A machine learning-based adaptive control method for high-pressure and high-flow hydraulic systems

Abstract

The application discloses a kind of high-pressure large-flow hydraulic system self-adaptive control methods based on machine learning, it is related to mechanical engineering field, the method includes: real-time acquisition of hydraulic system pressure, flow, temperature and valve position data by sensor;Adopt sliding average filtering and linear interpolation to carry out two-stage data preprocessing;Three-layer long short-term memory network prediction model is constructed, and the time series prediction of pressure and flow is realized;According to the prediction result trigger pressure abnormal processing mechanism, adjust proportional relief valve and variable pump displacement;Based on flow prediction result, execute multiple actuator collaborative flow control;Through accumulator dynamic compensation, buffer valve parameter adjustment and variable pump response characteristic optimization collaborative suppression pressure impact.The application can realize accurate predictive control and fast dynamic response by the above technical scheme, control pressure overshoot in lower range, improve the operation reliability of high-pressure large-flow hydraulic system.

Description

Technical Field

[0001] This invention relates to the field of mechanical engineering, and in particular to an adaptive control method for high-pressure, high-flow hydraulic systems based on machine learning. Background Technology

[0002] With the rapid development of modern industry, hydraulic transmission and control technology has gradually become a key supporting technology in the field of modern mechanical engineering, playing an irreplaceable role in many industrial fields such as engineering machinery, metallurgical equipment, ship propulsion, and aerospace. Among them, the application of high-pressure, high-flow hydraulic systems is becoming increasingly widespread. These systems typically operate under high-pressure conditions of 20 to 40 MPa, with flow rates ranging from 200 to 1000 L / min. They are characterized by strong nonlinearity, drastic load changes, complex system coupling, and high energy consumption. Their control performance directly affects the operating efficiency, reliability, and service life of the entire equipment.

[0003] Achieving adaptive and precise control of high-pressure, high-flow hydraulic systems has always been a core issue in this field. Traditional hydraulic system control methods mainly rely on empirical parameter adjustment or simple feedback control strategies, which are insufficient to cope with complex and ever-changing operating conditions. With the rapid development of artificial intelligence technology, machine learning algorithms have provided a new technical path for the intelligent control of hydraulic systems. Existing technologies construct control models by acquiring historical and real-time operating data of the hydraulic system, realizing start-stop control and fault control functions, focusing on system operating condition identification and fault diagnosis. However, this approach is mainly aimed at coal mill hydraulic systems, whose working pressure and flow are relatively low. It lacks effective handling of pressure fluctuation suppression under high-pressure conditions, flow distribution control under high-flow conditions, and system stability during rapid high-pressure switching. Meanwhile, existing technologies also use distributed parameter models to describe the dynamic characteristics of hydraulic systems, combine Kalman filtering for state estimation, and utilize deep reinforcement learning algorithms to output control strategies, achieving real-time tracking of operating condition changes. However, this approach mainly focuses on general adaptive control under operating conditions and does not fully consider the special technical challenges faced by high-pressure, high-flow hydraulic systems. It has shortcomings in handling multivariable coupled control, rapid dynamic response, and system energy efficiency optimization under high-pressure, high-flow conditions.

[0004] This invention provides an adaptive control method for high-pressure, high-flow hydraulic systems based on machine learning, which improves the operational reliability of high-pressure, high-flow hydraulic systems under harsh working conditions, enables accurate prediction of system state within a preset time period, enhances control initiative, shortens response time, and solves the shortcomings of existing technologies in terms of rapid dynamic response. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive control method for high-pressure, high-flow hydraulic systems based on machine learning, in order to solve the problems of existing technologies in high-pressure maintenance, high-flow distribution, fast response, and energy consumption optimization.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] An adaptive control method for a high-pressure, high-flow hydraulic system based on machine learning, comprising:

[0008] Step 1, acquire hydraulic system operating data: Real-time acquisition of system operating parameters, including pressure data, flow data, temperature data and valve position opening data, through sensors deployed in key parts of the hydraulic system, and acquisition of historical data from the hydraulic system historical operating database;

[0009] Step 2, Preprocessing running data: The collected real-time data and historical data are preprocessed in two stages. In the first stage, the moving average filtering algorithm is used to smooth the data, and the window length is set to a preset value. In the second stage, the missing data periods are identified and the data is completed using the linear interpolation method.

[0010] Step 3, Construct a Long Short-Term Memory Network Prediction Model: Based on the preprocessed data, construct a three-layer Long Short-Term Memory Network model. Set the number of hidden units in each layer to a preset value. Use the Adam optimizer to train the model parameters. Set the initial learning rate to a preset value. Use early stopping to prevent the model from overfitting. After training, a prediction model that can be used for pressure and flow prediction is obtained.

[0011] Step 4, Pressure Prediction and Anomaly Handling: Based on the prediction model obtained in Step 3, input the current and historical operating data of the hydraulic system to obtain the pressure prediction value of the hydraulic system within a preset time in the future. When the pressure prediction value exceeds the preset safety threshold range, the corresponding pressure anomaly handling mechanism is triggered, including adjusting the proportional relief valve and the displacement of the variable pump.

[0012] Step 5, Flow Prediction and Distribution Control: Based on the prediction model obtained in Step 3, predict the flow demand of each actuator in the hydraulic system. Determine whether there is a flow distribution imbalance based on the prediction results. When the flow distribution is unbalanced, comprehensively consider the priority and flow demand of each actuator and formulate a flow distribution adjustment strategy. Achieve coordinated flow control of multiple actuators by adjusting the displacement of the variable pump and the opening of the proportional valve.

