Magnetic suspension compressor, control method and device thereof, storage medium and program product

Abstract

The invention discloses a magnetic suspension compressor and a control method and device thereof, a storage medium and a computer program product. The method comprises the steps that more than one disturbance is introduced, and control parameters of a displacement current double-closed-loop control system are adjusted; obtaining rotor displacement and stator coil current of the magnetic bearing, obtaining control parameters, the rotor displacement and the stator coil current corresponding to all disturbances, and recording the control parameters, the rotor displacement and the stator coil current as a first data set; training and testing by using more than two preset training models based on the first data set to obtain an initial control parameter prediction model; compressor operation parameters and unit operation parameters are obtained, and input and output of the initial control parameter prediction model, the compressor operation parameters and the unit operation parameters are recorded as a second data set; training and testing are conducted again based on the second data set, and the updated control parameter prediction model is obtained. According to the scheme, the control parameter prediction model is obtained by using more than two preset training models, the bearing control performance is improved, and the operation stability is ensured.

Description

Technical Field

[0001] The present invention belongs to the technical field of magnetic levitation compressors, and specifically relates to a control method, device, magnetic levitation compressor, storage medium and computer program product for a magnetic levitation compressor, and in particular to an adaptive optimal parameter control method, device, magnetic levitation compressor, storage medium and computer program product for a magnetic levitation compressor based on machine learning. Background Art

[0002] The magnetic levitation compressor has an active magnetic bearing (AMB)-rotor system. The active magnetic bearing-rotor system has attracted much attention due to its advantages over the mechanical bearing-rotor system, such as no friction and low energy consumption. However, the bearing control of the active magnetic bearing-rotor system involves a control algorithm, and the control parameters need to be adjusted to ensure excellent bearing control performance. It is usually impossible to manually find the optimal control parameters, and it is difficult to ensure the optimal bearing control performance.

[0003] The above contents are only used to assist in understanding the technical solution of the present invention and do not constitute an admission that the above contents are prior art. Summary of the invention

[0004] The purpose of the present invention is to provide a control method, device, magnetic levitation compressor, storage medium and computer program product for a magnetic levitation compressor, so as to solve the problem that it is usually impossible to manually find the optimal bearing control parameters and it is difficult to ensure the optimal bearing control performance in the bearing control of an active magnetic levitation bearing-rotor system, so as to achieve the effect of obtaining a control parameter prediction model of the magnetic levitation compressor by using an extreme learning machine model, a support vector machine model and a random forest model to improve the bearing control performance and ensure the operating stability of the magnetic levitation compressor.

[0005] The present invention provides a control method for a magnetic levitation compressor, wherein the magnetic levitation compressor has a magnetic bearing, and the control system of the magnetic bearing adopts a displacement current double closed-loop control system; the control method for the magnetic levitation compressor comprises: when the magnetic levitation compressor is running, introducing one or more preset disturbances into the displacement current double closed-loop control system; and adjusting the control parameters of the displacement current double closed-loop control system more than once; for each disturbance among the one or more preset disturbances introduced, when the magnetic bearing is running based on the control parameters of the displacement current double closed-loop control system adjusted each time, obtaining the rotor displacement of the magnetic bearing and the stator coil current of the magnetic bearing; thereby, obtaining the control parameters, rotor displacement and stator coil current of the displacement current double closed-loop control system corresponding to all disturbances among the one or more preset disturbances introduced, which are recorded as a first data group; based on the The first data group is trained and tested using more than two preset training models to obtain an initial control parameter prediction model of the magnetic levitation compressor, so as to control the operation of the magnetic bearing according to the initial control parameter prediction model of the magnetic levitation compressor; when the displacement current dual closed-loop control system operates according to the initial control parameter prediction model, the operating parameters of the magnetic levitation compressor are obtained, which are recorded as compressor operating parameters; and the operating parameters of the unit where the magnetic levitation compressor is located are obtained, which are recorded as unit operating parameters; the input and output of the initial control parameter prediction model, the compressor operating parameters and the unit operating parameters are recorded as a second data group; based on the second data group, training and testing are performed using more than two preset training models to obtain an updated control parameter prediction model of the magnetic levitation compressor, so as to control the operation of the magnetic bearing according to the updated control parameter prediction model of the magnetic levitation compressor.

[0006] In some embodiments, when the magnetic levitation compressor is running, one or more preset disturbances are introduced into the displacement current dual closed-loop control system, including: adding one or more interference sources to the feedback value of the rotor displacement of the displacement loop of the displacement current dual closed-loop control system to achieve the introduction of one or more preset disturbances into the displacement current dual closed-loop control system; wherein the one or more interference sources include: at least one of Gaussian white noise of a set frequency, sinusoidal interference signals of different amplitudes, and random pulse signals of different amplitudes; and / or, two or more preset training models, including: at least two of an extreme learning machine model, a support vector machine model, and a random forest model.

[0007] In some embodiments, based on the first data group, training and testing are performed using two or more preset training models to obtain an initial control parameter prediction model of the magnetic levitation compressor, including: based on each of the two or more preset training models, preprocessing the first data group to obtain first sample data under each training model; training and testing each training model using the first sample data under each training model to obtain a control parameter prediction model under each training model; thereby, obtaining a control parameter prediction model under each of the two or more preset training models; performing weighted averaging processing on the control parameter prediction model under each of the two or more preset training models to obtain a weighted average control parameter prediction model of the two or more preset training models as the initial control parameter prediction model of the magnetic levitation compressor.

[0008] In some embodiments, based on each of more than two preset training models, the first data group is preprocessed to obtain the first sample data under each training model, including: based on the first data group, determining the control parameters corresponding to the minimum rotor displacement and the minimum stator coil current under each of more than one preset disturbance introduced, as the optimal control parameters under each disturbance; thereby, obtaining the optimal control parameters, minimum rotor displacement and minimum stator coil current corresponding to all disturbances in more than one preset disturbance introduced, which are recorded as the initial data group; for the initial data group, based on the demand relationship between the input and output of each of the more than two preset training models, constructing input and output data pairs to obtain the first sample data under each training model.

[0009] In some embodiments, based on the second data group, two or more preset training models are used for training and testing to obtain an updated control parameter prediction model for the magnetic levitation compressor, including: when the displacement current dual closed-loop control system operates according to the initial control parameter prediction model, the operating parameters of the magnetic levitation compressor are obtained, recorded as compressor operating parameters; and the operating parameters of the unit to which the magnetic levitation compressor is located are obtained, recorded as unit operating parameters; the input and output of the initial control parameter prediction model, the compressor operating parameters and the unit operating parameters are recorded as the second data group; based on each of the two or more preset training models, the second data group is preprocessed to obtain second sample data under each training model; each training model is trained and tested using the second sample data under each training model to obtain an updated control parameter prediction model under each training model; thereby, an updated control parameter prediction model under each of the two or more preset training models is obtained; the updated control parameter prediction model under each of the two or more preset training models is weighted averaged to obtain a weighted average updated control parameter prediction model of the two or more preset training models as the updated control parameter prediction model for the magnetic levitation compressor.

[0010] In some embodiments, in the second data group, the compressor operating parameters include at least one of the following: the current of the motor in the compressor, the operating frequency of the motor in the compressor, and the bus voltage of the motor controller in the compressor; the unit where the magnetic levitation compressor is located includes an air-conditioning unit, and the unit operating parameters include at least one of the following: the shutdown or running state of the air-conditioning unit, the operating condition of the air-conditioning unit, the state of the cooling water flow switch in the air-conditioning unit, the state of the freezing water flow switch in the air-conditioning unit, the opening of the throttle valve in the air-conditioning unit, the load of the air-conditioning unit, the cold water inlet pressure in the air-conditioning unit, the freezing water outlet pressure in the air-conditioning unit, the current heat exchange amount of the air-conditioning unit, the condensing pressure of the condenser in the air-conditioning unit, and the evaporating pressure of the evaporator in the air-conditioning unit; based on each training model in more than two preset training models, the second data group is preprocessed , obtaining the second sample data under each training model, including: for the compressor operating parameters in the second data group, screening is performed according to the first corresponding setting parameter range, and the corresponding operating parameters within the first corresponding setting parameter range are obtained, which are recorded as the first screening operating parameters; for the unit operating parameters in the second data group, screening is performed according to the second corresponding setting parameter range, and the corresponding operating parameters within the second corresponding setting parameter range are obtained, which are recorded as the second screening operating parameters; the input and output of the initial control parameter prediction model, the first screening operating parameters and the second screening operating parameters are used as the screening data group; for the screening data group, based on the demand relationship between the input and output of each training model in more than two preset training models, a data pair of input and output is constructed to obtain the second sample data under each training model.

