Artificial intelligence-based multi-stage roots pump control method and system

Abstract

The application relates to the technical field of multi-stage Roots pumps and discloses a multi-stage Roots pump control method and system based on artificial intelligence, which comprises the following steps: collecting real-time operation data signals of a multi-stage Roots pump through Hall sensors, pressure sensors and angular velocity sensors, wherein the data signals comprise air outlet pressure, magnetic field intensity and rotating speed signals; performing dynamic characteristic extraction on the data signals, processing the data signals through a normalization model and generating standardized characteristic data; constructing a data twin model based on actual working condition data; and migrating the standardized characteristic data to the data twin model by using a transfer learning mechanism.The application can quickly migrate the standardized characteristic data to an actual working condition model by constructing a data twin model, shorten a training cycle and improve the model generalization capability; and through light-weight edge deployment, real-time data acquisition and prediction deviation calculation are realized, so that the phenomenon of local overload or efficiency loss can be avoided.

Description

Technical Field

[0001] This invention relates to the field of multistage Roots pump technology, and in particular to a multistage Roots pump control method and system based on artificial intelligence. Background Technology

[0002] As a core component of vacuum systems, the operating efficiency and stability of multistage Roots pumps directly impact the performance of industrial production. Traditional control methods for multistage Roots pumps often rely on fixed-parameter PID control or empirical rules, which struggle to address the pressure coupling issues of individual pumping chambers under complex dynamic conditions.

[0003] An existing patent discloses a variable phase angle multistage Roots vacuum pump and its operating method (publication number CN116517828A). The variable phase angle multistage Roots vacuum pump includes a pump casing, a right rotor 1, and a left rotor 2, wherein the right rotor 1 and left rotor 2 rotate in opposite directions around their respective axes of rotation. In the technology disclosed in this patent, due to the uneven pressure between stages during the operation of the multistage Roots pump, and the difficulty in dynamically controlling the pressure of each individual pumping chamber, local overload or efficiency loss is easily caused. Summary of the Invention

[0004] This invention provides a multi-stage Roots pump control method and system based on artificial intelligence to solve existing technical problems, thereby resolving the issue of pressure differences in Roots pumps at different stages.

[0005] To solve the above-mentioned technical problems, according to one aspect of the present invention, more specifically, a multi-stage Roots pump control method based on artificial intelligence, comprising:

[0006] S1. Real-time operating data signals of the multi-stage Roots pump are collected through Hall sensors, pressure sensors, and angular velocity sensors. The data signals include outlet pressure, magnetic field strength, and rotational speed signals. Dynamic features are extracted from the data signals, and standardized feature data is generated by processing them through a normalization model.

[0007] S2. Construct a data twin model based on actual working condition data, and use a transfer learning mechanism to transfer standardized feature data to the data twin model;

[0008] S3. With the joint optimization objectives of maximizing the energy efficiency ratio and minimizing the deviation, a deep deterministic strategy gradient algorithm is used to train reinforcement learning to generate the optimal decoupling control strategy for pressure distribution between multi-stage Roots pump stages.

[0009] S4. According to the decoupling control strategy, the speed of each level of motor is adjusted in real time by the PID controller;

[0010] S5. Lightweight edge deployment is carried out based on the data twin model to collect control effect data in real time and calculate the prediction deviation; when the prediction deviation exceeds the preset threshold, the incremental update mechanism of the single Roots pump control model is triggered.

[0011] Furthermore, the normalization model is used to perform min-max normalization on each sensor signal individually, scaling it to the [0, 1] interval, as follows:

[0012]

[0013] In the formula, c represents the standard data after normalization processing of the i-th type of sensor; c i c represents the raw data from the i-th type of sensor; min c max These represent the minimum and maximum theoretical values ​​detected by the i-th type of sensor, respectively.

[0014] Furthermore, the data twin model calculates the prediction deviation based on the deviation of the outlet pressure over the periodic time, and the deviation between the outlet pressure change period and the magnetic field strength change period, as follows:

[0015]

[0016] In the formula, g represents the deviation of the multi-stage Roots pump calculated by the data twin model; m represents the deviation of the outlet pressure within the period; and a represents the deviation between the outlet pressure change period and the magnetic field strength change period.

[0017] Furthermore, when g ≥ 78%, it indicates that the prediction deviation exceeds the preset threshold, thereby triggering the incremental update mechanism of the single Roots pump control model;

[0018] When g < 78%, it means that there is no need to trigger the model incremental update mechanism.

[0019] Furthermore, in step S3, the reward function of the deep deterministic policy gradient algorithm is defined as a weighted inverse proportional function of energy efficiency ratio and deviation.