[0013] Step 6, Load mutation identification and pressure shock suppression: Based on the pressure prediction value and the flow prediction value, calculate the corresponding change rate. When the change rate of the pressure prediction value exceeds the preset pressure prediction value change rate threshold or the change rate of the flow prediction value exceeds the preset flow prediction value change rate threshold, it is determined to be a load mutation. The pressure shock suppression subroutine is started, and pressure shock is suppressed in a coordinated manner through accumulator dynamic compensation, buffer valve parameter adjustment and variable pump response characteristic optimization.

[0014] Step 7, Control Command Execution and Feedback: The control command execution module receives the decision commands generated in steps 4, 5 and 6, executes control actions by driving the proportional valve and variable pump, and monitors the actual response of the actuator in real time. It uses a dual closed-loop feedback mechanism to correct control deviations and ensure control accuracy.

[0015] Preferably, in step 1, the pressure sensor configured in the data acquisition module has a preset measurement range, a preset accuracy level, a preset response time, and an adjustable sampling frequency within a preset sampling frequency range; the flow sensor is a turbine sensor with a preset measurement range and a preset accuracy level; the temperature sensor is a platinum resistance temperature sensor with a preset temperature measurement range; and the valve position sensor is an angle sensor with a preset angle measurement range.

[0016] Preferably, in step 2, the first stage of data smoothing uses a moving average filtering algorithm. When the absolute value of the difference between a point value in the data sequence and the average value of a preset number of adjacent points exceeds a preset percentage threshold, the point is marked as a suspected abnormal point and replaced with the weighted average value of the surrounding data. In the second stage of data completion, when the number of consecutive missing data points exceeds a preset range, the extrapolated average value of the end values ​​of the preceding valid data segment is used for filling.

[0017] Preferably, in step 3, the long short-term memory network model introduces an attention mechanism to dynamically adjust the feature weights according to the contribution of each input feature to the prediction result, and expands the number of hidden units to a preset value to enhance the feature extraction capability.

[0018] Preferably, in step 4, the safety threshold range for judging pressure anomalies is dynamically set based on the rated working pressure of the hydraulic system. The lower limit is set to a first preset ratio of the rated pressure, the upper limit is set to a second preset ratio of the rated pressure, and the anomaly judgment response delay time is set to a preset delay time to avoid false triggering.

[0019] Preferably, in step 5, the flow distribution control module is configured with execution mechanism priority sorting logic, the lower limit of flow supply to high-priority execution mechanisms is not lower than the preset proportion of their rated flow, and when the total system flow is insufficient, the flow demand of high-priority execution mechanisms is guaranteed first.

[0020] Preferably, in step 6, the load change identification adopts a dual-threshold detection mechanism. The pressure change rate threshold is a preset pressure change rate threshold, and the flow change rate threshold is a preset flow change rate threshold. When either threshold is exceeded, it is determined to be a load change. The accumulator compensation control adopts an adaptive adjustment algorithm. The compensation coefficient is adjusted in real time according to the pressure deviation and the deviation change rate, and the adjustment range is within the preset adjustment range.

[0021] Preferably, it also includes an expert rule base module, which stores control rules based on the experience of domain experts, covering load prediction rules, fault diagnosis rules and safety protection rules. The total number of rules is a preset number, which is used to assist in handling control decisions under complex working conditions.

[0022] Preferably, it also includes a fault tolerance processing module, which uses the physical correspondence between the variable pump speed and displacement to perform data calculation and replacement when a sensor fault is detected; when a communication interruption is detected, the controller caches the control instruction sequence of the most recent preset time period and executes it according to the original plan, and performs instruction synchronization after the interruption is recovered.

[0023] Compared with the prior art, the beneficial technical effects of the present invention are as follows:

[0024] This invention effectively eliminates anomalies and noise in hydraulic system operating data by constructing a two-stage data preprocessing process, providing high-quality data input for subsequent prediction models. It also employs a three-layer long short-term memory network to establish a time-series prediction model for pressure and flow, enabling accurate prediction of the system state within a preset time period and providing a forward-looking basis for control decisions. Compared with the ex-post response mode of existing technologies, the initiative of control is improved.

[0025] This invention enables coordinated control of multiple actuators through a flow distribution control module based on flow prediction results, ensuring the rationality of flow supply for each actuator under high pressure and high flow conditions. The pressure shock suppression module effectively addresses load change scenarios through the synergistic effect of accumulator dynamic compensation, buffer valve parameter adjustment, and variable pump response characteristic optimization, controlling the pressure overshoot within a preset percentage range and shortening the response time to within a preset response time, thus solving the shortcomings of existing technologies in terms of rapid dynamic response.

[0026] This invention integrates a preset number of domain expert control rules through an expert rule base module, providing experience support for complex working conditions; the fault tolerance processing module maintains the basic operating capability of the system when sensor failure or communication interruption occurs through data calculation, instruction caching and synchronization mechanisms, improving the operational reliability of the high-pressure and high-flow hydraulic system under harsh working conditions and meeting the strict requirements of modern industry for intelligent control of high-pressure and high-flow hydraulic systems. Attached Figure Description

[0027] Figure 1This is a schematic diagram of the overall technical architecture of an adaptive control method for high-pressure, high-flow hydraulic systems based on machine learning proposed in this invention.

[0028] Figure 2 A schematic diagram illustrating the core principle framework of introducing the attention mechanism into the three-layer long short-term memory network prediction model in this invention;

[0029] Figure 3 This is a flowchart illustrating the logical flow of the two-stage data preprocessing in this invention.

[0030] Figure 4 This is a flowchart illustrating the logical flow framework of the coordinated control of pressure prediction and flow prediction allocation, as well as the identification of load mutations and suppression of pressure shocks in this invention.

[0031] Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow of the control command execution, dual closed-loop feedback mechanism, and expert rule base-assisted decision-making in this invention. Detailed Implementation

[0032] 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.