[0011] Matching the above method, the present invention provides a control device for a magnetic levitation compressor on the other hand. The magnetic levitation compressor has a magnetic bearing, and the control system of the magnetic bearing adopts a displacement current double closed-loop control system; the control device of the magnetic levitation compressor includes: a control unit, which is configured to introduce one or more preset disturbances in the displacement current double closed-loop control system when the magnetic levitation compressor is running; and adjust the control parameters of the displacement current double closed-loop control system more than once; an acquisition unit, which is configured to obtain the rotor displacement of the magnetic bearing and the stator coil current of the magnetic bearing for each disturbance among the one or more preset disturbances introduced, when the magnetic bearing is running based on the control parameters of the displacement current double closed-loop control system adjusted each time; thereby, the control parameters, rotor displacement and stator coil current of the displacement current double closed-loop control system corresponding to all disturbances among the one or more preset disturbances introduced are obtained, which are recorded as a first data group; the control unit , is also configured to perform training and testing based on the first data group using two or more preset training models to obtain an initial control parameter prediction model of the magnetic levitation compressor, so as to control the operation of the magnetic bearing according to the initial control parameter prediction model of the magnetic levitation compressor; the control unit is also configured to obtain the operating parameters of the magnetic levitation compressor when the displacement current dual closed-loop control system operates according to the initial control parameter prediction model, which are recorded as compressor operating parameters; and obtain the operating parameters of the unit where the magnetic levitation compressor is located, which are recorded as unit operating parameters; the input and output of the initial control parameter prediction model, the compressor operating parameters and the unit operating parameters are recorded as a second data group; the control unit is also configured to perform training and testing based on the second data group using two or more preset training models to obtain an updated control parameter prediction model of the magnetic levitation compressor, so as to control the operation of the magnetic bearing according to the updated control parameter prediction model of the magnetic levitation compressor.

[0012] In some embodiments, the control unit, when the magnetic levitation compressor is running, introduces one or more preset disturbances into the displacement current dual closed-loop control system, including: adding one or more interference sources to the feedback value of the rotor displacement of the displacement loop of the displacement current dual closed-loop control system to introduce one or more preset disturbances into the displacement current dual closed-loop control system; wherein the one or more interference sources include: at least one of Gaussian white noise of a set frequency, sinusoidal interference signals of different amplitudes, and random pulse signals of different amplitudes; and / or, more than two preset training models, including: at least two of an extreme learning machine model, a support vector machine model, and a random forest model.

[0013] In some embodiments, the control unit, based on the first data group, uses two or more preset training models for training and testing to obtain an initial control parameter prediction model of the magnetic levitation compressor, including: preprocessing the first data group based on each of the two or more preset training models to obtain first sample data under each training model; training and testing each training model using the first sample data under each training model to obtain a control parameter prediction model under each training model; thereby, obtaining a control parameter prediction model under each of the two or more preset training models; performing weighted averaging processing on the control parameter prediction model under each of the two or more preset training models to obtain a weighted average control parameter prediction model of the two or more preset training models as the initial control parameter prediction model of the magnetic levitation compressor.

[0014] In some embodiments, the control unit preprocesses the first data group based on each training model in more than two preset training models to obtain the first sample data under each training model, including: based on the first data group, determining the control parameters corresponding to the minimum rotor displacement and the minimum stator coil current under each disturbance in more than one preset disturbance introduced, as the optimal control parameters under each disturbance; thereby, obtaining the optimal control parameters, minimum rotor displacement and minimum stator coil current corresponding to all disturbances in more than one preset disturbance introduced, which are recorded as the initial data group; for the initial data group, constructing the input and output data pairs based on the demand relationship between the input and output of each training model in more than two preset training models to obtain the first sample data under each training model.

[0015] In some embodiments, the control unit, based on the second data group, uses two or more preset training models for training and testing to obtain an updated control parameter prediction model for the magnetic levitation compressor, including: when the displacement current dual closed-loop control system operates according to the initial control parameter prediction model, obtaining the operating parameters of the magnetic levitation compressor, recorded as compressor operating parameters; and obtaining the operating parameters of the unit to which the magnetic levitation compressor is located, recorded as unit operating parameters; recording the input and output of the initial control parameter prediction model, the compressor operating parameters and the unit operating parameters as the second data group; based on each of the two or more preset training models, preprocessing the second data group to obtain second sample data under each training model; using the second sample data under each training model to train and test each training model to obtain an updated control parameter prediction model under each training model; thereby, obtaining an updated control parameter prediction model under each of the two or more preset training models; performing weighted average processing on the updated control parameter prediction model under each of the two or more preset training models to obtain a weighted average updated control parameter prediction model of the two or more preset training models as the updated control parameter prediction model for the magnetic levitation compressor.

[0016] In some embodiments, in the second data group, the compressor operating parameters include at least one of the following: the current of the motor in the compressor, the operating frequency of the motor in the compressor, and the bus voltage of the motor controller in the compressor; the unit where the magnetic levitation compressor is located includes an air-conditioning unit, and the unit operating parameters include at least one of the following: the shutdown or running state of the air-conditioning unit, the operating condition of the air-conditioning unit, the state of the cooling water flow switch in the air-conditioning unit, the state of the freezing water flow switch in the air-conditioning unit, the opening of the throttle valve in the air-conditioning unit, the load of the air-conditioning unit, the cold water inlet pressure in the air-conditioning unit, the freezing water outlet pressure in the air-conditioning unit, the current heat exchange rate of the air-conditioning unit, the condensing pressure of the condenser in the air-conditioning unit, and the evaporating pressure of the evaporator in the air-conditioning unit; the control unit, based on each of more than two preset training models, performs Preprocessing is performed to obtain second sample data under each training model, including: for the compressor operating parameters in the second data group, screening is performed according to the first corresponding setting parameter range, and the corresponding operating parameters within the first corresponding setting parameter range are obtained, which are recorded as first screening operating parameters; for the unit operating parameters in the second data group, screening is performed according to the second corresponding setting parameter range, and the corresponding operating parameters within the second corresponding setting parameter range are obtained, which are recorded as second screening operating parameters; the input and output of the initial control parameter prediction model, the first screening operating parameters and the second screening operating parameters are used as a screening data group; for the screening data group, based on the demand relationship between the input and output of each training model in more than two preset training models, a data pair of input and output is constructed to obtain the second sample data under each training model.

[0017] Matching the above device, the present invention provides a magnetic levitation compressor on another aspect, including: the control device of the magnetic levitation compressor described above.

[0018] Matching the above method, the present invention provides a storage medium on another aspect, wherein the storage medium includes a stored program, wherein when the program is running, the device where the storage medium is located is controlled to execute the steps of the control method of the magnetic levitation compressor described above.

[0019] Matching the above method, the present invention provides a computer program product on another aspect, including a computer program, which implements the steps of the above-mentioned control method of the magnetic levitation compressor when executed by a processor.

[0020] Therefore, the scheme of the present invention obtains a control parameter prediction model of the magnetic levitation compressor by performing two model trainings on the magnetic bearing control system; wherein, in the first model training, the displacement feedback value of the displacement loop and the current feedback value of the current loop in the displacement current double closed-loop control system of the magnetic bearing control system are used as input, and the control parameters of the control mode (such as PID control mode) of the magnetic bearing control system are used as output, and various possible disturbances that may occur during the actual operation of the magnetic levitation compressor are added to the displacement current double closed-loop control system (such as adding disturbances to the displacement feedback value in the displacement loop), and the extreme learning machine model, the support vector machine model and the random forest model are used for training to obtain the initial control parameters of the magnetic levitation compressor. A number prediction model is provided; in the second model training, the input and output in the initial control parameter prediction model, the operating parameters of the magnetic levitation compressor, and the operating parameters of the unit where the magnetic levitation compressor is located (such as an air-conditioning unit) are used as input, and the output in the initial control parameter prediction model is used as output. The extreme learning machine model, the support vector machine model and the random forest model are used for training to obtain an updated control parameter prediction model of the magnetic levitation compressor as the control parameter prediction model of the required magnetic levitation compressor; thus, the control parameter prediction model of the magnetic levitation compressor is obtained by using the extreme learning machine model, the support vector machine model and the random forest model to predict, which is beneficial to improving the bearing control performance and ensuring the operating stability of the magnetic levitation compressor.

[0021] Other features and advantages of the present invention will be set forth in the description which follows, and in part will be apparent from the description, or may be learned by practice of the present invention.

