Hydraulic element flow characteristic online self-calibration method and system based on multi-sensor information fusion

Abstract

This invention discloses a method and system for online self-calibration of flow characteristics of hydraulic components based on multi-sensor information fusion. First, valve spool opening or pump speed, differential pressure, oil temperature, and flow rate under healthy conditions are collected to construct a source domain flow soft measurement model based on a backpropagation neural network. Second, a velocity sequence excitation trajectory is designed to traverse typical operating conditions with multiple openings and differential pressures, and a high-quality target domain dataset is extracted based on an adaptive sliding window and continuous steady-state constraints. Then, a transfer learning method with parameter constraints is used to fine-tune and update the source domain model, reconstructing the nonlinear flow mapping under wear conditions to obtain a target domain flow soft measurement model. Finally, the updated model is inverted and embedded into a backstepping controller to replace the degraded and failed physical model of the hydraulic component. This invention achieves online self-sensing and self-calibration of flow characteristics of hydraulic components under wear conditions without relying on physical flow meters, effectively improving the flow control accuracy throughout the system's life cycle.

Description

Technical Field

[0001] This invention belongs to the field of fluid transmission and control technology, specifically relating to an online self-calibration method and system for the flow characteristics of hydraulic components based on multi-sensor information fusion. Background Technology

[0002] Hydraulic components, as the core control and power elements of hydraulic systems, are crucial for ensuring the control accuracy, response speed, and operational stability of the entire system through their precise flow control characteristics. During long-term service, the combined effects of high pressure, high-speed fluid erosion, oil contamination particle wear, and mechanical friction inevitably cause wear and degradation in key internal moving parts (such as valve core-valve sleeve and cylinder-plunger), leading to increased clearances and internal leakage. This gradual physical wear directly causes a significant drift in the flow characteristics of the hydraulic components (i.e., the nonlinear mapping relationship between output flow and valve core opening / pump speed and pressure difference), resulting in decreased control accuracy, sluggish response, and even oscillations in the hydraulic system.

[0003] To address the aforementioned issues, the flow rate of valve-controlled (or pump-controlled) systems in the prior art is typically obtained using the following two methods.

[0004] (1) Direct measurement via physical flow sensors. While this method can provide relatively direct flow measurement results, high-precision, fast-response flow sensors (especially those suitable for high-flow conditions) generally suffer from prominent problems such as high cost, low resistance to contamination, large size, and stringent installation requirements (e.g., requiring long straight pipe sections). In application scenarios with compact structures and harsh environments (such as mines and aerospace), its application is greatly limited.

[0005] Secondly, indirect estimation is based on mathematical models with fixed parameters. This method typically establishes a mathematical model based on the calibration data of hydraulic components at the time of manufacture or their physical parameters under healthy conditions. However, such fixed-parameter models are difficult to accurately characterize the flow characteristic drift caused by factors such as the nonlinear increase in fit clearance and complex changes in internal leakage during the actual wear evolution of hydraulic components. Especially under complex operating conditions such as high pressure differential, high temperature, and small opening, the model parameter mismatch problem is particularly prominent, resulting in a significant decrease in flow estimation accuracy and failing to meet the requirements of precise control.

[0006] Furthermore, most existing valve-controlled (or pump-controlled) system control methods rely on the aforementioned fixed mathematical models or idealized physical parameters to design control laws (such as PID control). These controllers generally lack the ability to dynamically perceive and adapt to the wear and degradation process of hydraulic components. When the actual flow characteristics of a component change due to wear, the internal model on which the controller relies gradually becomes ineffective, and the control system is prone to problems such as increased tracking errors and decreased stability, making it difficult to maintain high-performance control throughout the entire life cycle of the component.

[0007] In summary, there is an urgent need to propose a high-precision flow sensing and self-calibration method that is independent of or has minimal reliance on physical flow sensors and can adapt to the wear and degradation of hydraulic components. This sensing capability should be deeply integrated with the control method to maintain high-precision control performance of the hydraulic system under long-term service conditions. However, current technologies lack a systematic technical solution that can effectively address all of the aforementioned problems. Summary of the Invention

[0008] In view of this, the purpose of the present invention is to provide a method and system for online self-calibration of the flow characteristics of hydraulic components based on multi-sensor information fusion, which aims to achieve online self-sensing and reconstruction of the flow characteristics of hydraulic components after wear without relying on physical flow meters, and to improve the flow control accuracy of the system under long-term service wear conditions.

[0009] To achieve the above objectives, the present invention provides the following technical solution: This invention first proposes an online self-calibration method for the flow characteristics of hydraulic components based on multi-sensor information fusion, comprising the following steps: Step 1: Collect data on valve spool opening or pump speed, system differential pressure, oil temperature and flow rate under the healthy state of hydraulic components, construct a source domain dataset, and establish a source domain flow soft measurement model based on backpropagation neural network. Step 2: Construct the excitation trajectory of the velocity sequence, and realize the traversal of hydraulic components under typical working conditions of multiple openings, multiple speeds, and multiple pressure differences through multi-speed reciprocating control of hydraulic cylinders to obtain the original data of the target domain; Step 3: Construct an adaptive sliding time window based on the hydraulic cylinder operating cycle, and combine standard deviation, coefficient of variation and continuous steady-state constraints to perform steady-state identification, sample screening and data reconstruction to construct the target domain sample set; Step 4: The source domain flow soft measurement model is fine-tuned and updated using a transfer learning method with parameter constraints. The nonlinear mapping relationship between valve core opening or pump speed, pressure difference, temperature and output flow under the wear state of hydraulic components is reconstructed to obtain the target domain flow soft measurement model. Step 5: Construct a backstepping control strategy that integrates the target domain flow soft measurement model. Embed the updated target domain flow soft measurement model into the backstepping controller to replace the hydraulic component physical model that is inaccurate due to wear and degradation, thereby realizing online self-calibration and control of flow characteristics under hydraulic component wear conditions.