[0033] Example 1

[0034] Please refer to Figure 1 , Figure 1 This is a schematic diagram of the overall technical architecture of an adaptive control method for a high-pressure, high-flow hydraulic system based on machine learning, provided by an embodiment of the present invention. It includes: a data acquisition module, a data preprocessing module, an LSTM prediction model, pressure prediction and anomaly handling, flow prediction and distribution control, load mutation identification, control command execution, dual closed-loop feedback, an expert rule base, and fault tolerance processing.

[0035] Step 1 acquires hydraulic system operating data. This step provides the data input foundation for the entire adaptive control method. Sensors are deployed at key locations within the hydraulic system to achieve comprehensive perception of the system's operating status. The data acquisition module is equipped with a pressure sensor with a measurement range of 0 to 50 MPa, an accuracy class of 0.25, a response time within 1 millisecond, and an adjustable sampling frequency from 500 Hz to 2000 Hz. This range covers the entire typical working pressure of 20 to 40 MPa for high-pressure, high-flow hydraulic systems, while retaining sufficient measurement margin to cope with sudden pressure shocks. The flow sensor uses a turbine-type sensor with a range of 0 to 1500 L / min and an accuracy class of 0.5, accurately capturing the fluid flow status under high-flow conditions. The temperature sensor uses a platinum resistance temperature sensor with a temperature measurement range of -40℃ to 150℃, meeting the hydraulic system's operating temperature monitoring requirements. The valve position sensor uses an angle sensor with a measurement range of 0 to 90 degrees, used to monitor the opening position of proportional valves and variable valve mechanisms.

[0036] In actual deployment, pressure sensors are placed at the hydraulic pump outlet, key nodes of the main pipeline, and oil inlets of each actuator, forming a pressure monitoring network covering the entire system. Flow sensors are installed on the pump outlet main pipeline, each branch circuit, and the oil inlet pipeline of the actuator to achieve accurate monitoring of the system's flow distribution. Temperature sensors are placed near the oil tank, pump housing outlet, and key hydraulic components to reflect the real-time temperature rise of the hydraulic oil. Valve position sensors are directly connected to the control mechanisms of each proportional valve and variable pump to obtain feedback on the actual valve opening. The raw data collected by the sensors is transmitted to the data acquisition module via an industrial fieldbus. The data acquisition module is equipped with a signal conditioning circuit to amplify, filter, and perform analog-to-digital conversion on the sensor signals. The analog-to-digital converter uses 16-bit resolution, and a sample-and-hold circuit ensures synchronous acquisition of multi-channel data. The data acquisition module sends the processed digital signal to the data preprocessing unit at a preset sampling period.

[0037] Meanwhile, the hydraulic system's historical operation database stores complete operating records of the system over the past few months to years, covering system start-up and shutdown processes, typical operating data, load change records, fault event data, and equipment maintenance cycle information. The historical database adopts a time-series database storage format, and each record includes fields such as timestamp, pressure value, flow rate value, temperature value, valve position value, load status identifier, and system operating condition code. The data acquisition module supports querying the historical database by time interval, filtering by operating condition type, and searching by key parameter range. The acquired historical data and real-time acquired data together constitute the input data source for the subsequent data preprocessing stage.

[0038] Step 2: Preprocess the runtime data; such as Figure 3As shown, this step performs two-stage preprocessing on the collected real-time and historical data to ensure the data quality input to the prediction model. The first stage uses a moving average filtering algorithm to smooth the data, with a window length set to 5 sampling points. This window length has been verified in engineering to effectively filter out high-frequency random noise while retaining the true dynamic response characteristics of the hydraulic system. In the moving average filtering process, for an original data sequence of length N, the i-th data point after filtering is calculated using the following formula:

[0039]

[0040] in, For the filtered first The value of each data point For the first in the original data sequence The value of each data point The length of the sliding window, when < At that time, take the previous one. The average value of each point is used as the output. The sliding window slides sequentially on the data sequence. Each time the window moves, the arithmetic mean of the data within the current window is calculated as the processing result at that position.

[0041] Based on the moving average filtering, an outlier detection mechanism operates synchronously. When the absolute value of the difference between a point in the data sequence and the average of its five neighboring points exceeds a 10% threshold, that point is marked as a suspected outlier and replaced with the weighted average of its three surrounding points. The weight allocation is set according to the inverse distance principle, with closer data points receiving greater weight. The specific weight calculation formula is as follows: , The index represents the neighboring data points used to calculate the weighted average. After normalization, it is used for the weighted average calculation. The second stage identifies the missing data periods. When the number of consecutive missing data points is detected to be in the range of 1 to 5, a linear interpolation method is used to complete the data. The linear interpolation constructs an interpolation line based on the 3 valid data points before and after the missing period, and calculates the completed value according to the relative position of the missing points. When the number of consecutive missing data points exceeds 5, the extrapolated average of the last 3 data points at the end of the preceding valid data segment is used for filling. The extrapolated average is equal to the average of the last 3 data points at the end of the preceding valid data segment to avoid the linear interpolation introducing too large a deviation in the case of long-term missing data.

[0042] The data after two-stage preprocessing is organized in a time series format to form a complete input sequence containing historical time window data and current time data. Each input sequence contains system operation data from the past 32 sampling periods, covering time series features in four dimensions: pressure, flow rate, temperature, and valve position. The preprocessing unit outputs data to the feature engineering module, which constructs derived features based on the physical meaning of each dimension, including pressure change rate, flow accumulation, temperature gradient, and valve position change rate. These derived features are concatenated with the original time series data to form the final input feature vector of the prediction model.

