Hydraulic system employing AI-based virtual pressure sensor

Abstract

The hydraulic system comprises a hydraulic pump and a plurality of actuators, hydraulic oil is supplied to each actuator through a corresponding hydraulic loop by the hydraulic pump, and a corresponding control valve is arranged in each hydraulic loop; the control unit comprises a virtual pressure sensor, the virtual pressure sensor comprises an AI model using a time sequence, input signals of the AI model comprise the pump rotating speed, the pump opening pressure, the pump displacement, control signals of the m control valves and data of the hydraulic oil temperature at sequence moments, and the sequence moments comprise the current moment and at least m previous moments; an output signal of the AI model comprises load end pressure prediction values of the m control valves.

Description

Technical Field

[0001] This application relates to a hydraulic system, particularly a hydraulic system for construction machinery, which employs an AI (artificial intelligence) based virtual pressure sensor. Background Technology

[0002] Engineering equipment employing hydraulic systems, such as excavators, aerial work platforms, cranes, and drilling rigs, typically utilizes a single hydraulic pump to supply hydraulic oil to multiple actuators via a multi-channel hydraulic system to achieve simultaneous multiple actions. In hydraulic systems, the control of each flow control valve relies on real-time readings from pressure sensors on the corresponding actuator side, or on real-time monitoring of system pressure using these pressure sensor readings. If one or more of these pressure sensors could be eliminated, and the same function could be achieved through virtual pressure sensors in software, it would help reduce the overall design and manufacturing costs of the hydraulic system. When using a LUDV (Load Independent Flow Distribution) system in a hydraulic system, the hardware is complex and the design and manufacturing costs are high. If the functionality of a LUDV system can be achieved using virtual pressure sensors combined with simple electro-hydraulic valves, the overall hardware cost of the system can be reduced.

[0003] When designing a virtual pressure sensor, a physical model can be established, and the pressure value to be detected can be calculated based on the input signal using the physical model. However, achieving high pressure accuracy using a virtual pressure sensor based on a physical model is not easy to implement in engineering, especially for hydraulic systems of engineering equipment with large pressure differential changes and fluctuations. Summary of the Invention

[0004] The purpose of this application is to use an AI-based virtual pressure sensor in a hydraulic system, which can provide high pressure accuracy and is relatively easy to implement in engineering.

[0005] To this end, this application provides a hydraulic system and its virtual pressure sensor in one aspect. The hydraulic system includes: a hydraulic pump and a plurality of actuators, each actuator being supplied with hydraulic oil by the hydraulic pump via a corresponding hydraulic circuit, each hydraulic circuit having a corresponding control valve; and a control unit comprising the virtual pressure sensor. The virtual pressure sensor includes an AI model using a time series, the input signals of which include pump speed, pump inlet pressure, pump displacement, control signals of m control valves (m being between 1 and the total number of control valves, preferably between 2 and the total number of control valves), and hydraulic oil temperature data at sequential times, the sequential times including the current time and at least the previous m times; the output signal of the AI ​​model includes predicted load-side pressure values ​​of the m control valves.

[0006] According to this application, a virtual pressure sensor is established based on an AI model that processes time series data. This virtual sensor provides high pressure accuracy and can be reliably applied to hydraulic systems, especially those in engineering equipment. It can replace traditional physical sensors, reducing the number of physical sensors used, lowering system costs, and decreasing the likelihood of system failures caused by physical sensors. Alternatively, it can serve as a backup for physical sensors, providing redundancy in measurement values ​​and allowing for correction or verification of physical sensor readings. Furthermore, the virtual pressure sensor based on the AI ​​model for processing time series data presented in this application provides relatively high-precision pressure values ​​in a manner easily implemented in engineering. Attached Figure Description

[0007] The foregoing and other aspects of this application will be more fully understood and appreciated through the following detailed description with reference to the accompanying drawings, in which:

[0008] Figure 1 This is a schematic diagram of a hydraulic system for engineering machinery according to this application;

[0009] Figure 2 This is a block diagram of the input and output signals of the AI ​​model of the control unit in the hydraulic system.

[0010] Figure 3 This is a block diagram of an exemplary AI model used in this application;

[0011] Figure 4 This is a block diagram of another exemplary AI model used in this application;

[0012] Figures 5-7 This is a graph showing the performance of the AI ​​model used in this application. Detailed Implementation

[0013] This application generally relates to the introduction of AI (artificial intelligence) based virtual pressure sensors into hydraulic systems, especially hydraulic systems of construction machinery.