[0010] Furthermore, in step one, the source domain flow soft measurement model uses valve core opening or pump speed, system pressure difference and oil temperature as input feature vectors, and output flow as supervision label; the backpropagation neural network includes an input layer, a hidden layer and an output layer, constructs an error loss function between predicted flow and measured flow, and uses gradient descent algorithm to iteratively optimize network weights and bias parameters.

[0011] Furthermore, in the backpropagation neural network, the nonlinear mapping from the input layer to the first hidden layer is as follows: in: x is the j-th neuron in the first hidden layer; v Valve core opening; The pressure difference between the valve inlet and outlet is T; T is the temperature. The weights from the valve core opening to the j-th neuron in the first hidden layer; The value of the pressure difference between the valve inlet and outlet is the weight of the j-th neuron in the first hidden layer; The weights from temperature to the j-th neuron in the first hidden layer; The threshold value is the value of the j-th neuron from the input layer to the first hidden layer. The activation function corresponding to the input layer to the first hidden layer; The nonlinear mapping between two adjacent hidden layers is: in: For the current number The output value of the j-th neuron in the hidden layer; For the first The output of the i-th neuron in the hidden layer; For the first The hidden layer has a neuron i pointing to the first neuron. The connection weights of the j-th neuron in the hidden layer; For the first The bias of the j-th neuron in the hidden layer b; This is the nonlinear activation function between two adjacent hidden layers; This represents the number of neurons in the hidden layer.

[0012] The nonlinear mapping from the last hidden layer to the output layer is: in: For the predicted valve flow rate; The weights from the j-th neuron in the last hidden layer to the output neuron are: The threshold of the output neuron; The constructed error loss function is as follows: in: Let it be the error loss function; The actual measured traffic volume is labeled; This is a flow soft measurement mapping function established based on a neural network; For the first The input feature vector of each sample includes valve core opening, pressure difference, and temperature; For the first The actual traffic label corresponding to each sample; The total number of training samples; The gradient descent algorithm is used to iteratively optimize the network weights and bias parameters. The method is as follows: in, The learning rate determines the step size for parameter updates; The gradient of the error with respect to the weights reflects the direction of backpropagation of the error. This is the momentum factor, used to accelerate convergence and prevent the model from getting trapped in local optima.

[0013] Furthermore, in step two, the method for constructing the velocity sequence excitation trajectory is as follows: Discretize the hydraulic cylinder's movement speed as The target velocities constitute a velocity set: in, The target uniform velocity is the i-th velocity; and These are the minimum and maximum constraint speeds at which the hydraulic cylinder can operate stably; Combining the physical stroke range of the hydraulic cylinder With safety margin Determine the effective range of motion and effective walking length per one trip : Regarding the first Each speed level constructs a smooth velocity trajectory including acceleration, steady-state constant speed, and deceleration. : in: This represents the total one-way travel time. For acceleration and deceleration duration; For a single-journey local time variable; Apply a closed-loop displacement constraint to the extension and retraction processes to ensure that the net displacement in a single cycle is zero. .

[0014] Furthermore, in step three, the method for constructing the target domain sample set is as follows: Based on the effective travel length of a single stroke of the hydraulic cylinder relative to target speed and sampling step size Calculate the total number of sampling points for a single trip. And adaptive sliding window length: in: This represents the total number of sampling points for a single trip. The proportion of the window to the length of a single-pass sample; The length of the sliding window corresponding to the i-th velocity; Minimum window length; Maximum window length; In the Within a sliding window, the coefficients of variation of hydraulic cylinder speed, valve core opening, and system pressure difference are constructed as steady-state indicators: in: This is an indicator of the speed stability of the hydraulic cylinder; The standard deviation of the velocity; For the first Average speed of each window; This is an indicator of valve core opening stability. The standard deviation of valve core opening; For the first The average valve core opening of each window; This is an indicator of system differential pressure stability. This represents the standard deviation of the system pressure difference. For the first Average system differential pressure across windows; Introducing a continuous steady-state constraint mechanism, only when continuous When all windows meet the steady-state conditions, the system is considered to have entered the effective steady-state region; The joint steady-state criterion for the sliding windows is defined as follows: in: , and These are the speed stability threshold, valve core opening stability threshold, and differential pressure stability threshold, respectively. Simultaneously extract valve opening or pump speed, pressure difference, temperature and flow data within the steady-state range to construct a target domain sample set.

[0015] Furthermore, in step four, the optimization objective function for transfer learning with parameter constraints is: in: Optimize the objective function for transfer learning with parameter constraints; These are the parameters of the soft measurement model, and , This is the weight matrix between the layers of the neural network. These are the bias vectors of neurons in each layer of the neural network; The model parameters are migrated and updated under the wear state of the target domain; For the target domain The actual traffic label corresponding to each sample; For the target domain The flow output predicted by the migration model for each sample; The total number of samples in the target domain; is the parameter constraint regularization coefficient, used to adjust the weight between the target domain prediction error and the source domain parameter preservation term.

[0016] Furthermore, in step five, the backstepping control strategy takes the flow tracking error as the control target, and decomposes the valve control or pump control system into multiple controllable subsystems by designing virtual control quantities in a hierarchical manner, and constructs control laws layer by layer based on Lyapunov stability theory; the target domain flow soft measurement model is embedded into the backstepping controller after inversion, which is used to directly convert the expected flow demand into the valve core target opening degree or pump speed control command.

[0017] Furthermore, the control law constructed by the backstepping control strategy is expressed as: in: The target pump output flow command calculated by the backstepping controller; and It is a nonlinear characteristic function that includes the system's volumetric elastic modulus, leakage coefficient, and current wear state; This refers to the velocity loop error; This refers to the flow loop error; Let be the system state variable vector, where For hydraulic cylinder displacement, For the speed of the hydraulic cylinder, These are state variables related to the pressure difference between the two chambers of the hydraulic cylinder or the pump output flow rate. The derivative of the virtual control variable in the third step; This is the gain for the flow loop error feedback.