[0043] Step 3: Construct a Long Short-Term Memory (LSTM) network prediction model, such as... Figure 2 As shown, a three-layer long short-term memory (LSTM) network model is constructed based on preprocessed data for time-series prediction of pressure and flow. This model can effectively capture long-term temporal dependencies and hidden states in the hydraulic system operation data. The specific configuration of the three-layer LSTM network model is as follows: The input layer receives a 32×8 dimension input feature vector, where 32 is the time step and 8 is the feature dimension of each time step, including pressure, flow, temperature, valve position, and corresponding rate of change features. The first LSTM network layer contains 256 hidden units and is configured with three gating mechanisms: input gate, forget gate, and output gate, as well as cell state memory units. The first layer is responsible for extracting low-level temporal features from the input sequence, mapping the 32×8 input sequence to a 32×256 feature representation. The second long short-term memory network layer contains 512 hidden units, receives the output sequence of the first layer, and further extracts mid-to-high-level temporal patterns. This layer is configured with a dropout regularization mechanism, with the dropout ratio set to 0.2 to prevent overfitting. The third long short-term memory network layer also contains 512 hidden units. Its output is aggregated with temporal information through a global average pooling layer and then passed to the fully connected output layer. The output dimension of the fully connected layer corresponds to the pressure value and flow value within the prediction time range.

[0044] The Long Short-Term Memory (LSTM) network model introduces an attention mechanism, dynamically adjusting feature weights based on the contribution of each input feature to the prediction result. The attention mechanism is implemented as follows: self-attention is applied to the output sequence of the third LSM network layer, and the self-attention weight matrix is ​​calculated using the following formula:

[0045]

[0046] in, , , These are the query matrix, key matrix, and value matrix, respectively, obtained from the output of the third layer through a linear transformation. The dimension of the key vector is used to scale the dot product result. The function normalizes the attention weights to a probability distribution form. The attention weight matrix is ​​multiplied by the value matrix to obtain the attention-weighted output. This output is concatenated with the original output of the third layer and then mapped to the prediction space through a fully connected layer. The attention mechanism enables the model to automatically identify the historical data points that have the greatest impact on the current prediction time and allocate more learning weights to these key time steps, thereby improving prediction accuracy.

[0047] Model parameters were trained using gradient descent optimization with the Adam optimizer. The initial learning rate was set to 0.001. The momentum parameter of the Adam optimizer was also set. Set to 0.9, second moment to estimate attenuation rate The numerical stability constant is set to 0.999. The model is set to 10 to the power of -8. During training, an early stopping method is used to prevent overfitting. The early stopping method is configured as follows: early stopping is triggered when the validation set loss does not decrease for 10 consecutive training epochs. After triggering, the model parameters with the lowest validation set loss are saved as the final model. The training data is divided into training set, validation set and test set in a ratio of 7:2:1. During training, the model is evaluated on the validation set every 5 training epochs. The validation set loss and root mean square error of pressure prediction are monitored. After the model training is completed, a prediction model that can be used for pressure and traffic prediction is obtained. The prediction model is exported in ONNX format and deployed to edge computing devices for real-time inference.

[0048] Please refer to Figure 4 , Figure 4 This is a logical flowchart of the coordinated control of pressure prediction and flow prediction allocation, as well as the identification of load mutations and suppression of pressure shocks in this invention, including: pressure prediction and anomaly handling, flow prediction and allocation control, and identification of load mutations and suppression of pressure shocks.

[0049] Step 4: Pressure Prediction and Anomaly Handling. Based on the prediction model obtained in Step 3, the current and historical operating data of the hydraulic system are input to predict the pressure change trend of the hydraulic system in the next 1 to 3 seconds. The prediction model takes the latest 32 time steps of data after preprocessing in Step 2 as input and outputs the pressure prediction values ​​for the next 32 time steps. The prediction time window covers the range of 1 to 3 seconds. The pressure prediction module adopts a rolling prediction strategy, triggering a prediction calculation every 200 milliseconds. Each prediction outputs the pressure prediction values ​​for the next 1 second, 2 seconds and 3 seconds.

[0050] The safety threshold range for pressure anomaly detection is dynamically set based on the rated working pressure of the hydraulic system. The rated working pressure is set to 31.5 MPa, the lower limit of the safety threshold is set to 85% of the rated pressure, i.e., 26.78 MPa, and the upper limit of the safety threshold is set to 110% of the rated pressure, i.e., 34.65 MPa. The dynamic threshold mechanism is adaptively adjusted according to the actual operating status of the system. When the system operates in the high-pressure range for a long time, the lower limit threshold can be appropriately increased to adapt to changes in operating conditions. The anomaly detection response delay time is set to 50 milliseconds to avoid false triggering caused by instantaneous pressure fluctuations. The delay detection mechanism requires that the predicted pressure exceeds the threshold range for three consecutive samplings before it is confirmed as a real anomaly.

[0051] When the predicted pressure exceeds the safety threshold, a corresponding pressure anomaly handling mechanism is triggered. The pressure anomaly handling mechanism is divided into two modes based on the type of over-limit: high-pressure anomaly handling and low-pressure anomaly handling. When the predicted pressure is detected to exceed the upper limit threshold of 34.65 MPa, a command is first sent to the proportional relief valve control module to quickly adjust the relief valve setting pressure to the current predicted pressure plus a safety margin of 2 MPa. At the same time, a command is sent to the variable pump displacement control module to reduce the variable pump displacement by 15% to 20% to reduce the pump source flow input. When the predicted pressure is detected to be below the lower limit threshold of 26.78 MPa, the variable pump displacement is first increased by 10% to 15%, and the opening of the proportional relief valve is appropriately closed to reduce the leakage flow. The pressure anomaly handling command is also sent to the control command execution module in step 7 for action execution, and the abnormal event is recorded in the system log for subsequent analysis.