[0022] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A schematic flow chart of an embodiment of a method for controlling a magnetic levitation compressor of the present invention;

[0024] Figure 2 A flowchart of an embodiment of the method of the present invention for training and testing using two or more preset training models based on the first data set;

[0025] Figure 3 is a flow chart of an embodiment of preprocessing the first data group in the method of the present invention;

[0026] Figure 4 A flowchart of an embodiment of the method of the present invention for training and testing using two or more preset training models based on the second data group;

[0027] Figure 5is a flow chart of an embodiment of preprocessing the second data group in the method of the present invention;

[0028] Figure 6 It is a structural schematic diagram of an embodiment of a control device for a magnetic levitation compressor of the present invention;

[0029] Figure 7 A schematic diagram of the working process of the control parameter prediction model of the magnetic levitation compressor;

[0030] Figure 8 This is a schematic diagram of the implementation process of the extreme learning machine model;

[0031] Fig. 9 This is a schematic diagram of the solution process of the random forest control parameter prediction model;

[0032] Fig.10 This is a schematic diagram of the regression solution process of the support vector machine model;

[0033] Fig.11 Schematic diagram of the process of solving problems for machine learning.

[0034] In conjunction with the accompanying drawings, the reference numerals in the embodiments of the present invention are as follows:

[0035] 102 - acquisition unit; 104 - control unit. DETAILED DESCRIPTION

[0036] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in combination with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the scheme of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0037] Considering that the control of active magnetic bearing-rotor system is difficult, the parameters of the bearing controller of active magnetic bearing-rotor system greatly affect the dynamic response speed, control bandwidth and stability of the bearing control system. When using PID, anti-disturbance control, model prediction and other control methods, it is difficult to adjust the control parameters. In the actual design process of the parameters, it is usually impossible to manually find the optimal control parameters, and it is difficult to ensure the optimal control performance.

[0038] Therefore, the scheme of the present invention proposes a control method for a magnetic levitation compressor, specifically a magnetic levitation compressor adaptive optimal parameter control method based on machine learning. Through optimization and model training, the optimal parameters within a certain parameter range are found, the control reliability is enhanced, the bearing control performance is improved, and the stability of the magnetic levitation compressor system is greatly improved.

[0039] According to an embodiment of the present invention, a control method for a magnetic levitation compressor is provided. Figure 1 The flow chart of an embodiment of the method of the present invention is shown in FIG. The magnetic levitation compressor has a magnetic bearing, and the control system of the magnetic bearing adopts a displacement current double closed loop control system, that is, a double closed loop control system in which the outer loop is a displacement loop and the inner loop is a current loop. In the scheme of the present invention, Figure 1 As shown, the control method of the magnetic levitation compressor includes: steps S110 to S150.

[0040] At step S110, when the magnetic levitation compressor is running, one or more preset disturbances are introduced into the displacement current dual closed-loop control system, such as simulating various disturbances that may occur during the actual operation of the magnetic levitation compressor, for example: introducing preset disturbance parameters into the displacement current dual closed-loop control system, there are more than one type of preset disturbance parameters, and there are more than one disturbance parameters for each of the preset disturbance parameters; and adjusting the control parameters of the displacement current dual closed-loop control system more than once, specifically, the control parameters of the displacement current dual closed-loop control system can be adjusted according to a set period, such as adjusting the PID control parameters of the displacement current dual closed-loop control system.

[0041] In some embodiments, in step S110, when the magnetic levitation compressor is running, one or more preset disturbances are introduced into the displacement current dual closed-loop control system, including: adding one or more interference sources to the feedback value of the rotor displacement of the displacement loop of the displacement current dual closed-loop control system to introduce one or more preset disturbances into the displacement current dual closed-loop control system; wherein the one or more interference sources include: at least one of Gaussian white noise of a set frequency, sinusoidal interference signals of different amplitudes, and random pulse signals of different amplitudes.

[0042] Figure 7The figure is a workflow diagram of the control parameter prediction model of the magnetic levitation compressor. In the actual process of designing the bearing parameters of the magnetic levitation compressor, it is usually impossible to manually find the optimal control parameters, and it is difficult to ensure the optimal control performance. Therefore, the scheme of the present invention provides an optimal control parameter predicted based on previous data. To this end, the present invention provides a method for obtaining initial control parameter prediction model training data, a method for installing the entire control parameter prediction model, and a structural device for optimizing the control parameter prediction model through actual operation data, such as using 5G communication as a real-time data transmission method, and the remote PC then trains the control parameter prediction model according to the specific training method of the control parameter prediction model given in the scheme of the present invention, which can specifically be the use of extreme learning machine, random forest and support vector machine weighted prediction control parameters.

[0043] like Figure 7 As shown in FIG. 1 , the workflow of the control parameter prediction model of the magnetic levitation compressor includes:

[0044] Step 11: Initial training data acquisition: On the experimental unit, when the magnetic bearing rotor is statically suspended, add high-frequency Gaussian white noise A to the bearing control loop. j +Sinusoidal interference of different amplitudes B j + Large random pulse signals of different amplitudes C j As the interference source, by determining a control mode (such as PID control mode), randomly changing the control parameters K1, K2, K3, the current value D is obtained. j And displacement accuracy E j The magnetic bearing control loop is a closed-loop system. When adding interference sources to the bearing control loop, they are added based on needs and convenience, such as adding them to the displacement sampling value.

[0045] In the scheme of the present invention, by introducing more than one preset disturbance into the displacement current dual closed-loop control system to simulate various possible disturbances that may occur during the actual operation of the magnetic levitation compressor, the optimal parameters under each disturbance are found in advance; thereby, all the data found in advance can be trained into an initial control parameter prediction model through machine learning, which is integrated into the magnetic bearing controller, and the optimal control parameters given by the control parameter prediction model are used under each operating condition or disturbance of the magnetic bearing compressor system, thereby improving the control effect and making the magnetic bearing rotor stably and suspended with high displacement accuracy.

[0046] At step S120, for each of the one or more preset disturbances introduced, the rotor displacement of the magnetic bearing is obtained, and the stator coil current of the magnetic bearing is obtained when the magnetic bearing operates based on the control parameters of the displacement current dual closed-loop control system adjusted each time; thereby, the control parameters, rotor displacement and stator coil current of the displacement current dual closed-loop control system corresponding to all disturbances among the one or more preset disturbances introduced are obtained and recorded as the first data group.

[0047] At step S130, based on the first data group, two or more preset training models are used for training and testing to obtain an initial control parameter prediction model of the magnetic levitation compressor, so as to control the operation of the magnetic bearing according to the initial control parameter prediction model of the magnetic levitation compressor.

[0048] In some embodiments, in step S130, based on the first data group, two or more preset training models are used for training and testing to obtain the specific process of the initial control parameter prediction model of the magnetic levitation compressor. Please refer to the following exemplary description.

[0049] Combine the following Figure 2 The flowchart of an embodiment of the method of the present invention for training and testing based on the first data group using more than two preset training models is shown, which further illustrates the specific process of training and testing based on the first data group using more than two preset training models in step S130, including: steps S210 to S230.

[0050] Step S210, for the first data group, preprocessing is performed based on each training model in more than two preset training models to obtain the first sample data under each training model; that is, based on each training model in more than two preset training models, the first data group is preprocessed to obtain the first sample data under each training model.

[0051] In some embodiments, in step S210, based on each of two or more preset training models, the first data group is preprocessed to obtain the specific process of the first sample data under each training model, see the following exemplary description.

[0052] Combine the following Figure 3 The flowchart of an embodiment of preprocessing the first data group in the method of the present invention further illustrates the specific process of preprocessing the first data group in step S210, including: step S310 to step S320.

[0053] Step S310, based on the first data group, determine the control parameters corresponding to the minimum rotor displacement and the minimum stator coil current under each disturbance among the one or more preset disturbances introduced, as the optimal control parameters under each disturbance; in this way, the optimal control parameters, minimum rotor displacement and minimum stator coil current corresponding to all disturbances among the one or more preset disturbances introduced are obtained, which are recorded as the initial data group.

[0054] Step S320, for the initial data group, based on the input and output requirement relationship of each training model in more than two preset training models, construct input and output data pairs to obtain the first sample data under each training model.

[0055] like Figure 7 As shown, the working process of the control parameter prediction model of the magnetic suspension compressor also includes: in step 11, by analyzing data, such as analyzing control parameters K1, K2, K3, obtaining the current value D j And displacement accuracy E j Under each disturbance, the smaller the displacement accuracy, the better, and the smaller the current, the better; that is, when the control current and displacement accuracy are the smallest, it is determined to be the optimal control parameter, and sufficient initial data is obtained through this method. Among them, [K1, K2, K3, D j 、E j ] is a set of data.