[0018] This invention also proposes a system for implementing the online self-calibration method of hydraulic component flow characteristics based on multi-sensor information fusion as described above, comprising: The data acquisition module is used to collect data on the health and wear status of hydraulic components, including valve opening or pump speed, system differential pressure, oil temperature, and flow rate. The source domain modeling module is used to build a soft measurement model of source domain traffic based on backpropagation neural network based on health status data. The excitation trajectory generation module is used to construct the velocity sequence excitation trajectory and traverse the typical working conditions of hydraulic components under multiple openings, multiple speeds, and multiple pressure differences. The steady-state data filtering module is used to extract a high-quality target domain sample set from the excitation trajectory running data based on an adaptive sliding time window and continuous steady-state constraints. The transfer learning update module is used to fine-tune the source domain model using a transfer learning method with parameter constraints to obtain a target domain flow soft measurement model that adapts to the current wear state. The backstepping control module is used to invert and embed the updated target domain flow soft measurement model into the backstepping controller, replacing the failed physical model to achieve high-precision flow control.

[0019] Furthermore, the steady-state data filtering module includes: An adaptive window length calculation unit is used to dynamically determine the sliding window length based on the single-stroke travel length of the hydraulic cylinder and the target speed; The steady-state index calculation unit is used to calculate the coefficient of variation of speed, opening degree or rotational speed, and pressure difference within a window. The continuous steady-state determination unit is used to confirm the effective steady-state interval when the threshold condition is met in multiple consecutive windows.

[0020] The beneficial effects of this invention are as follows: This invention presents an online self-calibration method for the flow characteristics of hydraulic components based on multi-sensor information fusion. First, a source domain flow soft measurement model based on a backpropagation neural network (BPNN) is established using health state data as a priori knowledge foundation. Second, a velocity sequence excitation trajectory is designed to actively traverse typical working conditions after wear, acquiring raw data from the target domain. Third, high-quality target domain samples are automatically selected through an adaptive sliding window and steady-state constraint mechanism. Further, transfer learning with parameter constraints is used to fine-tune and update the source domain model, adapting to the current wear state while retaining the physical prior of the health state. Finally, the updated soft measurement model is inverted and embedded into a backstepping controller, replacing the degraded physical model to achieve high-precision closed-loop control. In summary, this invention focuses on data-driven modeling, active excitation sampling, steady-state screening, transfer learning updates, and control inversion fusion. Its aim is to achieve online self-sensing, model self-calibration, and adaptive control of flow characteristics under hydraulic component wear conditions without the need for flow sensors, thereby maintaining high-performance system control throughout its entire lifespan. It addresses the problem of flow characteristic drift caused by wear and degradation of hydraulic components during long-term service, leading to decreased system control accuracy. Simultaneously, it overcomes the shortcomings of traditional solutions, such as high cost, poor adaptability, and the inability of fixed-parameter models to reflect degradation characteristics, which rely on physical flow sensors. Attached Figure Description

[0021] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration: Figure 1 This is a flowchart of the online self-calibration method for the flow characteristics of hydraulic components based on multi-sensor information fusion according to the present invention; Figure 2 This is a schematic diagram of the online self-calibration method for the flow characteristics of hydraulic components based on multi-sensor information fusion in this embodiment. Figure 3 This is a schematic diagram of an example hydraulic system.

[0022] Explanation of reference numerals in the attached figures: 1-Hydraulic cylinder; 2-Displacement sensor; 3-Load; 4-Pressure sensor; 5-Relief valve; 6-Servo valve; 7-Flow meter; 8-Bidirectional metering pump; 9-Unidirectional metering pump; 10-Servo motor; 11-Oil tank; 12-Liquid thermometer; 13-Valve core displacement sensor; 14-Speed ​​sensor. Detailed Implementation

[0023] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0024] like Figure 1-2As shown in this embodiment, the online self-calibration method for the flow characteristics of hydraulic components based on multi-sensor information fusion includes the following steps: Step 1: Collect data on valve spool opening or pump speed, system differential pressure, oil temperature and flow rate under the healthy state of hydraulic components, construct a source domain dataset, and establish a source domain flow soft measurement model based on backpropagation neural network.

[0025] Specifically, the source-domain flow soft-sensing model uses valve spool opening or pump speed, system pressure difference, and oil temperature as input feature vectors, and output flow rate as supervision label. The backpropagation neural network includes an input layer, hidden layers, and an output layer. It constructs an error loss function between the predicted and measured flow rates and uses a gradient descent algorithm to iteratively optimize the network weights and bias parameters, thereby obtaining a source-domain soft-sensing model of the flow characteristics of hydraulic components under healthy conditions. Specifically, such as... Figure 3 As shown, the valve core opening is acquired in real time via valve core displacement sensor 13; the upstream pressure is measured via pressure sensor 4.5, and the downstream pressure is measured via pressure sensors 4.3 and 4.4 respectively, and the system pressure difference is calculated; the oil temperature is acquired via liquid thermometer 12; and the actual flow rate data is recorded by flowmeter 7 under the healthy operating condition of the servo valve, serving as the output label for model training. By adjusting the upstream pressure of the relief valve 5.3 and applying different load conditions in conjunction with the pump control system, the downstream back pressure is adjusted, thereby achieving controllable changes in the system pressure difference and acquiring data samples covering multiple operating conditions.

[0026] Because the physical quantities have different dimensions and significantly different values, in order to avoid neuron output saturation and accelerate network convergence, the input vector... and output tags Perform linear normalization: in: These are the normalized eigenvalues; and These represent the maximum and minimum values ​​of the corresponding dimension in the source dataset. After training, the predicted output needs to be inversely normalized using the corresponding inverse operation to restore it to the actual physical traffic.