[0052] Step 5: Flow prediction and distribution control. Based on the prediction model obtained in Step 3, the flow demand of each actuator in the hydraulic system is predicted. A typical configuration of the hydraulic system includes 3 to 5 actuators. Each actuator is equipped with an independent flow sensor and a proportional valve. The flow prediction module takes the historical flow data of each actuator as input and predicts the flow demand curve of each actuator in the next 1 to 3 seconds. The flow prediction results are output in matrix form, with rows corresponding to the actuator number and columns corresponding to the prediction time nodes.

[0053] The flow distribution control module determines whether there is a flow distribution imbalance based on the prediction results. The criteria for determining a flow distribution imbalance are: the deviation between the predicted flow demand of any actuator and the current allocated flow exceeds 10% of its rated flow, or the total predicted flow demand of the system exceeds 95% of the maximum output flow of the variable pump. When a flow distribution imbalance is determined, the flow distribution control module comprehensively considers the priority and flow demand of each actuator and formulates a flow distribution adjustment strategy. The priority ranking logic of the actuators is determined according to the preset working conditions. In the typical priority configuration, the actuators used for safety braking function have the highest priority, the actuators used for main operation function have the second highest priority, and the actuators used for auxiliary function have the lowest priority. The lower limit of the flow supply for high-priority actuators is not less than 70% of their rated flow. When the total flow of the system is insufficient, the flow demand of high-priority actuators is guaranteed first.

[0054] The specific calculation process of the flow allocation adjustment strategy is as follows: First, calculate the comprehensive adjustment coefficient based on the priority weight of each actuator, the current flow deviation, and the predicted flow demand; second, allocate the total available flow of the system to each actuator according to the comprehensive adjustment coefficient; finally, convert the allocation result into the variable pump displacement setpoint and the proportional valve opening setpoint. The flow allocation control module achieves coordinated flow control of multiple actuators by adjusting the variable pump displacement and the proportional valve opening. The variable pump displacement adjustment range is 20% to 100% of its rated displacement, and the proportional valve opening adjustment range is 0% to 100% of its rated opening. The flow allocation control command is sent to the control command execution module in step 7, and the execution effect of the flow allocation strategy is fed back to the prediction model in step 3 for model fine-tuning.

[0055] Step 6: Load mutation identification and pressure shock suppression. Based on the prediction results of Steps 4 and 5, load mutation characteristics are identified. Load mutation identification adopts a dual-threshold detection mechanism: the threshold for the rate of change of the pressure prediction value is 5 MPa / second, and the threshold for the rate of change of the flow prediction value is 50 L / min / second. When either threshold is exceeded, it is determined to be a load mutation. After the load mutation event is triggered, the system enters the pressure shock suppression mode. The load mutation identification module continuously monitors the pressure prediction value and the flow prediction value, and calculates the corresponding rate of change. When the rate of change of the pressure prediction value exceeds the threshold for three consecutive sampling periods, or the rate of change of the flow prediction value exceeds the threshold for three consecutive sampling periods, the load mutation event is confirmed and the pressure shock suppression subroutine is started.

[0056] The pressure shock suppression subroutine employs a multi-level collaborative control strategy to effectively suppress pressure shocks. The first level is accumulator dynamic compensation control. The accumulator pre-charge pressure is set to 60% of the system's rated pressure. When a sudden pressure drop is detected, the accumulator releases hydraulic oil to the system for pressure compensation. The compensation amount is calculated in real time based on the pressure deviation and the rate of change of the deviation. The accumulator compensation control uses an adaptive adjustment algorithm, and the compensation coefficient is adjusted in real time based on the pressure deviation and the rate of change of the deviation, with an adjustment range of 0.5 to 2.0. When the pressure deviation is large, the compensation coefficient is increased to accelerate pressure recovery; when the pressure deviation is small, the compensation coefficient is decreased to avoid pressure oscillations caused by overcompensation. The dynamic response time of the accumulator is controlled within 10 milliseconds to ensure timely response to rapid pressure changes.

[0057] The second level is the buffer valve parameter adjustment control. The buffer valve is located near the inlet and outlet of the actuator to absorb pressure shock energy. When a sudden load change event is triggered, the throttle opening of the buffer valve is adjusted in real time according to the current pressure change rate. The adjustment range of the throttle opening is 30% to 100% of its maximum opening. When the pressure change rate is positive, i.e., the pressure is rising, the throttle opening of the buffer valve is increased to accelerate the discharge of hydraulic oil and suppress pressure overshoot. When the pressure change rate is negative, i.e., the pressure is falling, the throttle opening of the buffer valve is decreased to slow down the flow of hydraulic oil and prevent cavitation.

[0058] The third level is the variable pump response characteristic optimization control. The displacement adjustment response time of the variable pump directly affects the system pressure recovery speed. After a load change event is triggered, the response characteristic parameters of the variable pump are temporarily adjusted. The maximum rate of change of pump displacement is increased from 5% / second under normal conditions to 15% / second, and the integral gain of the pump controller is increased from 0.5 under normal conditions to 2.0 to accelerate the displacement adjustment speed. At the same time, the pressure-flow characteristic curve of the variable pump is compensated in real time, and the pump power limiting parameters are adjusted according to the deviation between the current pressure and the target pressure to maximize the pump response speed while ensuring system safety.

[0059] After the control commands at the three levels are generated in tandem, they are sent to the control command execution module in step 7. At the same time, step 6 monitors the effect of pressure shock suppression. When the pressure overshoot is controlled within 5% of the rated pressure and the response time is shortened to within 500 milliseconds, the pressure shock suppression is confirmed to be successful, and the variable pump response characteristic parameters gradually return to the normal set value.