[0056] Step 12: Data preprocessing:

[0057] According to the input and output requirements of different types of prediction models, corresponding input-output data pairs are constructed, and the sorted input-output data pairs are divided into model training sets and final model test sets according to a certain ratio. For example: in the input-output data pair, the current value D j And displacement accuracy E j is the input data, and the control parameters K1, K2, and K3 are the output data; for a set of data [K1, K2, K3, D j 、E j ], if there are 1000 sets of data, set 800 sets as training data and the remaining 200 sets as test data.

[0058] ① Determine the control target: Determine the optimal control parameters to minimize the displacement accuracy of the magnetic bearing and achieve high-precision and stable operation of the bearing rotor.

[0059] ② Determine the magnetic bearing control parameters K1, K2, and K3 as output parameters, and other parameters as input parameters. Among them, the control parameters K1, K2, and K3 have different values ​​for different control modes.

[0060] ③ Perform parameter correlation analysis to determine the main input parameters, such as current value D j And displacement accuracy E j As input, it is used as the main input of the model of different prediction methods. Among them, through correlation analysis, we can know the correlation between input data and output data. The higher the correlation, the more suitable they are as a set of training data pairs.

[0061] In the scheme of the present invention, based on each of two or more preset training models, the first data group obtained by introducing disturbance is preprocessed to obtain the first sample data under each training model, and then based on the first sample data, training and testing are performed using two or more preset training models to more accurately obtain the initial control parameter prediction model of the magnetic levitation compressor, so as to find the optimal parameters under each disturbance in advance; then all the data found in advance can be trained into an initial control parameter prediction model through machine learning, and the initial control parameter prediction model can be integrated into the magnetic bearing controller, and the optimal control parameters given by the control parameter prediction model can be used under each working condition or disturbance of the magnetic bearing compressor system, thereby improving the control effect, so that the magnetic bearing rotor is stably and suspended with high displacement accuracy.

[0062] Step S220, using the first sample data under each training model to train and test each training model, to obtain a control parameter prediction model under each training model; thereby, a control parameter prediction model under each training model in more than two preset training models is obtained.

[0063] Step S230, performing weighted average processing on the control parameter prediction model under each training model in more than two preset training models, and obtaining a weighted average control parameter prediction model of more than two preset training models as the initial control parameter prediction model of the magnetic levitation compressor.

[0064] like Figure 7 As shown, the workflow of the control parameter prediction model of the magnetic levitation compressor also includes: in step 12, by performing correlation analysis, it can be known that the correlation between the input data and the output data is higher, indicating that the two are more suitable as a set of training data pairs; a total of N=4000 processed data are obtained, and the data are divided into 3800 training sets and 200 test sets. The training set is used for training, and the test set is used for testing. When the accuracy of the test result reaches the preset accuracy requirement, the training and testing are stopped to determine the initial control parameter prediction model of the magnetic levitation compressor.

[0065] exist Figure 7 In the example shown, in the magnetic bearing control loop, the rotor displacement X detected by the displacement sensor is fb(k), respectively input to the first comparator and the control parameter prediction module, the control parameter prediction module starts the control parameter prediction function under the control of the bearing host computer to change the control parameters K1, K2, K3; the reference displacement X ref (k) and rotor displacement X fb (k outputs the reference displacement X after passing through the first comparator ref (k) and rotor displacement X fb The difference between the displacement error and the control parameters K1, K2 and K3 is the displacement error e(k). The controller outputs the current adjustment parameter u(k) based on the displacement error e(k) and the control parameters K1, K2 and K3. The reference current I is obtained based on the current adjustment parameter u(k). ref (k), the current sensor detects the coil current I fb (k); reference current I ref (k) and coil current I fb (k) After the second comparator, the reference current I is obtained ref (k) and coil current I fb The difference between (k) and (k) is the current error. The current error is passed through the current controller to obtain a PWM signal. The PWM signal is passed through the power amplifier to obtain a control signal of the coil current for controlling the coil current. The control parameter prediction module and the controller constitute a control parameter prediction model. The control parameter prediction model predicts the controller to adjust the control parameters through machine learning.

[0066] In the scheme of the present invention, based on the first data group obtained by introducing disturbance, training and testing are performed using more than two preset training models to obtain an initial control parameter prediction model of the magnetic levitation compressor, so as to find the optimal parameters under each disturbance in advance; then all the data found in advance can be trained into an initial control parameter prediction model through machine learning, which is integrated into the magnetic bearing controller, and the optimal control parameters given by the control parameter prediction model are used under each operating condition or disturbance of the magnetic bearing compressor system, thereby improving the control effect and making the magnetic bearing rotor stably and suspended with high displacement accuracy.

[0067] At step S140, when the displacement current dual closed-loop control system operates according to the initial control parameter prediction model, the operating parameters of the magnetic levitation compressor are obtained and recorded as compressor operating parameters; and the operating parameters of the unit in which the magnetic levitation compressor is located are obtained and recorded as unit operating parameters; the input and output of the initial control parameter prediction model, the compressor operating parameters and the unit operating parameters are recorded as a second data group.

[0068] At step S150, based on the second data group, two or more preset training models are used for training and testing to obtain an updated control parameter prediction model of the magnetic levitation compressor, so as to control the operation of the magnetic bearing according to the updated control parameter prediction model of the magnetic levitation compressor.

[0069] Bearing control involves control algorithms, and control parameters need to be adjusted to ensure excellent bearing control performance. It is difficult to adjust the optimal parameters manually, and the optimal control within a certain parameter range cannot be guaranteed. The scheme of adaptive optimal parameter control of magnetic bearings provided by the scheme of the present invention simulates various possible disturbances in the actual operation of the magnetic levitation compressor, and finds the optimal parameters under each disturbance in advance; all the data found in advance are trained by machine learning to become an initial control parameter prediction model, which is integrated into the magnetic bearing controller, and the optimal control parameters given by the control parameter prediction model are used under each working condition or disturbance of the magnetic bearing compressor system, thereby improving the control effect, so that the magnetic bearing rotor is stable and suspended with high displacement accuracy. At the same time, the displacement accuracy, impact force, control parameters used, etc. of each magnetic bearing compressor system in the actual operating conditions on site are obtained through the fifth generation mobile communication technology (5th Generation Mobile Communication Technology, referred to as 5G) module, and the data set is fed back to the aforementioned control parameter prediction model, and the control parameter prediction model is further trained to achieve optimal control.

[0070] In some embodiments, in step S150, based on the second data group, two or more preset training models are used for training and testing to obtain a specific process of updating the control parameter prediction model of the magnetic levitation compressor, as shown in the following exemplary description.

[0071] Combine the following Figure 4 The flowchart of an embodiment of the method of the present invention for training and testing based on the second data group using more than two preset training models is shown, which further illustrates the specific process of training and testing based on the second data group using more than two preset training models in step S150, including: step S410 to step S440.

[0072] Step S410, when the displacement current dual closed-loop control system operates according to the initial control parameter prediction model, the operating parameters of the magnetic levitation compressor are obtained, recorded as compressor operating parameters; and the operating parameters of the unit where the magnetic levitation compressor is located are obtained, recorded as unit operating parameters; the input and output of the initial control parameter prediction model, the compressor operating parameters and the unit operating parameters are recorded as a second data group.

[0073] Step S420, for the second data group, preprocessing is performed based on each training model in more than two preset training models to obtain second sample data under each training model; that is, based on each training model in more than two preset training models, the second data group is preprocessed to obtain second sample data under each training model.

[0074] In some embodiments, in the second data group, the compressor operating parameters include at least one of the following: the current of the motor in the compressor, the operating frequency of the motor in the compressor, and the bus voltage of the motor controller in the compressor; the unit where the magnetic levitation compressor is located includes an air-conditioning unit, and the unit operating parameters include at least one of the following: the shutdown or running state of the air-conditioning unit, the operating conditions of the air-conditioning unit, the state of the cooling water flow switch in the air-conditioning unit, the state of the freezing water flow switch in the air-conditioning unit, the opening of the throttle valve in the air-conditioning unit, the load of the air-conditioning unit, the cold water inlet pressure in the air-conditioning unit, the freezing water outlet pressure in the air-conditioning unit, the current heat exchange capacity of the air-conditioning unit, the condensing pressure of the condenser in the air-conditioning unit, and the evaporating pressure of the evaporator in the air-conditioning unit.

[0075] In step S420, based on each of the two or more preset training models, the second data group is preprocessed to obtain the specific process of the second sample data under each training model. Please refer to the following exemplary description.

[0076] Combine the following Figure 5 The flowchart of an embodiment of preprocessing the second data group in the method of the present invention further illustrates the specific process of preprocessing the second data group in step S420, including: steps S510 to S540.