[0014] like Figure 1 As shown, the hydraulic system involved in this application includes a hydraulic pump (main pump) 2 driven by a power source 1 (such as an engine or motor). The hydraulic pump 2 is a variable displacement pump. The output end of the hydraulic pump 2 is connected to a high-pressure oil circuit for supplying hydraulic oil to multiple actuators C1, C2... connected in parallel via the high-pressure oil circuit and corresponding hydraulic circuits. Figure 1 The image shows two actuators as examples, but it is understood that a hydraulic system may include more actuators supplied with hydraulic oil by a common hydraulic pump 2. The actuators may be hydraulic cylinders, hydraulic motors, or other forms of hydraulic actuators.

[0015] Each actuator C1, C2... is equipped with a corresponding control valve (usually an electro-proportional control valve) V1, V2... in its respective hydraulic circuit, which is used to control the hydraulic pump 2 to supply hydraulic oil to the corresponding actuator.

[0016] Figure 1 The hydraulic system in the diagram can be configured as a LUDV (Load Independent Flow Distribution) system. To construct a LUDV system, appropriate hydraulic components and circuits need to be added to the hydraulic system, which is well known to those skilled in the art and will not be described further here.

[0017] exist Figure 1 In the example shown, the actuator is a bidirectional actuator (such as a hydraulic motor, hydraulic cylinder, etc.). Therefore, each actuator is equipped with two ports, each connected to one of the two working ports of the corresponding control valve. Each control valve also has an inlet and a drain port, with the inlet connected to the output of hydraulic pump 2 and the drain port connected to the oil tank. Each control valve has three positions: a neutral position, and first and second working positions. In the neutral position, the control valve's inlet is not connected to either working port. In the first working position, the control valve's inlet is connected to the first working port, and the drain port is connected to the second working port, thereby driving the corresponding actuator to operate in a first direction. In the second working position, the control valve's inlet is connected to the second working port, and the drain port is connected to the first working port, thereby driving the corresponding actuator to operate in a second direction.

[0018] It is understandable that, based on the functional settings of the actuator, the corresponding control valve can have different valve positions and port settings. For example, for a unidirectional actuator, the control valve may only have one working valve position.

[0019] Back Figure 1 The hydraulic system also includes a control unit 3, which controls at least the operation of the power source 1, the hydraulic pump 2, and each control valve. Furthermore, to control these operations, the control unit 3 is configured to receive various detection signals from the hydraulic system. The control unit 3 needs to acquire the load-side pressure of each control valve for hydraulic system control (e.g., controlling the displacement of the hydraulic pump 2, the drive current of each control valve, etc.) or for hydraulic system status monitoring. The load-side pressure of a control valve is the pressure at the working port connected to the inlet in a certain working valve position (one of the first or second working valve positions).

[0020] In existing technologies, each control valve is typically equipped with a physical pressure sensor at or near its working port to detect the load-side pressure of the control valve in real time. In the hydraulic system of this application, virtual pressure sensors for one or more (or even all) control valves are implemented through software, while the original physical pressure sensors of the control valves can be eliminated or retained.

[0021] like Figure 2 As shown, the control unit 3 is equipped with an AI-based virtual pressure sensor model (hereinafter referred to as AI model) M. This AI model M is configured to estimate the load-side pressure of the control valve of interest (one or more, or all, of the aforementioned control valves in the hydraulic system) based on the input signal (input data).

[0022] The AI ​​model receives input signals that are correlated with the load-side pressure of the control valve. Since the control unit 3 can acquire a lot of data from the hydraulic system (setting data, operating data, etc.), it is necessary to identify those signals that are correlated with the load-side pressure of the control valve in order to make the AI ​​model as accurate and simple as possible.

[0023] In order to determine the signal that is correlated with the load-side pressure of the control valve, this application first constructs a physical model for the virtual pressure sensor.

[0024] First refer to Figure 1 The hydraulic system shown is assumed to have two control valves V1 and V2 equipped with virtual pressure sensors. During a certain action of the construction machinery, the combined action of actuators C1 and C2 is involved. For this purpose, control unit 3 controls the control valves V1 and V2 to be in a desired working position (one of the first and second working positions), and each has a corresponding opening area A (the flow area between the oil inlet and the connected first or second working oil port).