[0027] Specifically, in this embodiment, the input vector of the source domain soft measurement model and with one-dimensional flow data As the output label. The optional source domain model training network is a backpropagation neural network (BPNN), which includes an input layer, hidden layers, and an output layer. In the backpropagation neural network, the nonlinear mapping from the input layer to the first hidden layer is: in: x is the j-th neuron in the first hidden layer;v Valve core opening; The pressure difference between the valve inlet and outlet is T; T is the temperature. The weights from the valve core opening to the j-th neuron in the first hidden layer; The value of the pressure difference between the valve inlet and outlet is the weight of the j-th neuron in the first hidden layer; The weights from temperature to the j-th neuron in the first hidden layer; The threshold value is the value of the j-th neuron from the input layer to the first hidden layer. This is the activation function corresponding to the input layer and the first hidden layer.

[0028] In a backpropagation neural network, the nonlinear mapping between two adjacent hidden layers is: in: For the current number The output value of the j-th neuron in the hidden layer; For the first The output of the i-th neuron in the hidden layer; For the first The hidden layer has a neuron i pointing to the first neuron. The connection weights of the j-th neuron in the hidden layer; For the first The bias of the j-th neuron in the hidden layer b; This is the nonlinear activation function between two adjacent hidden layers; This represents the number of neurons in the hidden layer.

[0029] The nonlinear mapping from the last hidden layer to the output layer is: in: For the predicted valve flow rate; The weights from the j-th neuron in the last hidden layer to the output neuron are: The threshold of the output neuron.

[0030] To minimize the predicted flow Compared with the measured traffic label Based on the deviation between the samples, a loss function is constructed with the sample mean squared error as the objective. Specifically, the constructed error loss function is as follows: in: Let it be the error loss function; The actual measured traffic volume is labeled; This is a flow soft measurement mapping function established based on a neural network; For the first The input feature vector of each sample includes valve core opening, pressure difference, and temperature; For the first The actual traffic label corresponding to each sample; This represents the total number of training samples.

[0031] The gradient descent algorithm is used to minimize the loss function. The weights and bias parameters of each network layer are iteratively updated along the direction of backpropagation of the error. The method is as follows: in, The learning rate determines the step size for parameter updates; The gradient of the error with respect to the weights reflects the direction of backpropagation of the error. This is the momentum factor, used to accelerate convergence and prevent the model from getting trapped in local optima.

[0032] Finally, by using the optimal weight and threshold matrix obtained through training, a source domain flow soft measurement benchmark model based on BPNN is established as a reference benchmark for subsequent transfer learning models under target domain operating conditions.

[0033] Step 2: Construct the excitation trajectory of the velocity sequence, and realize the traversal of hydraulic components under typical working conditions of multiple openings, multiple speeds, and multiple pressure differences through multi-speed reciprocating control of hydraulic cylinders, and obtain the original data of the target domain.

[0034] Specifically, in the velocity sequence excitation trajectory, the motion velocity of the hydraulic cylinder is first discretized into... The effective motion range of the hydraulic cylinder is determined by identifying target speeds and considering stroke boundary constraints and safety margin requirements. For each target speed, a smooth speed trajectory is constructed, including acceleration, steady-state constant speed, and deceleration phases, ensuring continuous, shock-free speed switching while retaining the steady-state data sampling range. Displacement closed-loop constraints are applied during the extension and retraction processes to ensure precise actuator reset after each complete reciprocating cycle. By traversing different combinations of target speeds, comprehensive coverage of typical operating conditions with multiple valve opening degrees (or multiple pump speeds) and multiple pressure differentials is achieved, forming the set of excitation trajectories required for the target domain sample.

[0035] In this embodiment, a velocity sequence excitation trajectory is constructed to achieve full traversal and dynamic excitation under typical operating conditions of multiple valve core openings (or multiple pump speeds) and multiple pressure differentials. Specifically, the method for constructing the velocity sequence excitation trajectory in this embodiment is as follows.

[0036] First, the speed of the hydraulic cylinder within its stable operating range is discretized into... The target velocities constitute a velocity set: in, The target uniform velocity is the i-th velocity; and These are the minimum and maximum constraint speeds at which the hydraulic cylinder can operate stably, respectively.

[0037] To ensure that the excitation trajectory meets the actuator motion boundary constraints, the physical stroke range of the hydraulic cylinder is considered. Safety margins reserved at both ends Determine the effective range of motion : Therefore, the effective travel length of the hydraulic cylinder in a single stroke can be obtained. for: To avoid sudden speed changes, hydraulic shocks, and flow spikes during multi-target speed switching, a smooth, shock-free transition strategy is adopted for each speed level. Specifically, for the first... Each speed level constructs a smooth velocity trajectory including acceleration, steady-state constant speed, and deceleration. : in: This represents the total one-way travel time. For acceleration and deceleration duration; For a single-journey local time variable, the smoothed velocity trajectory is shown. It consists of three parts: an acceleration section, a constant speed steady-state section, and a deceleration section. It can maintain a stable constant speed range while ensuring the continuity of speed and acceleration, which can be used for subsequent steady-state data acquisition.

[0038] The hydraulic cylinder extends using a forward velocity. The retraction process uses a symmetrical negative velocity. To ensure that the actuator strictly resets after each complete reciprocating cycle, this invention applies a position closed-loop constraint to the single-cycle excitation trajectory, that is, applies a displacement closed-loop constraint to the extension and retraction processes, ensuring that the net displacement in a single cycle is zero. .

[0039] Thus, the total displacement of the extension is strictly equal to and opposite in direction to the total displacement of the retraction, thereby ensuring that the net displacement in a single cycle is always zero.