[0077] Step S510, filtering the compressor operating parameters in the second data group according to the first corresponding setting parameter range, and obtaining corresponding operating parameters within the first corresponding setting parameter range, which are recorded as first filtered operating parameters.

[0078] Step S520, filtering the unit operating parameters in the second data group according to a second corresponding setting parameter range, and obtaining corresponding operating parameters within the second corresponding setting parameter range, which are recorded as second filtered operating parameters.

[0079] Step S530: taking the input and output of the initial control parameter prediction model, the first screening operation parameter and the second screening operation parameter as a screening data group.

[0080] Step S540, for the screening data group, based on the input and output requirement relationship of each training model in more than two preset training models, construct input and output data pairs to obtain second sample data under each training model.

[0081] To achieve the best effect of the control parameter prediction model, such as Figure 7 As shown, the current operating data of the unit system is obtained in real time through the 5G module of the unit system controller through the remote PC host computer, where each data includes the current state of the unit, the operating condition of the unit, the cooling water flow switch state, the chilled water flow switch state, the throttle valve conduction opening, the unit load, the chilled water inlet pressure, the chilled water outlet pressure, the current cooling capacity, the condensing pressure, the evaporating pressure, the magnetic levitation compressor current, the magnetic levitation compressor motor operating frequency, the magnetic levitation controller bus voltage, the magnetic levitation bearing displacement accuracy, the magnetic levitation bearing current, the magnetic bearing control parameters K1, K2, K3, a total of 19 feature categories.

[0082] Among them, the current state of the unit refers to the state of the unit being shut down or running, and the operating condition of the unit refers to the cooling water flow switch state, the chilled water flow switch state, the throttle valve conduction opening, the unit load, the chilled water inlet pressure, the chilled water outlet pressure, the current cooling capacity, the condensing pressure, the evaporating pressure, the magnetic levitation compressor current, the magnetic levitation compressor motor operating frequency, the magnetic levitation controller bus voltage, the magnetic levitation bearing displacement accuracy, the magnetic levitation bearing current, and the set values ​​of the magnetic bearing control parameters K1, K2, and K3 under each condition.

[0083] The data of the magnetic bearing unit during operation is collected, and each piece of data contains the 19 feature categories of the above unit operation data. The obtained data is preprocessed as mentioned above, for example: abnormal data is deleted, including data that is obviously incorrect and inconsistent with reality, data with artificially found data records, and data with some missing information.

[0084] In the scheme of the present invention, based on each of more than two preset training models, the second data group obtained based on the initial control parameter model is preprocessed so that the second sample data obtained under each training model is more accurate, and then, the control parameter prediction model is further trained to achieve optimal control.

[0085] Step S430, using the second sample data under each training model to train and test each training model, to obtain an updated control parameter prediction model under each training model; thereby, an updated control parameter prediction model under each training model in more than two preset training models is obtained.

[0086] Step S440, performing weighted averaging processing on the updated control parameter prediction model under each training model in more than two preset training models, and obtaining a weighted average updated control parameter prediction model of more than two preset training models as the updated control parameter prediction model of the magnetic levitation compressor.

[0087] See also Figure 7 In the example shown, the main controller of the magnetic levitation compressor system communicates with the magnetic levitation compressor system through the 5G communication module: receiving the real-time operating current and displacement accuracy data of the magnetic levitation bearing transmitted by the magnetic levitation compressor system, sending the current operating conditions and pressure difference of the compressor system to the bearing controller, and updating the latest control parameter prediction model to the main control chip of the bearing. The main controller of the magnetic levitation compressor system interacts with the remote PC host computer (including the unit operation data monitoring system and the control parameter prediction model training system): the remote PC host computer transmits the latest control parameter prediction model to the main controller of the magnetic levitation compressor system, and the main controller of the magnetic levitation compressor system feeds back the pressure difference, working conditions and received magnetic bearing data of the magnetic levitation compressor system to the remote PC host computer.

[0088] See also Figure 7 In the example shown, the training architectures of the three models are updated simultaneously while preprocessing the acquired data, such as re-introducing the newly acquired data into the model for training:

[0089] a. The number of nodes in the hidden layer of the extreme learning machine is 10.

[0090] b. The number of random features of random forest is mtry=4.

[0091] c. The support vector machine inputs 16 values ​​and outputs 3 values. Among them, the input 16 values, such as: the current state of the unit, the cooling water flow switch state under each working condition, the chilled water flow switch state, the throttle valve guide vane opening, the unit load, the chilled water inlet pressure, the chilled water outlet pressure, the current cooling capacity, the condensing pressure, the evaporation pressure, the magnetic levitation compressor current, the magnetic levitation compressor motor operating frequency, the magnetic levitation controller bus voltage, the magnetic levitation bearing displacement accuracy, the magnetic levitation bearing current; output 3 values, such as: magnetic bearing control parameters K1, K2, K3.

[0092] The data is distinguished into training set and test set, and the previous generation of control parameter prediction model is further optimized and iterated. This cycle is repeated until each magnetic levitation compressor can achieve the expected operating effect under different working conditions, and then the iteration is stopped. When the components of the unit are damaged and the parameters need to be re-matched, the control parameter prediction model can be iterated again.

[0093] The optimal parameter adaptive control scheme and control parameter training device of the magnetic levitation compressor constructed by the scheme of the present invention finds the optimal parameters within a certain parameter range through optimization and model training, enhances the reliability of control, improves the bearing control performance, and greatly improves the stability of the magnetic levitation compressor system. The magnetic bearing control parameter prediction method provided by the scheme of the present invention can be used not only in the magnetic levitation bearing system, but also in other control occasions.

[0094] In some embodiments, the two or more preset training models include: at least two of an extreme learning machine model, a support vector machine model, and a random forest model. Of course, preferably, the two or more preset training models include: an extreme learning machine model, a support vector machine model, and a random forest model.

[0095] Figure 8 This is a schematic diagram of the implementation flow of the extreme learning machine model structure. Figure 8 As shown in Figure 1, the implementation process of the extreme learning machine model structure is as follows: Extreme Learning Machine (ELM) is also a neural network, which belongs to a feedforward model with a single hidden layer. ELM randomly initializes the weights of the input layer and the hidden layer and the bias of the nodes, and finally calculates the weights of the hidden layer and the output layer.

[0096] Assume there are n sample points (x j ,t j ), then the extreme learning machine model can be written as follows:

[0097]

[0098] Among them, g(x) is the activation function, W i =[W i,1 ,W i,2 ,..,W i,n ] T is the input weight, β i is the weight of the output layer (beta), b i is the bias of the i-th hidden node. i *X j W i and X j The inner product of O. j Indicates output, X j represents input, and N represents the number of neural network layers.

[0099] The learning goal of the extreme learning machine is to minimize the error between the predicted value and the actual value of the output:

[0100]

[0101] Among them, t j Indicates actual value.

[0102] That is, there is β i , W i and b i , so that:

[0103]

[0104] It can be expressed as a matrix:

[0105] H*β=T (4).

[0106] Among them, H is the output of the hidden layer node, β is the weight of the connection between the hidden layer and the output layer, and T is the output label in the training data pair, which is specifically expressed as follows:

[0107]

[0108]

[0109] In order to be able to train a single hidden layer neural network, we want to get b i and β i , so that:

[0110]

[0111] Among them, H represents the output of the hidden layer node, β represents the weight of the connection between the hidden layer and the output layer, T represents the output label in the training data pair, and b i represents the bias of the i-th hidden node, W i =[W i,1 ,W i,2 ,..,W i,n ] T represents the input weight. i=1,...,L, which is equivalent to minimizing the loss function E:

[0112]

[0113] In the ELM algorithm, when we randomly determine the input weights of the input layer and the hidden layer, W i and bias, b i , the output matrix H of the hidden layer can be uniquely determined. The solution to the single hidden layer neural network model can be transformed into solving a linear equation:

[0114] H*β=T (8).

[0115] Among them, formula (8) is the same as formula (4).

[0116] And the output weight ③β can be determined:

[0117]

[0118] where H + is the generalized inverse of the matrix H, and it can be proved that the solution has the minimum and unique norm; β represents the weight connecting the hidden layer and the output layer, and T represents the output label in the training data pair.

[0119] The implementation steps of the ELM algorithm are as follows:

[0120] (1) Select the number of neurons in the hidden layer. The number of neurons is calculated according to the formula where m is the number of neurons in the hidden layer, n is the number of input nodes, l is the number of output nodes, and α is a constant between 1 and 10.