[0060] Step 7: Control command execution and feedback, such as... Figure 5As shown, the control command execution module receives the decision commands generated in steps 4, 5, and 6, and constructs a unified control command queue. The control command queue is sorted according to timestamp and priority. Commands with the same timestamp are executed sequentially in the order of steps 4, 5, and 6, while commands with different timestamps are executed in chronological order. The control command execution module executes control actions by driving a proportional valve and a variable pump. The proportional valve control signal is output using pulse width modulation with a modulation frequency of 500Hz and a control accuracy of 0.1%. The variable pump displacement control signal is output using analog voltage with a voltage range of 0 to 10V, corresponding to a displacement adjustment range of 20% to 100%.

[0061] A dual-loop feedback mechanism is used to ensure control accuracy and system stability. The outer loop is the position loop, which uses the target value in the control command as the setpoint and the actual displacement or rotation angle of the actuator as the feedback value. The proportional gain of the position loop is set to 5, and the integral time is set to 0.5 seconds. The inner loop is the velocity loop, which uses the output of the position loop as the setpoint and the actual speed of the actuator as the feedback value. The proportional gain of the velocity loop is set to 10, and the integral time is set to 0.2 seconds. The outputs of the position loop and the velocity loop are limited and then sent to the actuator as the actual control quantities of the proportional valve and the variable pump. The feedback data is collected in real time by the angle sensor and speed sensor installed on the actuator, and the sampling frequency is synchronized with the control cycle at 200 milliseconds.

[0062] The dual closed-loop feedback mechanism monitors the actual response of the actuator in real time and corrects the control deviation. When the control deviation exceeds the allowable range, the feedback adjustment module automatically adjusts the control gain to compensate for the time-varying characteristics of the system. When abnormal feedback data is detected or the control command execution timeout exceeds 100 milliseconds, the fault detection module triggers the fault handling process and notifies the fault tolerance processing module in step 6.

[0063] The expert rule base module stores control rules based on the experience of domain experts, covering load prediction rules, fault diagnosis rules, and safety protection rules. The expert rule base is configured with a total of 50 rules. The rule format adopts the form of condition-action pairs. Each rule includes a rule number, a list of antecedent conditions, and a list of consequent actions. The load prediction rule predicts the future load change trend based on the characteristics of the current operating condition and historical similar operating condition data. The fault diagnosis rule judges the fault type and severity based on the abnormal patterns of sensor data and the degree of deviation of system operating parameters. The safety protection rule triggers protective actions such as emergency shutdown or power reduction when a dangerous operating condition is detected. The expert rule base module matches the real-time operating condition through a rule reasoning engine. When a rule that meets the triggering conditions is matched, the consequent action of the rule is sent to the control instruction execution module as a supplementary control instruction.

[0064] The fault-tolerant processing module provides the system with fault-tolerant operation capability in the event of sensor failure or communication interruption. When a sensor failure is detected, the fault-tolerant processing module uses the physical correspondence between the variable pump speed and displacement to perform data estimation and substitution. There is an approximately linear relationship between the displacement and speed of the variable pump. The displacement is equal to the pump's geometric displacement multiplied by the speed and then multiplied by the volumetric efficiency. The flow rate value is calculated by monitoring the motor speed and known pump parameters to replace the measured value of the faulty sensor. When a communication interruption is detected, the controller caches the control command sequence of the most recent 5 seconds and executes it according to the original plan. After the interruption is restored, the command is synchronized. The actual execution results and sensor data during the communication interruption are recorded to the data recording module.

[0065] In a specific application scenario, assuming the high-pressure, high-flow hydraulic system is configured with a three-actuator parallel structure, the system rated pressure is 31.5MPa, the maximum displacement of the variable pump is 500L / min, and the rated flow rates of the three actuators are 200L / min, 150L / min, and 150L / min, respectively. During system operation, the data acquisition module in step 1 acquires data from the pressure sensor, flow sensor, temperature sensor, and valve position sensor at a sampling frequency of 1000Hz. The analog-to-digital converter inside the data acquisition module converts the analog signal into a 16-bit digital quantity, which is then transmitted to the data preprocessing unit via industrial Ethernet. The data preprocessing unit in step 2 performs a moving average filter on the raw data, with a window length of 5 sampling points, i.e., a processing window of 5 milliseconds. The outlier detection threshold is set to 10% of the average of the five adjacent points. Outliers are replaced by the weighted average of the surrounding three points, and missing data is filled by linear interpolation. The preprocessed data is organized into an input sequence in units of 32 time steps and input into the three-layer long short-term memory network prediction model deployed on the edge computing device in step 3.

[0066] After receiving a new input sequence, the prediction model performs forward inference calculations. The calculation process includes temporal feature extraction of three long short-term memory network layers, calculation and weighted output of self-attention weights, and output mapping of the fully connected layer. The prediction model outputs the pressure prediction value for the next 32 time steps and the flow prediction value for the three actuators, with the prediction time domain covering a range of 1 to 3 seconds. The pressure prediction and anomaly handling module in step 4 judges the system pressure status based on the prediction results. When the predicted pressure exceeds the dynamic safety threshold range, the corresponding pressure anomaly handling mechanism is triggered. The flow prediction and allocation control module in step 5 executes the flow allocation strategy based on the flow prediction value of the three actuators. The high-priority actuators are guaranteed a flow supply limit of no less than 70% of their rated flow.

[0067] In the load change identification and pressure shock suppression scenario, when the pressure change rate is detected to exceed 5 MPa / s, the load change identification module confirms the load change event and starts the pressure shock suppression subroutine. The accumulator compensation control calculates the compensation coefficient based on the pressure deviation, with an adaptive adjustment range of 0.5 to 2.0. The accumulator compensation response time is controlled within 10 milliseconds. The buffer valve parameter adjustment control adjusts the throttle opening according to the direction of the pressure change rate, with an adjustment range of 30% to 100% of the maximum opening. The variable pump response characteristic optimization control increases the maximum rate of pump displacement change to 15% / s and the integral gain to 2.0. The coordinated control of the three levels effectively suppresses the pressure shock caused by the load change, controls the pressure overshoot within 5% of the rated pressure, and shortens the response time to within 500 milliseconds.