[0025] Using conventional flow formulas in this field, the flow rate Q1 from pump 2 through control valve V1 to actuator C1, and the flow rate Q2 from control valve V2 to actuator C2 are expressed as follows:

[0026]

[0027] The rate of pressure change from the output end (referred to as the pump port) of pump 2 to the main pipeline of control valves V1 and V2, that is, the first derivative of the pump port pressure with respect to time. Represented as:

[0028]

[0029] Substitute Q1 and Q2 into The formula yields the following result:

[0030]

[0031] The symbols appearing in the above formulas have the following meanings:

[0032] α1, α2: Flow coefficients of control valves V1 and V2, representing the capacity of hydraulic oil to flow through the control valves;

[0033] ρ: density of hydraulic oil;

[0034] K: Bulk elastic modulus of hydraulic oil;

[0035] V: Total volume of the pipeline cavity from the pump inlet to control valves V1 and V2;

[0036] p p Pump inlet pressure;

[0037] The first derivative of the pump inlet pressure with respect to time;

[0038] n p Pump speed;

[0039] V g Pump displacement;

[0040] Q p Pump output flow rate is equal to pump speed multiplied by pump displacement.

[0041] A1: The opening area of ​​control valve V1 from the oil inlet to the working oil port connected to the oil inlet;

[0042] A2: The opening area of ​​control valve V2 from the oil inlet to the working oil port connected to the oil inlet;

[0043] p1: The pressure of the working port of control valve V1 connected to the oil inlet, that is, the load end pressure of control valve V1;

[0044] p2: The pressure of the working port of control valve V2 connected to the oil inlet, that is, the load end pressure of control valve V2;

[0045] Q le The total leakage in the pipeline from the pump inlet to control valves V1 and V2, relative to the pump inlet pressure p. p It is related to the hydraulic oil temperature T, and can be expressed as Q. le =f(p p The fixed form of ,T), when p is known p Q can be calculated directly in the case of T. le ;

[0046] Q ot The relief valve flow rate refers to the flow rate of the relief valve at the output end of hydraulic pump 2 (or the connected high-pressure oil circuit) when it is open. This flow rate can be determined by the characteristic curve of the relief valve and the pump port pressure p. p get.

[0047] Assign superscripts to the parameters in equation (2) that change with the control cycle, with k representing the current control cycle (or the current time) and k-1 representing the previous control cycle (the previous time), then equation (2) can be rewritten as the first derivative of the pump inlet pressure under the current cycle. Represented as:

[0048]

[0049] Given a fixed hydraulic system, α1, α2, ρ, K, and V can be constants in equation (3). Alternatively, ρ and K can account for the effect of hydraulic oil temperature.

[0050] Pump inlet pressure p p The pump inlet pressure of the current control cycle can be measured by the pump inlet pressure sensor. This information can be collected by the control unit during the current control cycle.

[0051] It can be obtained in various ways, such as by calculating it using the following linear method:

[0052]

[0053] Where Δt is the time of each control cycle of the control unit. It is the pump inlet pressure collected by the control unit in the previous control cycle.

[0054] Data collected by the control unit during the current control cycle.

[0055] The signal is acquired by the control unit during the current control cycle, for example, via a pump oscillation sensor or a control signal from the EP pump. get.

[0056] The control signal from the previous control cycle of control valve V1 Confirmed, therefore It can be obtained by the control unit in the current control cycle, that is, it can be represented as A fixed form.

[0057] The control signal from the previous control cycle of control valve V2 Confirmed, therefore It can be obtained by the control unit in the current control cycle, that is, it can be represented as A fixed form.

[0058] To represent as The fixed form in which T can be collected by the control unit, so This can be considered as something that can be obtained by the control unit in the current control cycle.

[0059] The flow rates of other valves can be calculated using flow rate formulas or measured by flow sensors. The flow rate of the relief valve can be obtained from the relief valve's characteristic curve and the pump inlet pressure. This can be considered as something that can be obtained by the control unit in the current control cycle.

[0060] The pressure at the working port of control valve V1 needs to be obtained through a virtual pressure sensor in the current control cycle. The pressure at the working port of control valve V2 needs to be obtained through a virtual pressure sensor for the current control cycle.

[0061] Taking the derivative of equation (3) with respect to time yields the second derivative of the pump inlet pressure with respect to time in the current control loop.