[0020] Furthermore, the specific process of the PID controller in step S4 is as follows:

[0021] 1) Compare the rotational speed signal fed back by the angular velocity sensor with the input signal of the motor speed in the control model for a single Roots pump;

[0022] 2) The deviation between the input signal and the actual output after comparison is added to the input signal.

[0023] Furthermore, the decoupling control strategy is used to perform customized control based on the state of different individual Roots pump control models in a multi-stage Roots pump.

[0024] An AI-based multi-stage Roots pump control system includes:

[0025] The multi-source data sensing module is used to collect real-time operating data signals of the multi-stage Roots pump, including outlet pressure, magnetic field strength and rotation speed signals.

[0026] The data twin modeling module is used to construct a data twin model based on actual working condition data and to transfer standardized feature data to the data twin model using a transfer learning mechanism.

[0027] The hierarchical reinforcement learning module uses a deep deterministic policy gradient algorithm to train reinforcement learning and generate the optimal decoupling control strategy for pressure distribution between multi-stage Roots pump stages.

[0028] The interstage decoupling control module adjusts the speed of each stage motor in real time through a PID controller;

[0029] The lightweight edge deployment module collects control effect data in real time and calculates prediction deviations; when the prediction deviation exceeds a preset threshold, it triggers an incremental update mechanism for the control model of a single Roots pump.

[0030] The multi-stage Roots pump control method and system based on artificial intelligence provided by this invention have the following advantages compared with the prior art:

[0031] 1. This invention generates an optimal pressure allocation strategy between stages by using a deep deterministic strategy gradient algorithm with the joint optimization objectives of maximizing energy efficiency ratio and minimizing deviation. This significantly reduces coupling effects and improves overall energy efficiency by about 15%-20%. Furthermore, the decoupling control strategy supports customized regulation of multi-stage Roots pumps to adapt to the differences in operating states of different levels of equipment.

[0032] 2. This invention constructs a data twin model to quickly transfer standardized feature data to the actual working condition model, shortening the training cycle and improving the model's generalization ability; through lightweight edge deployment, it achieves real-time data acquisition and prediction deviation calculation, ensuring that there will be no local overload or efficiency loss.

[0033] 3. This invention optimizes the Roots pump model only for abnormal levels by triggering incremental updates based on the prediction deviation threshold, thereby reducing computational resource consumption and improving system robustness. Furthermore, this mechanism can improve model update efficiency by 40% and reduce the risk of failure downtime.

[0034] 4. This invention eliminates dimensional differences and improves data consistency by performing dynamic feature extraction and min-max normalization on Hall sensor, pressure sensor and angular velocity signals, providing high-quality input for subsequent algorithms. Attached Figure Description

[0035] Figure 1 This is a flowchart of the control method in this invention;

[0036] Figure 2 This describes the pumping principle of the Roots pump in this invention.

[0037] Figure 3 This is a schematic diagram showing the distribution of sensors in this invention;

[0038] Figure 4 This is a schematic diagram of the periodic changes in magnetic field and pressure in this invention;

[0039] Figure 5 This is a schematic diagram of the deviation g and deviation m in this invention;

[0040] Figure 6 This is a schematic diagram of the deviation g and deviation a in this invention;

[0041] Figure 7 This is a model diagram of the deviation g, deviation m, and deviation a in this invention.

[0042] In the diagram: 1. Hall sensor; 2. Pressure sensor; 3. Magnetic rotor. Detailed Implementation

[0043] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0044] like Figure 2 , 3 As shown in Figure 4, the present invention provides a multi-stage Roots pump control method or system based on artificial intelligence, which is applied to multi-stage Roots pumps. A multi-stage Roots pump is composed of multiple (usually an even number) Roots pump control models (such as...). Figure 3 As shown, two Roots pump control models are connected in series to form a multi-stage Roots pump. Furthermore, two magnetic rotors form a separate extraction chamber.

[0045] The single Roots pump control model includes two magnetic rotors 3 (the magnetic rotors are permanent magnets with different magnetic poles at both ends), a Hall sensor 1, and a pressure sensor 2 located at the exhaust port. The pressure sensor 2 is used to monitor the air pressure at the exhaust port.