[0121] In the solution of the present invention, the number of nodes in the input layer of the initial control parameter prediction model is 5, and the number of nodes in the output layer is 3. After repeated adjustment, the number of nodes in the hidden layer is finally selected as 8.

[0122] (2) Randomly set the connection weights weight (0 < weight < 1) between the input layer and the hidden layer and the biases bias (0 < bias < 1) of the neurons.

[0123] (3) Select an activation function. Due to the good performance of the sigmoid function, the function H = 1 / (1 + exp(-H)) is selected in the solution of the present invention. Calculate the matrix H of the hidden layer output layer, and calculate the output layer weight β using the formula.

[0124] Through multiple calculations and tests, the number of neurons is finally selected as 8. Import 3800 groups of processed data to solve the prediction model of the extreme learning machine.

[0125] Set the model parameters well and test with 200 groups of test data. If the accuracy rate reaches 90%, integrate it into the upper computer.

[0126] Fig. 9 It is a schematic flow chart for solving the random forest control parameter prediction model. As Fig. 9 shown, the implementation steps of the random forest regression prediction model structure are as follows:

[0127] ① There are ntree trees in the random forest, and each regression tree corresponds to a training set. To construct ntree regression trees, a corresponding number of training sets need to be generated. In the generation process of the random forest algorithm, bagging sampling technology is mainly used to generate ntree training subsets from the original training set. Each sampling is random and replaced.

[0128] ②Then train a decision tree base learner for each new data set in turn according to the random attribute selection method. Assuming that the number of independent variables in the original sample is p, a subset of mtry independent variables is randomly selected from the independent variable set of each decision tree, and then an optimal independent variable is selected as the partition attribute. In the random forest regression algorithm, mtry = log 2 p.

[0129] ③ Thus, ntree regression decision trees are obtained. Each tree starts to recursively generate child nodes from top to bottom, and determines when the decision tree will terminate the split by setting the minimum threshold of the number of samples contained in the leaf node (the regression tree can be 5).

[0130] ④ From ntree decision trees, ntree prediction values ​​can be obtained. The weighted average of ntree prediction values ​​is taken using the ensemble learning idea as the regression prediction result, which will be the final output result of the algorithm.

[0131] The regression tree mathematical model is:

[0132]

[0133] Where: M is the number of samples in the subset of the regression tree model; c m is the mean response value of the data samples in each subset; R m are the subsets divided; I(x∈Rm) is the characteristic function, when x∈R m The value is 1 when it is set, otherwise the value is 0.

[0134] From the original training sample set of 3800 random forests (RF), 3800 training samples were randomly selected with replacement to obtain 1 sub-sample set, denoted as M=3800, and one regression tree was constructed for one sample; the experiment was repeated 300 times to obtain 500 sub-sample sets (the 500 sub-sample sets were independent of each other and the samples could be repeated), and 500 regression decision trees were constructed, denoted as ntree=300.

[0135] Since the selected independent variable of the initial control parameter prediction model in the solution of the present invention is p=5, a regression tree learner is trained by randomly selecting feature numbers. Generally, the random feature number mtry=log 2p, so in the solution of the present invention, mtry=2. 500 regression decision trees are trained, and finally the results of the 500 regression subtrees are averaged as the prediction result.

[0136] The accuracy of the established random forest regression model is verified by using the remaining 200 groups of validation samples. If the accuracy rate exceeds 90%, the random forest model is integrated into the host computer.

[0137] Fig.10 Schematic diagram of the process of solving support vector machine regression. Fig.10 In the example shown, Support Vector Machine (SVM) is considered to be a very beautiful, practical and powerful learning algorithm. SVM is a classification algorithm. The basic model is a linear classifier with the maximum margin in the feature space, that is, maximizing the minimum margin, and finally transforming it into a convex optimization problem to solve. Simply put, it is to find a segmentation surface, that is, a hyperplane, in an n-dimensional space to correctly classify the feature points in the space. The distance of the points on both sides of the hyperplane from the hyperplane also indicates the accuracy of the segmentation. The points closest to the hyperplane on both sides are called support vectors (SV), and the distance between these points and the hyperplane is called the margin. SVM maximizes this margin. In other words, the segmentation problem can essentially be converted into a regression problem.

[0138] like Fig.10 As shown in Figure 1, the implementation steps of the support vector machine structure are as follows: Support Vector Machine (SVM) is a binary classification model that maps the feature vector of an instance (taking two dimensions as an example) to some points in space, such as Fig.10 The red and black points in the figure belong to two different categories. The purpose of SVM is to draw a line that "best" distinguishes the two categories of points, so that if new points appear in the future, this line can also make a good classification. There are countless lines that can be drawn, and the difference lies in the effect. Each line can be called a dividing hyperplane. The best line we hope to find is the "partitioning hyperplane with the largest margin". From Fig.10 It can be seen that the distances from the points on the dotted line to the dividing hyperplane are the same. In fact, only these points jointly determine the position of the hyperplane, so they are called "support vectors."

[0139] The partitioning hyperplane can be defined as a linear equation:

[0140] ω T x+b=y;

[0141] Where ω={ω 1 ;ω2 ;...;ω d} is a normal vector that determines the direction of the hyperplane; d is the number of eigenvalues; x is the training sample; b is the displacement term that determines the distance between the hyperplane and the origin; as long as the normal vector ω and the displacement b are determined, a partitioning hyperplane can be uniquely determined. x represents the input of the training sample, y represents the output of the training sample, and the superscript T represents the transposition symbol.

[0142] Since the SVM regression model (SVR) is used in the solution of the present invention, the hyperplane decision boundary obtained when the support vector is used for classification is the regression model when the support vector is used for regression:

[0143] y i =ω T x i +b (11).

[0144] The distance between any point on the partitioning hyperplane and the marginal hyperplanes on both sides is So the minimum margin is

[0145] The constraint condition in SVR regression is to make the points in each training set as concentrated as possible near the support vector; in SVR, in order to obtain the distance from all feature sample points to the hyperplane, the sample points closest to the hyperplane are found at the same time.

[0146] The function interval is defined: for the selected training data set T and hyperplane (ω, b), the hyperplane and the feature sample point (x i ,y i ) is:

[0147] r i =y i (ω T x i +b) (12).

[0148] In order to find the support vector, we can calculate the minimum distance r of all sample points. min The one that:

[0149] r min =min(r) (13).

[0150] There is a serious problem with the function interval, that is, when ω and b increase or decrease at a given ratio, the hyperplane equation does not change, but the function interval becomes a multiple of the original ratio. In order to avoid this situation, a concept of geometric interval is introduced, namely:

[0151] r i =y i ((ωT / ||ω||)x i +b / ||ω||) (14).

[0152] This ensures that if ω and b change proportionally, the hyperplane will also change according to the given proportional multiple. Since the learning purpose of SVM is to maximize the corresponding geometric interval, the goal is to maximize the geometric interval, and the expression is as follows:

[0153]

[0154] In terms of optimization objective function, SVR and SVM are consistent. However, under the constraint condition, SVM makes each point in the training set as far away from the support vector of its own category (sign) as possible; for the regression model, our indicator is to make each feature sample point (x i ,y i ), try to fit a linear model as much as possible:

[0155] f(X)=ω T X+b.

[0156] In ε-SVR, if the sample point is close enough to the regression model and falls within the interval of the regression model, the loss of the sample point will not be calculated. The loss function is defined as follows:

[0157]

[0158] Therefore, the objective function of SVR can be expressed as follows:

[0159]

[0160] By introducing slack variables and Lagrange multipliers, the problem is transformed into a dual problem. The coefficient ω of the support vector machine model is solved using the Karush-Kuhn-Tucker conditions (KKT) and the sequential minimal optimization algorithm (SMO). T , b, determine the input 5 values ​​of the support vector machine and output 3 values. Input 5 values, such as: unit load, magnetic levitation compressor current, magnetic levitation compressor motor operating frequency, magnetic levitation bearing displacement accuracy, magnetic levitation bearing current; output 3 values, such as: magnetic bearing control parameters K1, K2, K3.

[0161] The 3800 groups of original training sample sets are divided to train the SVR model, and the remaining 200 groups are used as validation samples. If the validation accuracy reaches 90%, they are integrated into the host computer as a regression prediction model.

[0162] Fig.11 A flow chart of the process of solving machine learning problems. Fig.11 As shown in the figure, the schematic diagram of heterogeneous (i.e. different types of training models) integrated learning solution integrates the extreme learning machine, support vector machine and random forest regressor established above into the host computer data monitoring system. Input the real-time data into the three models trained previously to obtain three prediction results. According to the prediction accuracy of the previous model, the weights are assigned and the average is taken as the control parameter prediction output; among them, the weighted average prediction has better effect. For example: 0.1*prediction value of extreme learning machine + 0.5*prediction value of support vector machine + 0.4*prediction value of random forest regressor = final output value.

