist motor, 1 time / 10 seconds, ±0.5 V), torque sensor (drive shaft, 1 time / 15 seconds, ±1 N·m), displacement sensor (gate, 1 time / 5 seconds, ±0.005 m), temperature sensor (motor, bearing, 1 time / 20 seconds, ±0.5℃);

[0076] Environmental meteorological parameters: Rainfall sensor (river side weather station, 1 time / 10 minutes, ±0.1 mm), wind speed sensor (same as weather station, 1 time / 15 minutes, ±0.1 m / s).

[0077] Data processing and transmission: sensors transmit data to edge computing nodes (1.5 GHz main frequency, 4 GB memory) through LoRaWAN modules (rate 30 kbps, distance 8 km); 3σ criterion is used to remove some wind speed outliers (deviation from mean value 3.5 times standard deviation), and linear interpolation is used to fill in 1 time rainfall missing data; after conversion to JSON format, time and space are aligned (time error ≤90 ms), and synchronized to the cloud platform (storage 20 TB, read and write 150 MB / s).

[0078] Second step, historical data preprocessing and deep feature extraction:

[0079] Historical data preprocessing: retrieve historical data of the past 5 years (150 GB), remove duplicates (0.4% repetition rate), smooth water flow velocity and torque data using moving average method (window 10 points), and divide into training set (105 GB), validation set (30 GB), and test set (15 GB) after Z-Score standardization according to 7:2:1.

[0080] Feature extraction: 1D-CNN processes water level and water flow velocity time series data, outputting 256-dimensional time series features; MLP processes rainfall and wind speed numerical data, outputting 64-dimensional numerical features; one-hot encoding "emergency drainage scene" into a binary vector; attention weight calculation (water flow velocity weight 0.8, pH value weight 0.1), fusion into 512-dimensional intermediate features and standardization; after fine-tuning of stacked autoencoder (reconstruction error 0.03), output 128-dimensional deep feature vector.

[0081] Third step, intelligent decision-making model construction based on deep features:

[0082] 1. Scene recognition: improved CNN model input 128-dimensional features, scene recognition accuracy after training 97%; after real-time input of features, output "emergency drainage scene" (probability 0.95), confidence 0.95 (≥0.8, valid).

[0083] 2. Control parameter calculation: bidirectional LSTM predicts initial parameters: gate speed 0.3 m / min, opening 80%, power 1.8 kW (error 2.8%); Q-Learning optimization, reward function calculation gets scheduling accuracy score 88 (<90), energy saving score 82, priority adjustment gate opening to 85% (step +5%), scheduling accuracy score improves to 91, output target parameters: speed 0.3 m / min, opening 85%, power 1.8 kW.

[0084] 3. Safety warning: SVM model input torque (real-time 125 N·m, rated 100 N·m), temperature (45℃) features, determine "abnormal"; XGBoost model combined with historical failure data, output failure risk level "medium risk" (failure probability 45%, 30% ≤ 45% ≤ 70%), warning information: "torque overrated by 25%, medium risk, estimated failure time 1.5 hours".

[0085] Fourth step, dynamic adaptive control strategy execution:

[0086] Execute emergency + medium risk control in failure:

[0087] Emergency control: water flow speed real-time 3.2 m / s (> threshold 3 m / s), 1 hour rainfall 18 mm (≥ threshold 16 mm), start emergency mode, gate speed increases to 0.42 m / min (0.3 x 1.4 times), opening degree remains 85%; Real-time monitoring torque 128 N·m (overrated by 28%), automatically reduce power to 1.44 kW (down 20%), start hydraulic damper (buffer time 0.8 s); Linkage with 2 upstream hoists, share water level data through cloud platform, upstream gate opens 60% first, this gate runs synchronously, shortens drainage time.

[0088] Medium risk control in failure: automatically reduce operating load, power from 1.44 kW to 1.22 kW (down 15%), speed from 0.42 m / min to 0.38 m / min (down 10%); Send warning information to management personnel through cloud platform ("torque overrated by 28%, medium risk, suggest checking within 1 hour"), start backup torque sensor (accuracy ±1 N·m) cross verification, data consistent, confirm risk effective.

[0089] Fifth step, energy saving optimization and data interaction:

[0090] Energy saving optimization: real-time calculation of energy consumption (current 8.2 A, voltage 220 V, \cos\varphi = 0.88), upload every 1 minute; Particle swarm optimization algorithm allocates load between this gate and upstream gate, this gate power 1.22 kW, upstream gate average power 1.3 kW (load deviation 4%); Hoist continues to run, no idle, no need to sleep.