[0046] The pumping principle of a single Roots pump is as follows:

[0047] a. such as Figure 2 As shown in Figure (1), the magnetic field strength detected by Hall sensor 1 at this time is as follows: Figure 4 As shown in a(1) in the figure, and the pressure detected by pressure sensor 2 is as follows: Figure 4 As shown in b(1);

[0048] b, such as Figure 2 As shown in (2), the magnetic field strength detected by Hall sensor 1 at this time is as follows: Figure 4 As shown in a(2) in the figure, and the pressure detected by pressure sensor 2 is as follows Figure 4 As shown in b(2);

[0049] c. For example Figure 2 As shown in (3), the magnetic field strength detected by Hall sensor 1 at this time is as follows: Figure 4 As shown in a(3) in the figure, and the pressure detected by pressure sensor 2 is as follows Figure 4 As shown in b(3);

[0050] d. For example Figure 2 As shown in (3), the magnetic field strength detected by Hall sensor 1 at this time is as follows: Figure 4 As shown in a(3) in the figure, and the pressure detected by pressure sensor 2 is as follows Figure 4 As shown in b(3);

[0051] e. For example Figure 2 As shown in (4), the magnetic field strength detected by Hall sensor 1 at this time is as follows: Figure 4 As shown in a(4) in the figure, and the pressure detected by pressure sensor 2 is as follows Figure 4 As shown in b(4);

[0052] f. For example Figure 2 As shown in (5), the magnetic field strength detected by Hall sensor 1 at this time is as follows: Figure 4 As shown in a(5) in the figure, and the pressure detected by pressure sensor 2 is as follows Figure 4 As shown in b(5) in the figure.

[0053] The AI-based multi-stage Roots pump control system includes: a multi-source data sensing module for acquiring real-time operating data signals of the multi-stage Roots pump, including outlet pressure, magnetic field strength, and rotational speed signals; a data twin modeling module for constructing a data twin model based on actual operating data and using transfer learning to transfer standardized feature data to the data twin model; a hierarchical reinforcement learning module for training reinforcement learning using a deep deterministic policy gradient algorithm to generate the optimal decoupling control strategy for inter-stage pressure distribution of the multi-stage Roots pump; an inter-stage decoupling control module for adjusting the speed of each stage motor in real time using a PID controller; and a lightweight edge deployment module for acquiring control effect data in real time and calculating prediction deviations; and triggering an incremental update mechanism for the individual Roots pump control model when the prediction deviation exceeds a preset threshold.

[0054] Example 1

[0055] Real-time operating data signals of the multi-stage Roots pump are collected using Hall effect sensors, pressure sensors, and angular velocity sensors. The data signals include outlet pressure, magnetic field strength, and rotational speed. Dynamic features are extracted from the data signals, and standardized feature data are generated by processing them through a normalization model.

[0056] The normalization model is used to perform min-max normalization on each sensor signal individually, scaling it to the [0, 1] interval, as follows:

[0057]

[0058] In the formula, c represents the standard data after normalization processing of the i-th type of sensor; c i c represents the raw data from the i-th type of sensor; min c max These represent the minimum and maximum theoretical values ​​detected by the i-th type of sensor, respectively.

[0059] Specifically, the deviation of the outlet pressure within the cycle time is calculated. For the pressure sensor in the first cycle (e.g.) Figure 4 The pressure in b (2) is calculated using the normalized standard data within the time period. The pressure of the pressure sensor in the first cycle is taken as c1 = 0.8 (MPa), and the minimum and maximum theoretical values ​​detected by the pressure sensor are c1 and c2 respectively. min =0.1 (MPa), c max =1 (MPa). Therefore:

[0060]

[0061] For the pressure sensor in the second cycle (e.g.) Figure 4 The pressure in b (4) is calculated using the normalized standard data within the time period. The pressure of the pressure sensor in the first cycle is taken as c2 = 0.7 (MPa). Then we have:

[0062]

[0063] The above calculations show that the deviation of the outlet pressure between the first and second time cycles is m = 70% - 60% = 10%.

[0064] Specifically, the deviation between the outlet pressure change period and the magnetic field strength change period is calculated. The magnetic field strength change period (e.g., ...) is also considered. Figure 4 The calculation is performed using the normalized standard data of the time intervals (1)-(5) within the middle a. The period of magnetic field strength change is taken as c3 = 8 (ms), and the minimum and maximum theoretical values ​​of the magnetic field change period are c3 and c4 respectively. min =2(ms), c max=15 (ms). Therefore:

[0065]

[0066] For pressure change period (e.g.) Figure 4 The calculation is performed using the normalized standard data of the time intervals (1)-(5) within the middle b. The pressure change period time is taken as c4 = 12 (ms), and the minimum and maximum theoretical values ​​of the magnetic field change period are c4 and c5 respectively. min =5 (ms), c max =16 (ms). Therefore:

[0067]

[0068] The above calculations show that the deviation between the pressure change period and the magnetic field strength change period is a = 63% - 46.1% = 16.9%.