[0163] In the scheme of the present invention, at least two, preferably three, of the extreme learning machine model, the support vector machine model and the random forest model are used for training and testing, and the three prediction results corresponding to the three models trained previously are input with real-time data and weighted averaged to be used as the control parameter prediction output, so that the training is more comprehensive, thereby making the output more accurate and stable.

[0164] By adopting the technical solution of this embodiment, the control parameter prediction model of the magnetic levitation compressor is obtained by performing two model trainings on the magnetic bearing control system; wherein, in the first model training, the displacement feedback value of the displacement loop and the current feedback value of the current loop in the displacement current double closed-loop control system of the magnetic bearing control system are used as input, and the control parameters of the control mode of the magnetic bearing control system (such as the PID control mode) are used as output, and various possible disturbances that may occur during the actual operation of the magnetic levitation compressor are added to the displacement current double closed-loop control system (such as adding disturbances to the displacement feedback value in the displacement loop), and the extreme learning machine model, the support vector machine model and the random forest model are used for training to obtain the initial control of the magnetic levitation compressor. parameter prediction model; in the second model training, the input and output in the initial control parameter prediction model, the operating parameters of the magnetic levitation compressor, and the operating parameters of the unit where the magnetic levitation compressor is located (such as an air-conditioning unit) are used as input, and the output in the initial control parameter prediction model is used as output, and the extreme learning machine model, support vector machine model and random forest model are used for training to obtain an updated control parameter prediction model of the magnetic levitation compressor as the control parameter prediction model of the required magnetic levitation compressor; thus, the control parameter prediction model of the magnetic levitation compressor is obtained by using the extreme learning machine model, support vector machine model and random forest model to predict, which is beneficial to improving the bearing control performance and ensuring the operating stability of the magnetic levitation compressor.

[0165] According to an embodiment of the present invention, a control device for a magnetic levitation compressor corresponding to the control method for the magnetic levitation compressor is also provided. Figure 6The structure diagram of an embodiment of the device of the present invention is shown in FIG. The magnetic levitation compressor has a magnetic bearing, and the control system of the magnetic bearing adopts a displacement current double closed loop control system, that is, a double closed loop control system in which the outer loop is a displacement loop and the inner loop is a current loop. In the solution of the present invention, Figure 6 As shown, the control device of the magnetic levitation compressor includes: an acquisition unit 102 and a control unit 104.

[0166] Wherein, the control unit 104 is configured to introduce one or more preset disturbances into the displacement current double closed-loop control system when the magnetic levitation compressor is running, such as simulating various possible disturbances that may occur during the actual operation of the magnetic levitation compressor, for example: introducing preset disturbance parameters into the displacement current double closed-loop control system, the types of preset disturbance parameters are more than one, and the number of each type of disturbance parameters in the preset disturbance parameters is more than one; and adjusting the control parameters of the displacement current double closed-loop control system more than once, specifically, the control parameters of the displacement current double closed-loop control system can be adjusted according to a set period, such as adjusting the PID control parameters of the displacement current double closed-loop control system. The specific functions and processing of the control unit 104 refer to step S110.

[0167] In some embodiments, the control unit 104 introduces one or more preset disturbances into the displacement current dual closed-loop control system when the magnetic levitation compressor is running, including: the control unit 104 is specifically configured to add one or more interference sources to the feedback value of the rotor displacement of the displacement loop of the displacement current dual closed-loop control system to achieve the introduction of one or more preset disturbances into the displacement current dual closed-loop control system; wherein the one or more interference sources include: at least one of Gaussian white noise of a set frequency, sinusoidal interference signals of different amplitudes, and random pulse signals of different amplitudes.

[0168] Figure 7 The figure is a workflow diagram of the control parameter prediction model of the magnetic levitation compressor. In the actual process of designing the bearing parameters of the magnetic levitation compressor, it is usually impossible to manually find the optimal control parameters, and it is difficult to ensure the optimal control performance. Therefore, the scheme of the present invention provides an optimal control parameter predicted based on previous data. To this end, the present invention provides a method for obtaining initial control parameter prediction model training data, a method for installing the entire control parameter prediction model, and a structural device for optimizing the control parameter prediction model through actual operation data, such as using 5G communication as a real-time data transmission method, and the remote PC then trains the control parameter prediction model according to the specific training method of the control parameter prediction model given in the scheme of the present invention, which can specifically be the use of extreme learning machine, random forest and support vector machine weighted prediction control parameters.

[0169] like Figure 7 As shown in FIG. 1 , the workflow of the control parameter prediction model of the magnetic levitation compressor includes:

[0170] Step 11: Initial training data acquisition: On the experimental unit, when the magnetic bearing rotor is statically suspended, add high-frequency Gaussian white noise A to the bearing control loop. j +Sinusoidal interference of different amplitudes B j + Large random pulse signals of different amplitudes C j As the interference source, by determining a control mode (such as PID control mode), randomly changing the control parameters K1, K2, K3, the current value D is obtained. j And displacement accuracy E j The magnetic bearing control loop is a closed-loop system. When adding interference sources to the bearing control loop, they are added based on needs and convenience, such as adding them to the displacement sampling value.

[0171] In the scheme of the present invention, by introducing more than one preset disturbance into the displacement current dual closed-loop control system to simulate various possible disturbances that may occur during the actual operation of the magnetic levitation compressor, the optimal parameters under each disturbance are found in advance; thereby, all the data found in advance can be trained into an initial control parameter prediction model through machine learning, which is integrated into the magnetic bearing controller, and the optimal control parameters given by the control parameter prediction model are used under each operating condition or disturbance of the magnetic bearing compressor system, thereby improving the control effect and making the magnetic bearing rotor stably and suspended with high displacement accuracy.

[0172] The acquisition unit 102 is configured to acquire the rotor displacement of the magnetic bearing and the stator coil current of the magnetic bearing for each disturbance in the one or more preset disturbances introduced, when the magnetic bearing operates based on the control parameters of the displacement current double closed-loop control system adjusted each time; thereby, the control parameters, rotor displacement and stator coil current of the displacement current double closed-loop control system corresponding to all disturbances in the one or more preset disturbances introduced are obtained, which are recorded as the first data group. The specific functions and processing of the acquisition unit 102 refer to step S120.

[0173] The control unit 104 is further configured to perform training and testing based on the first data set using two or more preset training models to obtain an initial control parameter prediction model of the magnetic levitation compressor, so as to control the operation of the magnetic bearing according to the initial control parameter prediction model of the magnetic levitation compressor. The specific functions and processing of the control unit 104 are also shown in step S130.

[0174] In some implementations, the control unit 104, based on the first data set, uses two or more preset training models for training and testing to obtain an initial control parameter prediction model for the magnetic levitation compressor, including:

[0175] The control unit 104 is further configured to perform preprocessing on the first data group based on each of the two or more preset training models to obtain first sample data under each training model; that is, based on each of the two or more preset training models, the first data group is preprocessed to obtain first sample data under each training model. The specific functions and processing of the control unit 104 are also shown in step S210.

[0176] In some implementations, the control unit 104 preprocesses the first data group based on each of the two or more preset training models to obtain first sample data under each training model, including:

[0177] The control unit 104 is further configured to determine, based on the first data group, the control parameters corresponding to the minimum rotor displacement and the minimum stator coil current under each of the one or more preset disturbances introduced, as the optimal control parameters under each disturbance; thereby, the optimal control parameters, minimum rotor displacement and minimum stator coil current corresponding to all disturbances in the one or more preset disturbances introduced are obtained, which are recorded as the initial data group. The specific functions and processing of the control unit 104 are also referred to step S310.

[0178] The control unit 104 is further configured to construct an input and output data pair for the initial data set based on the input and output requirement relationship of each training model in the two or more preset training models, and obtain the first sample data under each training model. The specific functions and processing of the control unit 104 are also shown in step S320.

[0179] like Figure 7 As shown, the working process of the control parameter prediction model of the magnetic suspension compressor also includes: in step 11, by analyzing data, such as analyzing control parameters K1, K2, K3, obtaining the current value D j And displacement accuracy E j Under each disturbance, the smaller the displacement accuracy, the better, and the smaller the current, the better; that is, when the control current and displacement accuracy are the smallest, it is determined to be the optimal control parameter, and sufficient initial data is obtained through this method. Among them, [K1, K2, K3, D j 、E j ] is a set of data.