[0062]

[0063] in:

[0064]

[0065] They can be represented in the following forms:

[0066]

[0067] The pressure at the working port of control valve V1, obtained through the virtual pressure sensor in the previous control cycle (i.e., time k-1), is a known quantity.

[0068] The pressure at the working port of control valve V2, obtained through the virtual pressure sensor in the previous control cycle (i.e., time k-1), is a known quantity.

[0069] The pressure collected by the pump pressure sensor in the previous control cycle (i.e., time k-2) is a known quantity.

[0070] There are two unknowns in equations (3) and (4). and Combining equations (3) and (4), we obtain a result containing two unknowns. and The physical model of the virtual pressure sensor, composed of two equations, can be solved to obtain two unknowns. and Thus, using known values and the estimated value calculated from the previous control loop. The load-side pressures of control valves V1 and V2 in the current working cycle can be calculated. and

[0071] For any working port of control valves V1 and V2 connected to their inlet in any working valve position, the above equations (3) and (4) are solved simultaneously to obtain the two unknowns. and The algorithm is universal. Based on the valve positions of control valves V1 and V2 during a certain operation in the hydraulic system, it is possible to determine which working port of control valves V1 and V2 (and correspondingly which port of actuators C1 and C2) is the load end. The load end pressure at each moment, obtained through the virtual pressure sensor model described above, is used as the measurement value of the virtual pressure sensor.

[0072] To summarize further, assuming a hydraulic system contains n0 actuators (and corresponding n0 control valves), for a situation where n (n0≥n) actuators (and corresponding n control valves) need to perform a certain action in a hydraulic system, the control unit controls the corresponding n control valves to be in their respective required positions. Similar to the virtual pressure sensor model containing the above equations (3) and (4), a virtual pressure sensor model suitable for the coordinated operation of n actuators can be established, which includes the first to nth derivatives of the pump outlet pressure with respect to time. The expressions for the first and second derivatives of the pump outlet pressure with respect to time are:

[0073]

[0074] By taking the derivative of equation (6) with respect to time, we can obtain the expressions for the third to nth derivatives of the pump inlet pressure with respect to time, which will not be described in detail here.

[0075] The expression for the first to nth derivatives of the pump inlet pressure with respect to time is in the form of n equations, containing n unknowns, namely the working port pressures p of the n control valves. j This refers to the load-side pressure of n control valves. These n equations constitute a virtual pressure sensor physical model applicable to the linkage of n actuators in a hydraulic system. Using this virtual pressure sensor physical model, the load-side pressure of each control valve in the current control cycle can be solved.

[0076] It should be noted that the load-side pressures of each control valve determined by the aforementioned virtual pressure sensor physical model may be difficult to achieve with high engineering accuracy under conditions such as complex hydraulic system operation, a large number of actuators requiring linkage, and significant external interference. This is the initial intention of this application to construct the virtual pressure sensor using an AI model rather than a physical model. Therefore, the aforementioned physical model demonstrates which signals are necessary and feasible for calculating (actually estimating) the load-side pressure of the control valve of interest, thus providing a theoretical basis for determining the input data when constructing the AI ​​model of this application.

[0077] Based on the analysis of the above physical model, the input signal of the AI ​​model set in the control unit 3 of this application was determined. Figure 2 The five input signals required for this AI model are shown below, and they will be introduced one by one.

[0078] 1) Pump speed n p The rotational speed of hydraulic pump 2 can be collected in real time by the rotational speed sensor of hydraulic pump 2 or the rotational speed sensor of power source 1.

[0079] 2) Pump inlet pressure p p The output pressure of hydraulic pump 2 can be collected in real time by the pressure sensor at the output of hydraulic pump 2.

[0080] 3) Pump displacement V g The displacement of hydraulic pump 2 (the amount of oil output per revolution) can be acquired in real time by a pump sway angle sensor (for schemes where the sway angle of hydraulic pump 2 is controlled by a pressure sensing oil circuit (not shown) in the hydraulic system) or by a pump sway angle control current (for EP pumps).

[0081] 4) The control signal I of the control valve is the driving current of the electromagnet of the control valve. The control unit 3 can apply the control signal I to a certain control terminal of the control valve to determine the corresponding valve position and opening area A of the control valve. Here, the control signal I, as an input signal, can be the control signals of multiple control valves collected by the control unit itself, or the control signal commands of multiple control valves calculated and output by the control unit. The opening area of ​​the valve core of the control valve has a fixed relationship (A~I) after design, which is independent of the intermediate transmission process.