[0069] Example 2

[0070] A data twin model is built based on actual operating condition data, and a transfer learning mechanism is used to transfer standardized feature data to the data twin model. Lightweight edge deployment is then implemented based on the data twin model to collect control performance data in real time and calculate prediction deviations. When the prediction deviation exceeds a preset threshold, an incremental update mechanism for the single Roots pump control model is triggered. By building a data twin model and combining it with a transfer learning mechanism, standardized feature data is quickly transferred to the actual operating condition model, shortening the training cycle and improving the model's generalization ability. Lightweight edge deployment enables real-time data acquisition and prediction deviation calculation, ensuring low-latency control response.

[0071] The data twin model calculates the prediction bias based on the deviation of the outlet pressure over the period and the deviation between the outlet pressure change period and the magnetic field strength change period. The results are as follows:

[0072]

[0073] In the formula, g represents the deviation of the multi-stage Roots pump calculated by the data twin model; m represents the deviation of the outlet pressure within the period; and a represents the deviation between the outlet pressure change period and the magnetic field strength change period.

[0074] The above formula is an empirical analysis to address the issue of pressure differences in different stages of the Roots pump during multi-stage Roots pump control. The data derived from this analysis, along with the characteristic relationships between the data, constitute the external manifestation of the empirical formula. The reasoning process is as follows:

[0075] 1) Fit the formula for the deviation g of the multi-stage Roots pump calculated by the data twin model.

[0076] Among them, the data twin model calculates the deviation g of the multi-stage Roots pump, which can be represented by the proportion of pressure borne by the valves in the sample multi-stage Roots pump.

[0077] For example, data was collected from 100 samples of multi-stage Roots pump cabinets. If, after manual evaluation, the pressure borne by the valves of a particular multi-stage Roots pump (e.g., ...) is... Figure 3 The valve at the intermediate connection point, when the valve bears greater pressure, indicates a greater deviation between the two Roots pumps. If the deviation exceeds the data in the other 50 samples, then the data twin model calculates a deviation of g = 50% for the multi-stage Roots pump.

[0078] All other data were collected by the device itself for subsequent calculations and fitting statistics.

[0079] 2) Establish a mathematical model for the relationship between deviation g and deviation m (e.g.) Figure 5 As shown in the figure (where the red dots represent the distribution of the 100 collected samples), then:

[0080]

[0081] In Formula 1 above, k represents an empirical constant for adjusting the sensitivity of the model.

[0082] 3) Establish a mathematical model for the relationship between deviation g and deviation a (e.g., Figure 6 As shown in the figure (where the red dots represent the distribution of the 100 collected samples), then:

[0083]

[0084] In Formula 2 above, k represents an empirical constant for adjusting the sensitivity of the above model.

[0085] 4) Establish a mathematical model for the relationship between deviation g, deviation m, and deviation a (e.g., Figure 7 As shown, this model reveals a positive correlation between deviation m and deviation a. Combining this with the characteristic relationships derived from Formulas 1 and 2 above, we have:

[0086] g=(Formula 1)×(Formula 2);

[0087] Therefore, based on the above derivation, the condition coefficient g for the energy storage cabinet to call the backup power module is:

[0088]

[0089] Where the deviation of the outlet pressure between the first and second time periods is m = 10%, and the deviation of the pressure change period from the magnetic field strength change period is a = 16.9%, then:

[0090]

[0091] The above calculations show that the data twin model calculates the deviation of the multi-stage Roots pump as g = 83.9%, indicating that the prediction deviation exceeds the preset threshold, thus triggering the incremental update mechanism of the single Roots pump control model. Furthermore:

[0092] Table 1 shows the relationship between some detection parameters and control parameters.

[0093]

[0094]

[0095] As shown in Table 1 above, when the implemented data approaches infinity, the value of the deviation g can be used as a boundary to determine whether to perform an incremental update. When g ≥ 78%, it indicates that the prediction deviation exceeds the preset threshold, thereby triggering the incremental update mechanism of the single Roots pump control model (which means that the rotation speed of the magnetic rotor 3 in the detected single Roots pump needs to be increased); when g < 78%, it indicates that the model incremental update mechanism does not need to be triggered.

[0096] Example 3

[0097] With maximizing energy efficiency ratio and minimizing deviation as the joint optimization objectives, a deep deterministic policy gradient algorithm is used to train reinforcement learning to generate an optimal decoupled control strategy for interstage pressure distribution in a multi-stage Roots pump. This decoupled control strategy is used for customized control based on the states of different individual Roots pump control models within the multi-stage Roots pump system. The reward function of the deep deterministic policy gradient algorithm is defined as a weighted inverse proportional function of energy efficiency ratio and deviation. Based on the deep deterministic policy gradient (DDPG) algorithm, with maximizing energy efficiency ratio and minimizing deviation as the joint optimization objectives, an optimal interstage pressure distribution strategy is generated, significantly reducing coupling effects and improving overall energy efficiency.