[0068] Step 7's control command execution module receives all decision commands and executes them in timestamp order. It drives proportional valves and variable pumps to achieve control actions. A dual closed-loop feedback mechanism continuously monitors the actual response of the actuators. The gain parameters of the position loop and velocity loop are adaptively adjusted according to the system's operating status. The expert rule base module matches relevant control rules based on real-time operating conditions, and the rule matching results provide decision support as supplementary control strategies. The fault tolerance processing module monitors the sensor status and communication status in real time. When a sensor failure is detected, a flow estimation alternative is activated. When a communication interruption is detected, a control command caching execution scheme is activated to ensure the system's basic operational capability under fault conditions.

[0069] Example 2

[0070] To meet the needs of different application scenarios, this invention also provides a variant embodiment of the adaptive control method for high-pressure, high-flow hydraulic systems based on machine learning. This variant embodiment is basically the same as Embodiment 1 in terms of the core prediction model and control strategy; the main differences are as follows:

[0071] In the data acquisition stage, this variant expands the sensor configuration by adding a vibration sensor and an oil quality sensor. The vibration sensor is a piezoelectric accelerometer with a measurement range of ±50g and a frequency response range of 2Hz to 10kHz. It is used to monitor the vibration status of the hydraulic pump and actuator to assist in fault diagnosis. The oil quality sensor is used to monitor the contamination level and water content of the hydraulic oil online. The contamination level monitoring range is NAS 1638 level 6 to 16, and the water content monitoring range is 0 to 1000ppm. The vibration data and oil quality data, together with the original pressure, flow rate, temperature, and valve position data, constitute the expanded feature vector, which is input into the prediction model for comprehensive state prediction.

[0072] Regarding the prediction model, this variant adds a convolutional neural network feature extraction front-end to the three-layer long short-term memory network. The convolutional neural network front-end adopts a one-dimensional convolutional structure with a kernel size of 3, a kernel count of 16, a stride of 1, and the ReLU activation function. The convolutional layer extracts local features from the original time series data. The extracted feature maps are then concatenated with the output of the long short-term memory network layer after global average pooling and input together to the fully connected output layer. The joint structure of the convolutional front-end and the long short-term memory network can better capture local patterns and global dependencies in the time series data, improving the prediction accuracy after the fusion of multi-source heterogeneous data.

[0073] In terms of flow allocation control, this variant embodiment introduces a quadratic programming optimization algorithm for flow allocation calculation. The objective function of the flow allocation optimization problem is to minimize the weighted sum of squares of the flow deviations of each actuator. The constraints include total flow constraints, upper and lower limits of flow for each actuator, and priority guarantee constraints. The quadratic programming optimization algorithm solves the optimal flow allocation scheme in each control cycle, replacing the comprehensive adjustment coefficient allocation method in Embodiment 1, thereby improving the global optimality of flow allocation and the strictness of constraint satisfaction.

[0074] In terms of pressure shock suppression, this variant adds a fourth level of control, namely hydraulic cylinder buffer control. The hydraulic cylinder is equipped with a stroke end buffer device. When a sudden load change is detected, causing the actuator to move rapidly, the buffer device adjusts the buffer throttle opening in real time according to the movement speed and displacement to prevent the hydraulic cylinder from hitting the end position. The hydraulic cylinder buffer control judges the movement state by detecting the data change rate of the hydraulic cylinder displacement sensor. When the displacement change rate exceeds the threshold, the buffer control mode is triggered. The movement speed is slowed down by reducing the oil discharge throttle area of ​​the buffer chamber, thus achieving smooth deceleration.

[0075] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An adaptive control method for a high-pressure, high-flow hydraulic system based on machine learning, characterized in that, Includes the following steps: Step 1: Acquire hydraulic system operating data. Collect system operating parameters in real time through sensors, and at the same time acquire historical data of the hydraulic system. Step 2: Preprocess the running data. A two-stage preprocessing method combining moving average filtering and linear interpolation is used to preprocess the collected real-time data and historical data. Step 3: Construct a Long Short-Term Memory Network Prediction Model. Based on the preprocessed data, construct a three-layer Long Short-Term Memory Network model and use the Adam optimizer combined with the early stopping method to train the model parameters. Step 4, Pressure Prediction and Anomaly Handling: Input the current and historical operating data of the system into the prediction model to obtain the pressure prediction value within a preset time period in the future. When the pressure prediction value exceeds the preset safety threshold range, the corresponding pressure anomaly handling mechanism is triggered. Step 5, Flow forecasting and distribution control: The flow demand of each actuator is predicted by the forecasting model to obtain the flow forecast value. It is then determined whether there is a flow distribution imbalance. When there is a flow distribution imbalance, a flow distribution adjustment strategy is formulated to adjust the displacement of the variable pump and the opening of the proportional valve. Step 6, load change identification and pressure shock suppression: Based on the pressure prediction value and the flow prediction value, load changes are identified by dual thresholds of the rate of change of the pressure prediction value and the rate of change of the flow prediction value. When either rate of change exceeds the corresponding preset threshold, the pressure shock suppression subroutine is activated. Pressure shocks are suppressed in a coordinated manner through accumulator dynamic compensation, buffer valve parameter adjustment and variable pump response characteristic optimization. Step 7, Control command execution and feedback: The control command execution module receives the decision commands generated in steps 4, 5 and 6, and corrects the control deviation by controlling the proportional valve and variable pump in combination with real-time monitoring of the actual response of the actuator and using a double closed-loop feedback mechanism.