[0040] Based on the above-mentioned target speed division, boundary constraints, smooth transition, and extension / retraction closed-loop constraints, the motion speed of the hydraulic cylinder within a complete reciprocating cycle can be uniformly represented as a global continuous piecewise function: By traversing different target speeds This can generate a set of active excitation trajectories covering typical operating conditions with multiple valve core openings (multiple pump speeds) and multiple pressure differentials. Importing this trajectory into the controller drives the servo valve-controlled hydraulic cylinder to reciprocate, allowing the system to traverse the main operating range of the servo valve at different speeds, providing a sufficient excitation basis for target domain sample acquisition.

[0041] Step 3: Construct an adaptive sliding time window based on the hydraulic cylinder operating cycle, and combine standard deviation, coefficient of variation and continuous steady-state constraints to perform steady-state identification, sample screening and data reconstruction to construct the target domain sample set.

[0042] Specifically, this embodiment proposes a sliding time window mechanism based on the adaptive single-stroke travel length of the hydraulic cylinder and the target speed. By dynamically adjusting the window length according to the target speed, a consistent relative judgment scale is maintained under different speed conditions. Within each sliding window, dimensionless coefficients of variation of the hydraulic cylinder speed, valve core opening (or pump speed), and system pressure difference are constructed as steady-state indicators, and steady-state judgment is performed in combination with preset thresholds. At the same time, a continuous steady-state constraint mechanism is introduced. The system is considered to have entered the effective steady-state interval only when multiple consecutive windows meet the steady-state conditions, thereby effectively eliminating acceleration and deceleration phases, transient disturbances, and noise misjudgments. On this basis, multi-source data such as valve opening (or pump speed), pressure difference, temperature, and flow rate within the steady-state interval are extracted synchronously to form a high-quality target domain sample set covering typical working conditions.

[0043] Specifically, in this embodiment, the method for constructing the target domain sample set is as follows.

[0044] Effective travel length of a single stroke of the hydraulic cylinder To unify spatial constraints and combine them with the current target velocity The window length is dynamically determined so that the sliding window always covers a fixed proportion of the effective motion range at the current speed, thus ensuring consistent steady-state judgment criteria under different speed conditions. The total number of sampling points for a single stroke is: in: This represents the total number of sampling points for a single trip. This is the sampling step size.

[0045] To ensure that the window size has a consistent relative coverage ratio across different speed levels, then the... The adaptive sliding window length for each speed level is: in: The proportion of the window to the length of a single-pass sample; The length of the sliding window corresponding to the i-th velocity; Minimum window length; Maximum window length.

[0046] Within each sliding window, the standard deviation and dimensionless coefficient of variation of the hydraulic cylinder speed, valve core opening, and pressure difference are respectively constructed as steady-state criteria. The average speed of each window is: The standard deviation of the velocity is: The speed stability index is defined as follows: Similarly, the standard deviation of valve core opening Valve core opening variation coefficient With pressure difference standard deviation Coefficient of variation of pressure difference The stability index is defined as: in: This is an indicator of the speed stability of the hydraulic cylinder; The standard deviation of the velocity; For the first Average speed of each window; This is an indicator of valve core opening stability. The standard deviation of valve core opening; For the first The average valve core opening of each window; This is an indicator of system differential pressure stability. This represents the standard deviation of the system pressure difference. For the first Average system pressure difference across windows.

[0047] In this embodiment, a continuous steady-state constraint mechanism is introduced, which only applies when the continuous steady-state condition is met. When all windows meet the steady-state conditions, the system is considered to have entered the effective steady-state region. Specifically, the first... The joint steady-state criterion for the sliding windows is defined as follows: in: , and These are the speed stability threshold, valve core opening stability threshold, and differential pressure stability threshold, respectively. Simultaneously extract valve opening or pump speed, pressure difference, temperature, and flow rate data within the steady-state interval to construct a target domain sample set. Specifically, since a single sliding window can only reflect local stability, a continuous steady-state constraint mechanism is further introduced. The continuous steady-state length threshold is defined as... If and only if in succession The system is considered to have entered a valid steady-state operating state only when all steady-state criteria are met within each sliding window. When the continuity constraint is met, the first... The steady-state intervals are: After determining the steady-state interval, a time backtracking strategy is further employed to extract multi-source data within the steady-state interval to construct a high-quality training sample set for the target domain. Specifically, within the steady-state interval... Internally, characteristic variables such as valve core opening, system differential pressure, and temperature are extracted and used to construct supervised learning samples with the corresponding flow measurement values. in: The traffic volume label is calculated based on the measured speed; This represents the effective working area of ​​the hydraulic cylinder at the current speed.

[0048] Thus, this embodiment can automatically eliminate acceleration / deceleration transition sections and unsteady disturbance sections, accurately extract high-quality target domain samples covering multiple opening degrees and multiple pressure difference conditions, and provide a reliable data foundation for subsequent flow soft measurement model transfer learning and updates.

[0049] Step 4: Use a transfer learning method with parameter constraints to fine-tune and update the source domain flow soft measurement model, reconstruct the nonlinear mapping relationship between valve core opening or pump speed, pressure difference, temperature and output flow under the wear state of hydraulic components, and obtain the target domain flow soft measurement model.

[0050] Specifically, the soft measurement model update uses the parameters of the healthy state flow soft measurement model as initial parameters, introduces a transfer learning fine-tuning strategy with parameter constraints, and achieves stable updating of model parameters by minimizing the joint loss function composed of the target domain prediction error and the source domain parameter offset regularization term. This allows the model to gradually converge while retaining the knowledge structure of the source domain, and finally obtains a target domain flow soft measurement model that adapts to the current degradation state of hydraulic components.

[0051] In this embodiment, the target domain dataset is constructed using the multi-condition operating samples collected in step three under the current pump source degradation state: in, The input feature vector consists of valve core opening, system differential pressure, and oil temperature. This corresponds to the flow rate output at that moment.