[0091] Data interaction: optical fiber transmission (uplink rate 200 Mbps, time delay 12 ms), upload drainage data every 10 seconds; Management personnel issue "strengthen torque monitoring" instructions, respond within 1 second after encryption verification; Share water level and drainage volume data with river flood control dispatching system, adjust gate parameters cooperatively.

[0092] Sixth step, control effect feedback and model iteration:

[0093] Effect evaluation: The highest water level of the river channel is 10.3m (≤ safety line 10.5m) per day, and the drainage time is shortened by 23% (≥20%) compared with the historical same period; the failure occurrence rate is 0 times per 100 hours (after medium risk warning, hidden danger is checked in time, and is excluded), and the energy consumption reaches the standard rate of 88% (≥85%); the scene adaptability is evaluated every month, the flood control department scores "excellent", and the adaptation level is "excellent".

[0094] Model iteration: 30GB of new flood season data is added in this quarter, after preprocessing, it is added to the training set, the feature extraction layer is frozen, the top layer parameters of the scene recognition and safety warning module are fine-tuned, after training, the scene recognition accuracy is 97%, the fault warning accuracy is 90%; there is no new scene, and the model does not need to be upgraded; the verification period is 15 days, the scheduling accuracy deviation is 2.5%, the fault warning timeliness is 92%, and the energy consumption is reduced by 12%, and the iteration takes effect.

[0095] The above-described embodiments only express several embodiments of the present application, and the description is more specific and detailed, but it should not be understood as limiting the scope of the patent of the present application. It should be noted that for ordinary skilled persons in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A method for automatically controlling a water utilization intelligent hoist, characterized in that, Specifically comprising the following steps: S1, constructing a multi-dimensional parameter monitoring system: deploying multiple types of sensors on the hoist and its corresponding water conservancy facilities, real-time collection of water conservancy environmental parameters, equipment operation parameters, environmental meteorological parameters, and data cleaning, format conversion, time and space alignment, and synchronization of the processed data to the intelligent water conservancy cloud platform; S2, historical data preprocessing and deep feature extraction: obtaining historical data of multi-dimensional data, preprocessing in multiple dimensions, constructing deep feature extraction based on the preprocessed historical data and real-time collected data, and extracting deep features that can accurately reflect the water conservancy scene rules and equipment operation state; S3, intelligent decision-making model construction based on deep features: constructing an intelligent decision-making model that integrates multiple algorithms with deep features as input, realizing accurate decision-making of scene recognition, parameter calculation, and safety warning; S4, dynamic adaptive control strategy execution: according to the scene type, target control parameter, and safety warning result output by the intelligent decision-making model, using the hoist control system of the PLC controller to execute the dynamic adaptive control strategy; S5, energy saving optimization and data interaction: real-time monitoring of energy consumption, intelligent optimization of hoist start-stop timing, automatic entry into low-power sleep mode when the hoist is idle, real-time data upload, and management personnel issuing remote control instructions through the intelligent water conservancy cloud platform, and establishing data sharing protocols with other water conservancy systems in the basin; S6, control effect feedback and model iteration: evaluating the dispatching accuracy, equipment operation, and scene adaptation, periodically incrementally training the intelligent model, algorithm optimization and upgrading, and verifying the iteration effect.

2. The automatic control method of the gate machine for intelligent water utilization according to claim 1, characterized in that: In step S1, the water conservancy environmental parameters include water level, flow velocity, pH value, and turbidity; the equipment operation parameters include operating voltage, hoist output torque, gate lifting displacement, and temperature of the hoist and key components; the environmental meteorological parameters include rainfall and wind speed; data cleaning uses the 3σ criterion to remove outliers, and linear interpolation method is used to fill in missing data; format conversion: heterogeneous data output by different sensors are uniformly converted to JSON format; time and space alignment: using the system time of the edge computing node as the reference, synchronizing the time stamps of parameters with different collection frequencies, and associating the corresponding spatial position information to form a unified time and space data set, and synchronizing the processed data to the intelligent water conservancy cloud platform.