2. The adaptive control method for high-pressure, high-flow hydraulic systems based on machine learning according to claim 1, characterized in that, In step 1, the data acquisition module is equipped with a pressure sensor, a flow sensor, a temperature sensor, and a valve position sensor to acquire pressure data, flow data, temperature data, and valve position opening data. At the same time, based on the hydraulic system's historical operation database, it acquires system start-up and shutdown process data, typical operating condition data, load change records, fault event data, and equipment maintenance cycle information.

3. The adaptive control method for high-pressure, high-flow hydraulic systems based on machine learning according to claim 1, characterized in that, In step 2, the two-stage preprocessing method involves a first-stage data smoothing process using a moving average filtering algorithm. When the absolute value of the difference between a point in the data sequence and the average value of a preset number of adjacent points exceeds a preset percentage threshold, the point is marked as a suspected outlier and replaced with the weighted average of the surrounding data. In the second-stage data completion process, when the number of consecutive missing data points is within a preset range, a linear interpolation method is used for data completion. The linear interpolation constructs an interpolation line based on a preset number of valid data points before and after the missing time period. When the number of consecutive missing data points exceeds the preset range, the extrapolated average of the last preset number of data points at the end of the preceding valid data segment is used to fill the gap.

4. The adaptive control method for high-pressure, high-flow hydraulic systems based on machine learning according to claim 1, characterized in that, In step 3, the three-layer long short-term memory network model captures long-term temporal dependencies and hidden states of the network structure in the hydraulic system operation data to predict the temporal sequence of system pressure and flow. The specific network structure is as follows: the input layer includes eight feature dimensions, including system pressure, flow, temperature, valve position and corresponding rate of change features; the first long short-term memory network layer is used to extract low-dimensional basic temporal features. The second long short-term memory (LSM) network layer has more hidden units than the first LSM network layer and incorporates a random deactivation mechanism to randomly deactivate some hidden units during model training, reducing the risk of overfitting. The third LSM network layer uses the same number of hidden units as the second LSM network layer and incorporates a global average pooling layer to extract global temporal features of the system's operating state, while reducing the parameter scale of subsequent fully connected layers. The fully connected output layer maps the output results of the three LSM network layers to the actual predicted values.

5. The adaptive control method for high-pressure, high-flow hydraulic systems based on machine learning according to claim 4, characterized in that, The Long Short-Term Memory (LSTM) network model introduces an attention mechanism. After the third LSM network layer, the historical time-series features output by the third LSM network layer are dynamically weighted. This strengthens the model's attention to key time steps such as pressure surges, traffic anomalies, and load changes, thereby improving the model's accuracy in predicting system pressure and traffic and its dynamic response performance under complex operating conditions.

6. The adaptive control method for high-pressure, high-flow hydraulic systems based on machine learning according to claim 1, characterized in that, In step 6, the load mutation identification adopts a dual-threshold monitoring mechanism. The dual-threshold monitoring mechanism uses a pressure prediction value change rate threshold and a flow prediction value change rate threshold. When either threshold is exceeded, it is determined to be a load mutation. After the load mutation event is triggered, the system enters the pressure shock suppression mode. The load mutation identification module continuously monitors the pressure prediction value and the flow prediction value, and calculates the corresponding change rate. When it is detected that the pressure prediction value change rate exceeds the threshold for three consecutive sampling periods or the flow prediction value change rate exceeds the threshold for three consecutive sampling periods, the load mutation event is confirmed and the pressure shock suppression subroutine is started.

7. The adaptive control method for high-pressure, high-flow hydraulic systems based on machine learning according to claim 6, characterized in that, The pressure shock suppression subroutine employs a three-layer collaborative control strategy to suppress pressure shocks. The first layer uses accumulator dynamic compensation control, adjusting the compensation coefficient in real time based on pressure deviation and deviation change rate using an adaptive adjustment algorithm. The second layer uses a buffer valve for adjustment control, adjusting the buffer valve throttle opening in real time based on the current pressure change rate to absorb pressure shock energy. The third layer uses variable pump response characteristic optimization control, temporarily adjusting the response characteristic parameters of the variable pump to accelerate the displacement adjustment speed, while simultaneously compensating the pressure-flow characteristic curve of the variable pump in real time.

8. The adaptive control method for high-pressure, high-flow hydraulic systems based on machine learning according to claim 1, characterized in that, In step 7, the dual closed-loop feedback mechanism is divided into an outer loop and an inner loop. The outer loop is the position loop, which uses the target value in the control command as the set value and the actual displacement or rotation angle of the actuator as the feedback value to ensure that the actuator accurately reaches the set position and outputs the target speed signal to the inner loop. The inner loop is the speed loop, which uses the output of the position loop as the set value and the actual movement speed of the actuator as the feedback value. It is used for rapid response, adjusting hydraulic valves and variable pumps, suppressing pressure fluctuations, and improving the dynamic response speed of the system.

9. The adaptive control method for high-pressure, high-flow hydraulic systems based on machine learning according to claim 1, characterized in that, It also includes an expert rule base module, which stores control rules based on the experience of domain experts, covering load prediction rules, fault diagnosis rules and safety protection rules. The rules are represented in the form of condition-action pairs to assist in control decisions under complex operating conditions. The expert rule base module matches real-time operating conditions through a rule reasoning engine. When a rule that meets the triggering conditions is matched, the rule's follow-up action is sent to the control instruction execution module as a supplementary control instruction.

10. The adaptive control method for high-pressure, high-flow hydraulic systems based on machine learning according to claim 1, characterized in that, It also includes a fault tolerance processing module, which uses the physical correspondence between the speed and displacement of the variable pump to calculate and replace data when a sensor fault is detected; when a communication interruption is detected, the controller caches the control instruction sequence of the most recent preset time period and executes it according to the original plan. After the interruption is recovered, the instruction is synchronized and the actual execution results are supplemented.

Images