[0052] The parameters of the source domain flow soft measurement model established in step one are as follows: Based on this, a transfer learning strategy with parameter constraints is used to fine-tune the model, and its optimization objective function is defined as: in: Optimize the objective function for transfer learning with parameter constraints; These are the parameters of the soft measurement model, and , This is the weight matrix between the layers of the neural network. These are the bias vectors of neurons in each layer of the neural network; The model parameters are migrated and updated under the wear state of the target domain; For the target domain The actual traffic label corresponding to each sample; For the target domain The flow output predicted by the migration model for each sample; The total number of samples in the target domain; is the parameter constraint regularization coefficient, used to adjust the weight between the target domain prediction error and the source domain parameter preservation term.

[0053] In the optimization objective function, the first term is used to minimize the prediction error in the target domain, and the second term is a parameter regularization constraint term, which is used to limit the degree of deviation of the target model parameters relative to the source domain model.

[0054] For the target domain samples, the model output is: The gradient is calculated using the backpropagation algorithm, and the parameters are iteratively updated based on the target domain loss function: in: The transfer learning rate is typically smaller than the source domain training learning rate to ensure that the parameters converge stably in the vicinity of the source domain knowledge. When the validation error meets the convergence condition or the maximum number of iterations is reached, the updated target domain model is output. Through the above process, the healthy state model can be rapidly migrated and adaptively updated to the deterioration condition, thereby significantly improving the accuracy of flow prediction under wear conditions.

[0055] Step 5: Construct a backstepping control strategy that integrates the target domain flow soft measurement model. Embed the updated target domain flow soft measurement model into the backstepping controller to replace the hydraulic component physical model that is inaccurate due to wear and degradation, thereby realizing online self-calibration and control of flow characteristics under hydraulic component wear conditions.

[0056] Specifically, the backstepping control strategy takes flow tracking error as the control target. By designing virtual control quantities in a hierarchical manner, the valve control (or pump control) system is decomposed into multiple controllable subsystems. Based on Lyapunov stability theory, control laws are constructed layer by layer to make the system error converge asymptotically layer by layer. The target domain flow soft measurement model is inverted and embedded into the backstepping controller to replace the wear-degraded and inaccurate hydraulic component flow model, thereby realizing the self-calibration and high-precision closed-loop control of the nonlinear flow characteristics of the hydraulic component.

[0057] Specifically, the updated target domain flow soft measurement model from step four is used. This model replaces the hydraulic component model that fails due to wear in the control logic. It accurately captures the complex nonlinear relationship between valve spool opening, differential pressure, oil temperature, and actual output flow under wear conditions. This model reconstructs hydraulic components that experience wear and performance degradation into a complete and accurately predictable soft measurement model of healthy hydraulic components.

[0058] First, the dynamic equations of the valve-controlled system are transformed into a strict feedback form. Based on the physical model of the valve-controlled system, the state-space equations of the system can be described as follows: in: Represents actuator displacement; Represents actuator speed; Represents a stress-related state; and It is a nonlinear characteristic function that includes the system's volumetric elastic modulus, leakage coefficient, and current wear state.

[0059] To achieve precise motion control of the reconfigurable pump source system, a backstepping method is used to construct the control law. The position loop error is defined. Speed ​​loop error Flow loop error Construct the Lyapunov function for the entire system. To ensure that the system error converges exponentially, a dynamic convergence law for the error is established. At the flow control level, combining the error feedback from the first two levels, the following rigorous control law is designed: in: The target pump output flow command calculated by the backstepping controller; and It is a nonlinear characteristic function that includes the system's volumetric elastic modulus, leakage coefficient, and current wear state; This is a vector of system state variables; The desired displacement signal; The desired virtual speed control quantity is constructed from the displacement tracking error; The virtual control quantity for the desired flow rate or equivalent pressure is constructed from the speed tracking error; for The first derivative; A positive error feedback gain is used to adjust the convergence speed of the flow loop or equivalent pressure loop error.

[0060] This embodiment does not employ traditional physical formula compensation; instead, it directly embeds the soft-sensor model into the end-effector layer of the backstepping controller. Through inversion mapping technology, the desired flow rate demand is directly converted into the actual pump speed control input. The target valve core opening command x obtained by mapping v Send to the controller to drive the servo valve output and The actual traffic volume that matches the actual traffic volume.

[0061] This embodiment also proposes an online self-calibration system for the flow characteristics of hydraulic components based on multi-sensor information fusion, used to implement the online self-calibration method for the flow characteristics of hydraulic components based on multi-sensor information fusion described in this embodiment. Specifically, as follows... Figure 3 As shown, this embodiment of the online self-calibration system for hydraulic component flow characteristics based on multi-sensor information fusion includes a data acquisition module, a source domain modeling module, an excitation trajectory generation module, a steady-state data filtering module, a transfer learning update module, and a backstepping control module. Specifically, the data acquisition module is used to collect data on valve core opening or pump speed, system differential pressure, oil temperature, and flow rate under the health and wear states of hydraulic components; the source domain modeling module is used to establish a source domain flow soft measurement model based on a backpropagation neural network based on the health state data; the excitation trajectory generation module is used to construct a velocity sequence excitation trajectory, traversing typical operating conditions of hydraulic components under multiple openings, multiple speeds, and multiple differential pressures; the steady-state data filtering module is used to extract a high-quality target domain sample set from the excitation trajectory running data based on an adaptive sliding time window and continuous steady-state constraints; the transfer learning update module is used to fine-tune the source domain model using a transfer learning method with parameter constraints to obtain a target domain flow soft measurement model adapted to the current wear state; the backstepping control module is used to embed the updated target domain flow soft measurement model into the backstepping controller, replacing the failed physical model to achieve high-precision flow control.