[0082] 5) The hydraulic oil temperature T is collected in real time by an oil temperature sensor.

[0083] Furthermore, for cases where the output end (or connected high-pressure oil circuit) of hydraulic pump 2 is equipped with a relief valve, the input signal received by AI model M may also include the relief valve flow rate Q, which is not marked in the figure. ot It can be based on the pump inlet pressure p p Q can be obtained from the overflow valve characteristic curve.ot It should be noted that the hydraulic system described in this application may not include or involve the aforementioned relief valve flow rate Q. ot Therefore, when constructing the AI ​​model M, the input signal may or may not include the overflow valve flow rate Q. ot .

[0084] Furthermore, the total pipeline volume V from the pump inlet to each control valve can be considered as the total pipeline volume between the pump and each control valve of interest, and is generally considered a fixed value; therefore, it can be excluded as an input signal received by the AI ​​model M. If the total pipeline volume V is different at different times, then the input signal received by the AI ​​model M can include the total pipeline volume V.

[0085] Further analysis of the above physical model reveals that, when determining the m-th (n0≥m≥1, preferably n0≥m≥2) derivative of a certain parameter with respect to time, it is necessary to utilize the value of that parameter at least from time k to (km), that is, to include at least time k and the values ​​of the parameter at the preceding m (the first to the mth preceding times).

[0086] Therefore, for the AI ​​model M, assuming k represents any time of interest (not necessarily limited to the current time), its output signal is the load-side pressure of the m (n0≥m≥1) control valves of interest at time k, i.e., p j Let j = 1 to m, and the input signal be the value of each input signal at time k and at least the values ​​of the input signals at the preceding m times (k-1, k-2, ..., km). Here, I contains the control signals of all the m control valves of interest at time k and at least the preceding m times. Therefore, the AI ​​model M of the virtual pressure sensor in this application adopts an AI model that processes time series data.

[0087] As the input signal for the AI ​​model M, it must be obtainable by the control unit 3. p p p V g The five parameters I, T, and I are required and, as mentioned above, are available. The input signal may also include an optional Q. ot V may also include other signals that are considered potentially related to the load-side pressure of the control valve. For example, some control valves have a pressure sensor or flow meter at the load end, so the load-side pressure and / or flow rate measured by the pressure sensor and / or flow meter at each moment can also be included in the input signal of the AI ​​model M.

[0088] The AI ​​model M of the virtual pressure sensor in this application can adopt an existing AI model for processing time series data. There are many AI models for processing time series data in the field of AI, and those skilled in the art can choose one according to their own needs when designing a virtual pressure sensor.

[0089] According to an exemplary embodiment, the AI ​​model M of the virtual pressure sensor in this application adopts the Transformer Encoder model, such as... Figure 3 As shown. Assume this Transformer Encoder model is used to predict the load-side pressure of m control valves in a hydraulic system, where m is less than or equal to (preferably equal to) the total number of control valves n0 in the hydraulic system. The input signals of this Transformer Encoder model include the time-series data of each input signal of the AI ​​model M described above, i.e., the required n... p p p V g I, T (and possibly Q) ot The input data (including parameters such as V) at the current time (time k) and at least m times (from the 1st to the mth time) preceding time k. Optionally, the input data of the Transformer Encoder model includes the data of each input signal at the current time (time k) and x times (from the 1st to the xth time) preceding time k, where x > m, for example, x is 128, which is beneficial for the construction and operation of the AI ​​model M.

[0090] The output data (Outputs) of the Transformer Encoder model are the predicted values ​​of the load-side pressure of the m control valves.

[0091] This Transformer Encoder model contains:

[0092] The Input Embedding layer projects the feature dimensions of the input data onto the input feature dimensions of the encoder layer Nx. It can be a fully connected layer.

[0093] The Positional Encoding layer acquires and marks the order of the input data in time series to form the positional information of the input data. For example, it can be implemented using VanillaAPE (vanilla sinusoidal function-based encoding) encoding technology.

[0094] Multiple stacked encoder layers Nx are computed multiple times based on data from the Input Embedding module and the Positional Encoding module. Within each encoder layer Nx, the following steps are executed sequentially: Multi-Head Attention model, Add&Norm layer, Feed Forward model, Add&Norm layer;

[0095] Self-Attention Pooling layer: This is a pooling layer that focuses on self-attention.