[0098] Example 4

[0099] Based on the decoupling control strategy, the speed of each stage of the motor is adjusted in real time through a PID controller; the specific process of the PID controller is as follows:

[0100] 1) Compare the rotational speed signal fed back by the angular velocity sensor with the input signal of the motor speed in the control model for a single Roots pump;

[0101] 2) The deviation between the input signal and the actual output after comparison is added to the input signal.

[0102] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A multi-stage Roots pump control method based on artificial intelligence, characterized in that, include: S1. Real-time operating data signals of the multi-stage Roots pump are collected through Hall sensors, pressure sensors, and angular velocity sensors. The data signals include outlet pressure, magnetic field strength, and rotational speed signals. Dynamic features are extracted from the data signals, and standardized feature data is generated by processing them through a normalization model. S2. Construct a data twin model based on actual working condition data, and use a transfer learning mechanism to transfer standardized feature data to the data twin model; S3. With the joint optimization objectives of maximizing the energy efficiency ratio and minimizing the deviation, a deep deterministic strategy gradient algorithm is used to train reinforcement learning to generate the optimal decoupling control strategy for pressure distribution between multi-stage Roots pump stages. S4. According to the decoupling control strategy, the speed of each level of motor is adjusted in real time by the PID controller; S5. Lightweight edge deployment is carried out based on the data twin model to collect control effect data in real time and calculate the prediction deviation; when the prediction deviation exceeds the preset threshold, the incremental update mechanism of the single Roots pump control model is triggered.

2. The multi-stage Roots pump control method based on artificial intelligence according to claim 1, characterized in that: The normalization model is used to perform min-max normalization on each sensor signal individually, scaling it to the [0, 1] interval, as follows: In the formula, c represents the standard data after normalization processing of the i-th type of sensor; c i c represents the raw data from the i-th type of sensor; min c max These represent the minimum and maximum theoretical values ​​detected by the i-th type of sensor, respectively.

3. The multi-stage Roots pump control method based on artificial intelligence according to claim 1, characterized in that: The data twin model calculates the prediction deviation based on the deviation of the outlet pressure over the periodic time, and the deviation between the outlet pressure change period and the magnetic field strength change period, as follows: In the formula, g represents the deviation of the multi-stage Roots pump calculated by the data twin model; m represents the deviation of the outlet pressure within the period; and a represents the deviation between the outlet pressure change period and the magnetic field strength change period.

4. The multi-stage Roots pump control method based on artificial intelligence according to claim 3, characterized in that: When g≥78%, it means that the prediction deviation exceeds the preset threshold, thereby triggering the incremental update mechanism of the single Roots pump control model; When g < 78%, it means that there is no need to trigger the model incremental update mechanism.

5. The multi-stage Roots pump control method based on artificial intelligence according to claim 1, characterized in that: In step S3, the reward function of the deep deterministic policy gradient algorithm is defined as a weighted inverse proportional function of energy efficiency ratio and deviation.

6. The multi-stage Roots pump control method based on artificial intelligence according to claim 1, characterized in that: The specific process of the PID controller in step S4 is as follows: 1) Compare the rotational speed signal fed back by the angular velocity sensor with the input signal of the motor speed in the control model for a single Roots pump; 2) The deviation between the input signal and the actual output after comparison is added to the input signal.

7. The multi-stage Roots pump control method based on artificial intelligence according to claim 1, characterized in that: The decoupling control strategy is used to perform customized control based on the state of different individual Roots pump control models in a multi-stage Roots pump.

8. A multi-stage Roots pump control system based on artificial intelligence, characterized in that, The control system, applied to the control method according to any one of claims 1-7, comprises: The multi-source data sensing module is used to collect real-time operating data signals of the multi-stage Roots pump, including outlet pressure, magnetic field strength and rotation speed signals. The data twin modeling module is used to construct a data twin model based on actual working condition data and to transfer standardized feature data to the data twin model using a transfer learning mechanism. The hierarchical reinforcement learning module uses a deep deterministic policy gradient algorithm to train reinforcement learning and generate the optimal decoupling control strategy for pressure distribution between multi-stage Roots pump stages. The interstage decoupling control module adjusts the speed of each stage motor in real time through a PID controller; The lightweight edge deployment module collects control effect data in real time and calculates prediction deviations; when the prediction deviation exceeds a preset threshold, it triggers an incremental update mechanism for the control model of a single Roots pump.

Images