[0062] In this embodiment, the steady-state data filtering module includes an adaptive window length calculation unit, a steady-state index calculation unit, and a continuous steady-state determination unit. Specifically, the adaptive window length calculation unit is used to dynamically determine the sliding window length based on the single-stroke travel length of the hydraulic cylinder and the target speed; the steady-state index calculation unit is used to calculate the coefficient of variation of speed, opening degree or rotational speed, and pressure difference within the window; and the continuous steady-state determination unit is used to confirm the effective steady-state interval when multiple consecutive windows meet the threshold conditions.

[0063] This embodiment presents a method and system for online self-calibration of hydraulic component flow characteristics based on multi-sensor information fusion. It verifies the feasibility of online self-calibration and high-precision control of hydraulic component flow characteristics under wear conditions, even without flow sensors. By constructing a source domain benchmark model, designing velocity sequence excitation trajectories, employing adaptive sliding window steady-state screening, and parameter-constrained transfer learning, accurate modeling and online updating of flow characteristics are achieved. Embedding the updated flow soft measurement model into the backstepping controller allows for effective replacement of worn and degraded hydraulic components, significantly improving system position control accuracy and dynamic response performance. This method and system can be extended to various types of electro-hydraulic servo systems and hydraulic systems for construction machinery. It enables health status monitoring and control of hydraulic components without the need for additional flow sensors, demonstrating significant engineering application and promotional value.

[0064] Compared with the prior art, the present invention has the following beneficial effects: (1) By using speed excitation and adaptive sliding window mechanism, the traversal of hydraulic components with multiple opening degrees and multiple pressure differential conditions and the automatic extraction of steady-state data are realized, effectively eliminating non-steady-state interference data and improving the sample quality of the target domain; (2) A transfer learning method with parameter constraints is adopted to realize the rapid fine-tuning and updating of the flow soft measurement model under wear conditions, and to realize flow soft measurement prediction without relying on physical flow sensors. (3) The updated soft measurement model is embedded into the backstep controller to realize the online replacement of the degraded physical model and improve the flow control accuracy of hydraulic components under wear conditions.

[0065] The above-described embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A method for online self-calibration of the flow characteristics of hydraulic components based on multi-sensor information fusion, characterized in that: Includes the following steps: Step 1: Collect data on valve spool opening or pump speed, system differential pressure, oil temperature and flow rate under the healthy state of hydraulic components, construct a source domain dataset, and establish a source domain flow soft measurement model based on backpropagation neural network. Step 2: Construct the excitation trajectory of the velocity sequence, and realize the traversal of hydraulic components under typical working conditions of multiple openings, multiple speeds, and multiple pressure differences through multi-speed reciprocating control of hydraulic cylinders to obtain the original data of the target domain; Step 3: Construct an adaptive sliding time window based on the hydraulic cylinder operating cycle, and combine standard deviation, coefficient of variation and continuous steady-state constraints to perform steady-state identification, sample screening and data reconstruction to construct the target domain sample set; Step 4: The source domain flow soft measurement model is fine-tuned and updated using a transfer learning method with parameter constraints. The nonlinear mapping relationship between valve core opening or pump speed, pressure difference, temperature and output flow under the wear state of hydraulic components is reconstructed to obtain the target domain flow soft measurement model. Step 5: Construct a backstepping control strategy that integrates the target domain flow soft measurement model. Embed the updated target domain flow soft measurement model into the backstepping controller to replace the hydraulic component physical model that is inaccurate due to wear and degradation, thereby realizing online self-calibration and control of flow characteristics under hydraulic component wear conditions.

2. The online self-calibration method for the flow characteristics of hydraulic components based on multi-sensor information fusion according to claim 1, characterized in that: In step one, the source domain flow soft measurement model uses valve core opening or pump speed, system pressure difference and oil temperature as input feature vectors and output flow as supervision label; the backpropagation neural network includes an input layer, a hidden layer and an output layer, constructs an error loss function between the predicted flow and the measured flow, and uses a gradient descent algorithm to iteratively optimize the network weights and bias parameters.

3. The online self-calibration method for the flow characteristics of hydraulic components based on multi-sensor information fusion according to claim 2, characterized in that: In the backpropagation neural network, the nonlinear mapping from the input layer to the first hidden layer is as follows: in: x is the j-th neuron in the first hidden layer; v Valve core opening; The pressure difference between the valve inlet and outlet is T; T is the temperature. The weights from the valve core opening to the j-th neuron in the first hidden layer; The value of the pressure difference between the valve inlet and outlet is the weight of the j-th neuron in the first hidden layer; The weights from temperature to the j-th neuron in the first hidden layer; The threshold value is the value of the j-th neuron from the input layer to the first hidden layer. The activation function corresponding to the input layer to the first hidden layer; The nonlinear mapping between two adjacent hidden layers is: in: For the current number The output value of the j-th neuron in the hidden layer; For the first The output of the i-th neuron in the hidden layer; For the first The hidden layer has a neuron i pointing to the first neuron. The connection weights of the j-th neuron in the hidden layer; For the first The bias of the j-th neuron in the hidden layer b; This is the nonlinear activation function between two adjacent hidden layers; This represents the number of neurons in the hidden layer. The nonlinear mapping from the last hidden layer to the output layer is: in: For the predicted valve flow rate; The weights from the j-th neuron in the last hidden layer to the output neuron are: The threshold of the output neuron; The constructed error loss function is as follows: in: Let it be the error loss function; The actual measured traffic volume is labeled; This is a flow soft measurement mapping function established based on a neural network; For the first The input feature vector of each sample includes valve core opening, pressure difference, and oil temperature; For the first The actual traffic label corresponding to each sample; The total number of training samples; The gradient descent algorithm is used to iteratively optimize the network weights and bias parameters. The method is as follows: in, The learning rate determines the step size for parameter updates; The gradient of the error with respect to the weights reflects the direction of backpropagation of the error. This is the momentum factor, used to accelerate convergence and prevent the model from getting trapped in local optima.