[0096] Prediction Layer: This is a fully connected layer that projects the output feature dimension of the last encoder layer Nx onto the dimension of the control valve load pressure that needs to be predicted.

[0097] The number of stacked encoder layers Nx in this Transformer Encoder model can be set. The encoder layer Nx is a complete, general-purpose Transformer Encoder Layer with various parameters that can be configured, including:

[0098] The desired dimensions of the input features (required);

[0099] The number of attention heads required for multi-head attention;

[0100] The dimensions of Feed Forward (default = 2048);

[0101] Dropout value (default = 0.1);

[0102] The activation function for the intermediate layer can be either ReLU or GeLU (default = ReLU).

[0103] The Transformer Encoder model is a well-known AI model capable of processing time series data, and will not be described in detail here. After constructing the AI ​​model M of the virtual pressure sensor in this application using the Transformer Encoder model, the AI ​​model M needs to be trained.

[0104] Specifically, the hydraulic system that will use this AI model M will be built, for example... Figure 1The hydraulic system shown includes n0 actuators and corresponding n0 control valves. It is assumed that the load-side pressure of m control valves of interest is measured by virtual pressure sensors based on the AI ​​model M, where n0 ≥ m. A pressure sensor is installed at each working port of each of these m control valves of interest. Based on changes in pump displacement, pump speed, and the control of these m control valves of interest, the load-side pressure of each relevant working port is collected sequentially using the pressure sensors, acquiring the input signal data required by the AI ​​model M at each moment. Approximately 80% of the collected data is used as training data, and the remaining 20% ​​is used as test data.

[0105] During training, the input signal data at each time step in the training data is used as the input to the AI ​​model M, and the load end pressure values ​​of the m control valves of interest predicted by the model at the current time are used as the output of the AI ​​model M, thereby training the AI ​​model M.

[0106] The trained AI model M is integrated into the control unit 3, which then serves as a virtual pressure sensor for the control valve load side of this hydraulic system. It's important to note that if the hydraulic system hardware changes, retraining is required based on the steps described above.

[0107] The virtual pressure sensor of the AI ​​model M takes the input signal data from the current moment and the previous x moments (note: the duration of the input signal in actual use of the AI ​​model M is the same as the duration of the input signal used during training) as input. By running the Transformer Encoder model, it can obtain estimated values ​​of the load-side pressure of m control valves of interest. These estimated values ​​can be used by the control unit 3 to determine the control signals (control current) of each control valve in the current control cycle, or they can be used to monitor the load pressure to monitor the state of the hydraulic system.

[0108] To verify the performance of the AI ​​model M built based on the Transformer Encoder model, performance verification tests were conducted. Figure 5 As a comparative example, it is shown Figure 1 The inference result of the working pressure of the first working port of the control valve V1 in the hydraulic system shown. Figure 6 This is a prediction of the load-side pressure of the first working port of control valve V1 obtained by using the test dataset and the AI ​​model M built based on the Transformer Encoder model. Figure 5 , Figure 6 The horizontal axis represents time, and the vertical axis represents the pressure at the load end of the control valve. Comparing the data in the two graphs, we can see that the squared difference between the predicted and inferred results for the first working port of control valve V1 after 5000 seconds is 135.5 bar. 2Furthermore, the predicted results and the inference results generally follow the same trend. Similar test results were obtained for other working ports. The test results fully demonstrate the effectiveness and high accuracy of the AI ​​model M built based on the TransformerEncoder model in this application.

[0109] According to another exemplary embodiment, the AI ​​model M of the virtual pressure sensor in this application adopts a CNN model. A CNN model is a model capable of processing time series data; here, we take... Figure 4 The ResNet model shown is an example. Assume this ResNet model is used to predict the load-side pressure of m control valves in a hydraulic system, where m is less than or equal to (preferably equal to) the total number of control valves n0 in the hydraulic system. The input signals of this ResNet model include the time-series data of each input signal of the AI ​​model M described earlier, i.e., the required n... p p p V g I, T (and possibly Q) ot The input data (including parameters such as V) is the data at the current time (time k) and at least m times (the first to the mth times preceding time k). Optionally, the input data of the ResNet model includes the data of each input signal at the current time (time k) and x times (the first to the xth times preceding time k), where x>m, for example, x is 128, which is beneficial for the construction and operation of the AI ​​model M.