3. The automatic control method of the gate hoist for intelligent water utilization according to claim 2, characterized in that: Step S2 specifically comprises the following steps: S21, retrieve historical data from the intelligent water conservancy cloud platform, remove redundant data based on data time stamp and parameter identification, smooth the parameter data with large fluctuations using the moving average algorithm, standardize the parameter data with different dimensions using the Z-Score standardization formula, and divide the processed data into training set, validation set, and test set in the ratio of 7:2:1; S22, the one-dimensional convolutional neural network is used for extracting local features of the time sequence data, a change trend of parameters in the time dimension is captured, a time sequence primary feature vector with a dimension of 256 is output, a multi-layer perceptron is used for nonlinear mapping of the numerical data, numerical correlation features of the parameters are mined, a numerical primary feature vector with a dimension of 64 is output, a one-hot encoding is used for converting category information into a binary feature vector, and a category primary feature vector with a dimension consistent with the number of categories is output; S23, a scaled dot-product attention algorithm is used for calculating correlation weights between different types of primary feature vectors, the primary feature vectors are weighted and summed according to the attention weights, and then the time sequence, numerical and category features are fused through a splicing operation to form a middle-level fusion feature vector with a dimension of 512; a layer normalization technique is used for normalizing the middle-level fusion feature vector, with a mean of 0 and a variance of 1; S24, a stacked encoder is used for dimension compression and deep mining of the normalized middle-level fusion feature vector, the internal structure of the features is learned on the unlabeled data through the encoding-decoding process of the autoencoder, the parameters of the hidden layer of the autoencoder are fine-tuned in combination with the label information in the historical data, the feature expression capability is further optimized, and the output of the topmost hidden layer of the autoencoder is extracted to obtain a deep feature vector with a dimension of 128.

4. The automatic control method of the gate machine for intelligent water utilization according to claim 3, characterized in that: The step S3 specifically includes the following steps: S31, an improved convolutional neural network is used to construct a scene classification model, the input layer receives the deep feature vector, the convolutional layer extracts local correlation information of the features, the pooling layer reduces the feature dimension, the fully connected layer performs feature mapping, the output layer outputs probability values of various scenes, the cross-entropy function of chrysalis is used, the Adam optimizer is used to train the model on the training set, the number of convolutional kernels and the number of neurons in the fully connected layer are adjusted through the validation machine, so that the scene recognition accuracy is greater than or equal to 95%, when the real-time feature is input into the deep feature input model, the scene type with the maximum output probability is output as the current water conservancy scene, and the scene confidence is output at the same time, if the confidence is greater than or equal to 0.8, it is determined that the recognition is effective, otherwise, an artificial review process is started; S32, for different scene types, a hybrid model of a long short-term memory network combined with reinforcement learning is used to calculate the optimal target operating parameters of the gate, a bidirectional LSTM network is constructed, the time sequence correlation features in the deep feature vector are used as input, the initial control parameters of the gate are predicted, a Q-Learning reinforcement learning model is constructed with the highest water resource scheduling accuracy and the lowest energy consumption as a reward function, and the output parameters output by the LSTM are iteratively optimized: when the scheduling accuracy score is greater than or equal to 90 points and the energy saving score is greater than or equal to 85 points, the parameters remain unchanged, when the scheduling accuracy score is less than 85 points, the gate opening amplitude is adjusted preferentially, the adjustment step is ±5%, and the scheduling accuracy is improved, when the energy saving score is less than 80 points, the operation power adjustment step is optimized by ±100W, and the energy consumption is reduced, and the optimized target control parameters, including the gate lifting speed, the opening / closing amplitude and the operation power, are output. S33, based on the device state feature in the deep feature, a fusion algorithm of support vector machine and gradient boosting is used to construct a fault prediction and risk assessment model, an SVM binary classification model is constructed with the device running feature in the deep feature vector as input to judge whether the device has a fault risk, and an XGBoost multi-classification model is input for the samples judged as abnormal by SVM, combined with the fault verification degree label in the historical fault data: low risk: fault probability < 30%, medium risk: 30% ≤ fault probability ≤ 70%, high risk: fault probability > 70%.

5. The automatic control method of the gate machine for intelligent water utilization according to claim 4, characterized in that: In the step S4, according to the scene type, target control parameter and safety warning result output by the intelligent decision model, a start-stop machine control system of a PLC controller is used to execute a dynamic adaptive control strategy, which specifically includes: Normal scene control: the start-stop machine motor is started according to the station gate lifting speed and opening amplitude, frequency conversion speed regulation technology is used to realize smooth operation, and the water conservancy environment parameters are monitored in real time; when the water level fluctuation is within the threshold range and the water flow velocity change is less than the threshold value, the operation power and gate lifting speed are automatically adjusted; when the output scene confidence is less than the threshold value, the normal control is suspended, the scene review process is started, and the corresponding control is executed after the scene type is confirmed; Emergency scene control: when the scene is identified as an emergency drainage scene, and the water flow velocity exceeds the threshold or the rainfall reaches the threshold, the emergency control mode is immediately started, the gate lifting speed is increased, and the gate opening amplitude is increased, the torque is detected in real time, and when the torque exceeds the threshold, the operation power is automatically reduced, and the gate buffer device is started, a linkage mechanism with other start-stop machine systems of reservoirs and river channels is established, the water level and flow data are shared through the intelligent water conservancy cloud platform, and multi-gate collaborative scheduling is realized; Fault response control: low risk level: the start-stop machine is normally operated, the monitoring frequency is improved, and the parameter change is tracked in real time; medium risk level: the start-stop machine operation load is automatically reduced, the warning information is sent to the management personnel through the intelligent water conservancy cloud platform, and the poor verification data of the standby sensor is started; high risk level: the emergency stop protection is immediately triggered, if a standby start-stop device is configured, the standby device is automatically switched to operate, and the fault position, fault type and emergency treatment measures are synchronized to the intelligent water conservancy cloud platform and the management personnel mobile terminal; if no standby start-stop device is configured, the manual emergency process is started, the fault site disposal guide is pushed through the intelligent water conservancy cloud platform, and the surrounding water conservancy facilities are linked to adjust the operation state to avoid imbalance of the water conservancy system operation.