4. The online self-calibration method for the flow characteristics of hydraulic components based on multi-sensor information fusion according to claim 1, characterized in that: In step two, the method for constructing the velocity sequence excitation trajectory is as follows: Discretize the hydraulic cylinder's movement speed as The target velocities constitute a velocity set: in, The target uniform velocity is the i-th velocity; and These are the minimum and maximum constraint speeds at which the hydraulic cylinder can operate stably; Combining the physical stroke range of the hydraulic cylinder With safety margin Determine the effective range of motion and effective walking length per one trip : Regarding the first Each speed level constructs a smooth velocity trajectory including acceleration, steady-state constant speed, and deceleration. : in: This represents the total one-way travel time. For acceleration and deceleration duration; For a single-journey local time variable; Apply a closed-loop displacement constraint to the extension and retraction processes to ensure that the net displacement in a single cycle is zero. 。 5. The online self-calibration method for the flow characteristics of hydraulic components based on multi-sensor information fusion according to claim 1, characterized in that: In step three, the method for constructing the target domain sample set is as follows: Based on the effective travel length of a single stroke of the hydraulic cylinder relative to target speed and sampling step size Calculate the total number of sampling points for a single trip. And adaptive sliding window length: in: This represents the total number of sampling points for a single trip. The proportion of the window to the length of a single-pass sample; The length of the sliding window corresponding to the i-th velocity; Minimum window length; Maximum window length; In the Within a sliding window, the coefficients of variation of hydraulic cylinder speed, valve core opening, and system pressure difference are constructed as steady-state indicators: in: This is an indicator of the speed stability of the hydraulic cylinder; The standard deviation of the velocity; For the first Average speed of each window; This is an indicator of valve core opening stability. The standard deviation of valve core opening; For the first The average valve core opening of each window; This is an indicator of system differential pressure stability. This represents the standard deviation of the system pressure difference. For the first Average system differential pressure across windows; Introducing a continuous steady-state constraint mechanism, only when continuous When all windows meet the steady-state conditions, the system is considered to have entered the effective steady-state region; The joint steady-state criterion for the sliding windows is defined as follows: in: , and These are the speed stability threshold, valve core opening stability threshold, and differential pressure stability threshold, respectively. Simultaneously extract valve opening or pump speed, pressure difference, temperature and flow data within the steady-state range to construct a target domain sample set.

6. The online self-calibration method for the flow characteristics of hydraulic components based on multi-sensor information fusion according to claim 1, characterized in that: In step four, the optimization objective function for transfer learning with parameter constraints is: in: Optimize the objective function for transfer learning with parameter constraints; These are the parameters of the soft measurement model, and , This is the weight matrix between the layers of the neural network. These are the bias vectors of neurons in each layer of the neural network; The model parameters are migrated and updated under the wear state of the target domain; For the target domain The actual traffic label corresponding to each sample; For the target domain The flow output predicted by the migration model for each sample; The total number of samples in the target domain; is the parameter constraint regularization coefficient, used to adjust the weight between the target domain prediction error and the source domain parameter preservation term.

7. The online self-calibration method for the flow characteristics of hydraulic components based on multi-sensor information fusion according to claim 1, characterized in that: In step five, the backstepping control strategy takes the flow tracking error as the control target, decomposes the valve control or pump control system into multiple controllable subsystems by designing virtual control variables in a hierarchical manner, and constructs control laws layer by layer based on Lyapunov stability theory. The target domain flow soft measurement model is inverted and then embedded into a backstepping controller to directly convert the desired flow demand into a valve core target opening or pump speed control command.

8. The online self-calibration method for the flow characteristics of hydraulic components based on multi-sensor information fusion according to claim 7, characterized in that: The control law constructed by the backstepping control strategy is expressed as: in: The target pump output flow command calculated by the backstepping controller; and It is a nonlinear characteristic function that includes the system's volumetric elastic modulus, leakage coefficient, and current wear state; This refers to the velocity loop error; This refers to the flow loop error; Let be the system state variable vector, where For hydraulic cylinder displacement, For the speed of the hydraulic cylinder, These are state variables related to the pressure difference between the two chambers of the hydraulic cylinder or the pump output flow rate. The derivative of the virtual control variable in the third step; This is the gain for the flow loop error feedback.

9. A system for implementing the online self-calibration method for the flow characteristics of hydraulic components based on multi-sensor information fusion as described in any one of claims 1-8, characterized in that: include: The data acquisition module is used to collect data on the health and wear status of hydraulic components, including valve opening or pump speed, system differential pressure, oil temperature, and flow rate. The source domain modeling module is used to build a soft measurement model of source domain traffic based on backpropagation neural network based on health status data. The excitation trajectory generation module is used to construct the velocity sequence excitation trajectory and traverse the typical working conditions of hydraulic components under multiple openings, multiple speeds, and multiple pressure differences. The steady-state data filtering module is used to extract a high-quality target domain sample set from the excitation trajectory running data based on an adaptive sliding time window and continuous steady-state constraints. The transfer learning update module is used to fine-tune the source domain model using a transfer learning method with parameter constraints to obtain a target domain flow soft measurement model that adapts to the current wear state. The backstepping control module is used to invert and embed the updated target domain flow soft measurement model into the backstepping controller, replacing the failed physical model to achieve high-precision flow control.

10. The online self-calibration system for the flow characteristics of hydraulic components based on multi-sensor information fusion according to claim 10, characterized in that: The steady-state data filtering module includes: An adaptive window length calculation unit is used to dynamically determine the sliding window length based on the single-stroke travel length of the hydraulic cylinder and the target speed; The steady-state index calculation unit is used to calculate the coefficient of variation of speed, opening degree or rotational speed, and pressure difference within a window. The continuous steady-state determination unit is used to confirm the effective steady-state interval when the threshold condition is met in multiple consecutive windows.

Images