[0110] The output data of the ResNet model, Outputs, are the predicted values ​​of the load-side pressure of the m control valves.

[0111] This ResNet model includes:

[0112] Input Embedding layer: projects the feature dimensions of the input data onto the input feature dimensions of the ResNet; it can be a fully connected layer.

[0113] Positional Encoding layer: Acquires and marks the order of the input data in time series to form the positional information of the input data, for example, by using VanillaAPE (vanilla sinusoidal function-based encoding) encoding technology;

[0114] The ResNet module is a standard ResNet module that does not require pre-trained parameters. It can be ResNet18 or other models built on the ResNet base module.

[0115] The ResNet model, along with other CNN models capable of processing time series data, is well-known in the AI ​​field and will not be described in detail here. After constructing the AI ​​model M of the virtual pressure sensor in this application using the ResNet model, the AI ​​model M based on the ResNet model is trained using steps similar to those for the Transformer Encoder model. Then, the trained AI model M is integrated into the control unit 3, which can serve as a virtual pressure sensor for the load end of the control valve in the hydraulic system.

[0116] Similarly, performance verification tests were conducted to verify the performance of the AI ​​model M built on the ResNet model. Figure 7 This is a prediction result of the load end pressure of the first working port of control valve V1 obtained by using the test dataset and the AI ​​model M built based on the ResNet model. Figure 7 The horizontal axis represents time, and the vertical axis represents the pressure at the load end of the control valve. (Comparison) Figure 5 , Figure 7 The data in the figure shows that the squared difference between the predicted and inferred results for the first working port of control valve V1 over 5000 seconds is 76.7 bar. 2 Furthermore, the predicted results and the inference results generally follow the same trend. Similar test results were obtained for other working ports. The test results fully demonstrate the effectiveness and high accuracy of the AI ​​model M built based on the ResNet model in this application.

[0117] It can be expected that using other AI models capable of processing time series to construct the virtual pressure sensor for hydraulic systems described in this application will also be effective and highly accurate.

[0118] The hydraulic system control unit of this application incorporates a virtual pressure sensor built based on an AI model that processes time series data. This virtual sensor is used to predict the load-side pressure of some or even all control valves in the hydraulic system. This allows a single virtual pressure sensor to replace multiple physical sensors in the hydraulic system, reducing hardware costs. For some hydraulic systems, due to complexity, high integration, and limited installation space, it may be difficult to install certain physical sensors. Using an AI-based virtual pressure sensor to replace the original physical sensors allows for a higher degree of integration or a more complex system structure within a limited space. Alternatively, a virtual pressure sensor can be used simultaneously without replacing the original physical sensors. In this case, the detection value of the AI-based virtual pressure sensor can serve as a backup (redundancy) for the detection values ​​of the physical sensors, used to correct or verify their values. When a physical sensor malfunctions, the detection value of the virtual pressure sensor can be used to maintain the original functions of the hydraulic system.

[0119] Compared to virtual pressure sensors based on physical models, the AI ​​model-based virtual pressure sensor of this application offers improved detection accuracy and is easier to implement in engineering. Therefore, it can be more reliably applied to hydraulic systems, especially those in engineering equipment.

[0120] This application presents a solution using an AI-based virtual pressure sensor, applicable to various hydraulic systems, particularly those in construction machinery, including, but not limited to, LUDV (Load Independent Flow Distribution) systems. According to this application, the functionality of a hydraulic system, especially a LUDV system, is achieved through software and simplified hardware. The hydraulic system implements the pressure detection function of a physical sensor using appropriate software, with a virtual pressure sensor used on the actuator side to replace the original physical pressure sensor or as a backup.

[0121] It should be noted that, in order to achieve mutual coordination among multiple actuators in the hydraulic system of construction machinery, thereby enabling the working device to have high controllability, improve work efficiency and precise work trajectory, and reduce power source energy consumption loss, LUDV systems are increasingly being adopted. In LUDV systems, pressure compensation valves are used to increase the pressure at the port of the flow control valve of the low-load actuator to be equal to that of the highest load. Thus, the flow rate distributed by the hydraulic pump to each path is only related to the opening area of ​​the flow control valve of each path, achieving independent flow distribution for each path regardless of the load size. LUDV systems typically require complex hardware, resulting in high design and manufacturing costs. When implementing the LUDV system functionality using a virtual pressure sensor and a simple electro-hydraulic valve as described in this application, some corresponding physical sensors can be replaced, significantly reducing the design and manufacturing costs of the hydraulic system. Alternatively, it can serve as a backup (redundancy) for physical sensors, providing system functions with higher accuracy and reliability.