6. The automatic control method of the gate hoist for intelligent water utilization according to claim 5, characterized in that: In the step S5, based on the collected current and voltage data, the real-time operation energy consumption is calculated combined with the start-stop machine operation time, the consumption data is updated regularly and uploaded to the intelligent water conservancy cloud platform, the particle swarm optimization algorithm is used to optimize the start-stop time of the start-stop machine with the lowest energy consumption and scheduling as the objective function, when multiple start-stop machines are operated collaboratively, the loads of the devices are automatically distributed, and when the start-stop machine is in an idle state, the low-power sleep mode is automatically performed.

7. The automatic control method of the gate hoist for intelligent water utilization according to claim 6, characterized in that: In the step S6, the deviation of the water conservancy operation result from the target value is counted daily, the deviation reason is analyzed to generate a deviation analysis report, the failure rate, the average failure-free operation time, and the energy consumption compliance rate are counted weekly, the equipment maintenance reminder and the model optimization process are triggered if the failure rate is greater than a threshold or the energy consumption compliance is less than a threshold, the control strategy adaptation of different scenes is evaluated monthly, the strategy adaptation level is determined through user feedback combined with data indicators, the scenes with the lowest adaptation level are re-optimized for feature extraction weights and model parameters, the incremental training of the intelligent decision model is performed by processing the newly added historical data and adding the data to the training set every quarter, the bottom layer is frozen and the top layer is fine-tuned: the feature extraction layer parameters are frozen, and only the top layer fully connected layer parameters of the scene recognition, control parameter calculation, and safety warning module are fine-tuned, the deep feature extraction framework and the intelligent decision model are updated when a new scene or a new fault type appears, and the model is verified in the test set and the actual water conservancy scene after iteration, and if the scheduling accuracy deviation is not higher than a threshold, the fault warning rate is not lower than a threshold, and the energy consumption reduction is not lower than a threshold during the verification period, the iteration takes effect; if the requirements are not met, the feature extraction weights or algorithm parameters are re-adjusted, and the verification is re-performed until the requirements are met.

8. A hoist automatic control system for smart water utilization, characterized in that: The system is used to realize the automatic control method of the hoist of the intelligent water conservancy, and comprises a multi-dimensional parameter monitoring module, a historical data preprocessing and deep feature extraction module, an intelligent decision module, a dynamic adaptive control strategy module, an energy saving optimization and data interaction module, and a control effect feedback and model iteration module. The multi-dimensional parameter monitoring module is used to deploy multiple types of sensors on the hoist and its corresponding water conservancy facilities, to collect water conservancy environment parameters, device operation parameters, and environmental meteorological parameters in real time, to clean, format convert, and space-time align the data, and to synchronize the processed data to the intelligent water conservancy cloud platform. The historical data preprocessing and deep feature extraction module is used to obtain historical data of multi-dimensional data, to perform multi-dimensional preprocessing, to construct a deep feature extraction model based on the preprocessed historical data and the real-time collected data, and to extract deep features that can accurately reflect the water conservancy scene rules and device operation state. The intelligent decision module is used to construct an intelligent decision model that integrates multiple algorithms with deep features as input, to realize accurate decision-making of scene recognition, parameter calculation, and safety warning. The dynamic adaptive control strategy module is used to execute a dynamic adaptive control strategy by the hoist control system of the PLC controller according to the scene type, target control parameter, and safety warning result output by the intelligent decision model. The energy saving optimization and data interaction module is used to monitor energy consumption in real time, to intelligently optimize the hoist start-stop timing, to automatically enter a low-power sleep mode when the hoist is idle, to upload data in real time, and to issue remote control instructions through the intelligent water conservancy cloud platform by the management personnel, and to establish a data sharing protocol with other water conservancy systems in the basin. Control effect feedback and model iteration module: used for evaluating scheduling accuracy, device operation and scene adaptation, periodically incrementally training intelligent model, optimizing and upgrading algorithm, and verifying iteration effect.

Images