[0122] Furthermore, in applications where virtual pressure sensors are used to replace physical pressure sensors, the likelihood of system failures caused by physical failures of the pressure sensors can be reduced.

[0123] Furthermore, those skilled in the art will understand that the physical model for the virtual pressure sensor described above was not actually used in the AI ​​model of the virtual pressure sensor. However, the establishment and analysis of the physical model of the virtual pressure sensor contributes to the establishment of the AI ​​model of the virtual pressure sensor in this application, and can be considered the starting point for modeling the AI ​​model. Based on the analysis of the physical model, this application has determined what specific input signals the AI ​​model contains. The input signals of the AI ​​model may include only n p p p V gThe essential parameters are I, T, etc., or Q can be added according to customer needs. ot Optional parameters such as V are used. On the one hand, the determined input signals are necessary for obtaining the output signal through the AI ​​model, thus ensuring the accuracy of the output signal. On the other hand, the determined input signals are sufficient for obtaining the output signal through the AI ​​model, so the AI ​​model does not include or minimizes signals with low correlation to the output signal, ensuring that the AI ​​model has low computational cost and is easy to implement in engineering.

[0124] While this application has been described herein with reference to specific embodiments, the scope of this application is not limited to the details shown. Various modifications may be made to these details without departing from the basic principles of this application.

Claims

1. A virtual pressure sensor for a hydraulic system, the hydraulic system comprising: A hydraulic pump (2) and multiple actuators (C1, C2...), each actuator being supplied with hydraulic oil by the hydraulic pump (2) via a corresponding hydraulic circuit, each hydraulic circuit being equipped with a corresponding control valve (V1, V2...), the virtual pressure sensor being configured to predict the load end pressure of m control valves in the hydraulic system; The virtual pressure sensor includes an AI model using time series data, and the input signal of the AI ​​model includes the pump speed (n). p ), pump inlet pressure (p) p ), pump displacement (V) g The data of the control signals (I) and hydraulic oil temperature (T) of the m control valves at sequential times, wherein the sequential times include the current time and at least the previous m times; The output signal of the AI ​​model includes the predicted load-side pressure values ​​of the m control valves.

2. The virtual pressure sensor as described in claim 1, wherein, The output end of the hydraulic pump (2) is equipped with a relief valve, and the input signal of the AI ​​model also includes the relief valve flow rate (Q). ot The data at the specified time in the sequence.

3. The virtual pressure sensor as described in claim 1 or 2, wherein, The input signal of the AI ​​model also includes data on the total pipeline volume (V) from the pump inlet to the m control valves at the sequence time.

4. The virtual pressure sensor as described in any one of claims 1-3, wherein, If at least one of the control valves in the hydraulic system is equipped with a pressure sensor and / or flow meter at the load end, the input signal of the AI ​​model also includes the measurement data measured by the pressure sensor and / or flow meter at the sequence time.

5. The virtual pressure sensor as described in any one of claims 1-4, wherein, The AI ​​model is the TransformerEncoder model.

6. The virtual pressure sensor as described in claim 5, wherein, The Transformer Encoder model contains multiple stacked encoder layers, and each encoder layer has multiple configurable parameters.

7. The virtual pressure sensor as described in any one of claims 1-4, wherein, The AI ​​model is a CNN model, such as the ResNet model.

8. The virtual pressure sensor as described in any one of claims 1-7, wherein, m equals the total number of control valves in the hydraulic system.

9. The virtual pressure sensor as described in any one of claims 1-8, wherein, The virtual pressure sensor is integrated into the control unit (3) of the hydraulic system.

10. A control unit (3) for a hydraulic system, the hydraulic system comprising: A hydraulic pump (2) and multiple actuators (C1, C2...), each actuator is supplied with hydraulic oil by the hydraulic pump (2) via a corresponding hydraulic circuit, and a corresponding control valve (V1, V2...) is provided in each hydraulic circuit; The control unit (3) includes a virtual pressure sensor as described in any one of claims 1-9, for controlling the operation of the hydraulic system and / or monitoring the state of the hydraulic system based on the predicted load end pressure values ​​of m control valves predicted by the virtual pressure sensor.

Images