[0180] Step 12: Data preprocessing:

[0181] According to the input and output requirements of different types of prediction models, corresponding input-output data pairs are constructed, and the sorted input-output data pairs are divided into model training sets and final model test sets according to a certain ratio. For example: in the input-output data pair, the current value D j And displacement accuracy E j is the input data, and the control parameters K1, K2, and K3 are the output data; for a set of data [K1, K2, K3, D j 、E j ], if there are 1000 sets of data, set 800 sets as training data and the remaining 200 sets as test data.

[0182] ① Determine the control target: Determine the optimal control parameters to minimize the displacement accuracy of the magnetic bearing and achieve high-precision and stable operation of the bearing rotor.

[0183] ② Determine the magnetic bearing control parameters K1, K2, and K3 as output parameters, and other parameters as input parameters. Among them, the control parameters K1, K2, and K3 have different values ​​for different control modes.

[0184] ③ Perform parameter correlation analysis to determine the main input parameters, such as current value D j And displacement accuracy E j As input, it is used as the main input of the model of different prediction methods. Among them, through correlation analysis, we can know the correlation between input data and output data. The higher the correlation, the more suitable they are as a set of training data pairs.

[0185] In the scheme of the present invention, based on each of two or more preset training models, the first data group obtained by introducing disturbance is preprocessed to obtain the first sample data under each training model, and then based on the first sample data, training and testing are performed using two or more preset training models to more accurately obtain the initial control parameter prediction model of the magnetic levitation compressor, so as to find the optimal parameters under each disturbance in advance; then all the data found in advance can be trained into an initial control parameter prediction model through machine learning, and the initial control parameter prediction model can be integrated into the magnetic bearing controller, and the optimal control parameters given by the control parameter prediction model can be used under each working condition or disturbance of the magnetic bearing compressor system, thereby improving the control effect, so that the magnetic bearing rotor is stably and suspended with high displacement accuracy.

[0186] The control unit 104 is further configured to train and test each training model using the first sample data under each training model to obtain a control parameter prediction model under each training model; thereby, a control parameter prediction model under each training model in more than two preset training models is obtained. The specific functions and processing of the control unit 104 are also referred to step S220.

[0187] The control unit 104 is further configured to perform weighted average processing on the control parameter prediction model under each of the two or more preset training models to obtain a weighted average control parameter prediction model of the two or more preset training models as the initial control parameter prediction model of the magnetic levitation compressor. The specific functions and processing of the control unit 104 are also shown in step S230.

[0188] like Figure 7 As shown, the workflow of the control parameter prediction model of the magnetic levitation compressor also includes: in step 12, by performing correlation analysis, it can be known that the correlation between the input data and the output data is higher, indicating that the two are more suitable as a set of training data pairs; a total of N=4000 processed data are obtained, and the data are divided into 3800 training sets and 200 test sets. The training set is used for training, and the test set is used for testing. When the accuracy of the test result reaches the preset accuracy requirement, the training and testing are stopped to determine the initial control parameter prediction model of the magnetic levitation compressor.

[0189] exist Figure 7 In the example shown, in the magnetic bearing control loop, the rotor displacement X detected by the displacement sensor is fb (k), respectively input to the first comparator and the control parameter prediction module, the control parameter prediction module starts the control parameter prediction function under the control of the bearing host computer to change the control parameters K1, K2, K3; the reference displacement X ref (k) and rotor displacement X fb (k outputs the reference displacement X after passing through the first comparator ref (k) and rotor displacement X fb The difference between the displacement error and the control parameters K1, K2 and K3 is the displacement error e(k). The controller outputs the current adjustment parameter u(k) based on the displacement error e(k) and the control parameters K1, K2 and K3. The reference current I is obtained based on the current adjustment parameter u(k). ref (k), the current sensor detects the coil current I fb (k); reference current I ref (k) and coil current I fb (k) After the second comparator, the reference current I is obtained ref (k) and coil current I fbThe difference between (k) and (k) is the current error. The current error is passed through the current controller to obtain a PWM signal. The PWM signal is passed through the power amplifier to obtain a control signal of the coil current for controlling the coil current. The control parameter prediction module and the controller constitute a control parameter prediction model. The control parameter prediction model predicts the controller to adjust the control parameters through machine learning.

[0190] In the scheme of the present invention, based on the first data group obtained by introducing disturbance, training and testing are performed using more than two preset training models to obtain an initial control parameter prediction model of the magnetic levitation compressor, so as to find the optimal parameters under each disturbance in advance; then all the data found in advance can be trained into an initial control parameter prediction model through machine learning, which is integrated into the magnetic bearing controller, and the optimal control parameters given by the control parameter prediction model are used under each operating condition or disturbance of the magnetic bearing compressor system, thereby improving the control effect and making the magnetic bearing rotor stably and suspended with high displacement accuracy.

[0191] The control unit 104 is further configured to obtain the operating parameters of the magnetic levitation compressor, recorded as compressor operating parameters, when the displacement current double closed-loop control system operates according to the initial control parameter prediction model; and obtain the operating parameters of the unit in which the magnetic levitation compressor is located, recorded as unit operating parameters; and record the input and output of the initial control parameter prediction model, the compressor operating parameters and the unit operating parameters as a second data group. The specific functions and processing of the control unit 104 are also referred to step S140.

[0192] The control unit 104 is further configured to perform training and testing based on the second data set using two or more preset training models to obtain an updated control parameter prediction model of the magnetic levitation compressor, so as to control the operation of the magnetic bearing according to the updated control parameter prediction model of the magnetic levitation compressor. The specific functions and processing of the control unit 104 are also shown in step S150.

[0193] Bearing control involves control algorithms, and control parameters need to be adjusted to ensure excellent bearing control performance. It is difficult to adjust the optimal parameters manually, and the optimal control within a certain parameter range cannot be guaranteed. The scheme of adaptive optimal parameter control of magnetic bearings provided by the scheme of the present invention simulates various possible disturbances in the actual operation of the magnetic levitation compressor, and finds the optimal parameters under each disturbance in advance; all the data found in advance are trained by machine learning to become an initial control parameter prediction model, which is integrated into the magnetic bearing controller, and the optimal control parameters given by the control parameter prediction model are used under each working condition or disturbance of the magnetic bearing compressor system, thereby improving the control effect, so that the magnetic bearing rotor is stable and suspended with high displacement accuracy. At the same time, the displacement accuracy, impact force, control parameters used, etc. of each magnetic bearing compressor system in the actual operating conditions on site are obtained through the fifth generation mobile communication technology (5th Generation Mobile Communication Technology, referred to as 5G) module, and the data set is fed back to the aforementioned control parameter prediction model, and the control parameter prediction model is further trained to achieve optimal control.

[0194] In some implementations, the control unit 104, based on the second data set, uses two or more preset training models for training and testing to obtain an updated control parameter prediction model for the magnetic levitation compressor, including:

[0195] The control unit 104 is further configured to obtain the operating parameters of the magnetic levitation compressor, recorded as compressor operating parameters, when the displacement current double closed-loop control system operates according to the initial control parameter prediction model; and obtain the operating parameters of the unit in which the magnetic levitation compressor is located, recorded as unit operating parameters; and record the input and output of the initial control parameter prediction model, the compressor operating parameters and the unit operating parameters as a second data group. The specific functions and processing of the control unit 104 are also referred to step S410.

[0196] The control unit 104 is further configured to perform preprocessing on the second data group based on each of the two or more preset training models to obtain second sample data under each training model; that is, based on each of the two or more preset training models, the second data group is preprocessed to obtain second sample data under each training model. The specific functions and processing of the control unit 104 are also shown in step S420.

[0197] In some embodiments, in the second data group, the compressor operating parameters include at least one of the following: the current of the motor in the compressor, the operating frequency of the motor in the compressor, and the bus voltage of the motor controller in the compressor; the unit where the magnetic levitation compressor is located includes an air-conditioning unit, and the unit operating parameters include at least one of the following: the shutdown or running state of the air-conditioning unit, the operating conditions of the air-conditioning unit, the state of the cooling water flow switch in the air-conditioning unit, the state of the freezing water flow switch in the air-conditioning unit, the opening of the throttle valve in the air-conditioning unit, the load of the air-conditioning unit, the cold water inlet pressure in the air-conditioning unit, the freezing water outlet pressure in the air-conditioning unit, the current heat exchange capacity of the air-conditioning unit, the condensing pressure of the condenser in the air-conditioning unit, and the evaporating pressure of the evaporator in the air-conditioning unit.

[0198] The control unit 104 pre-processes the second data group based on each of the two or more preset training models to obtain second sample data under each training model, including:

[0199] The control unit 104 is further configured to filter the compressor operating parameters in the second data group according to the first corresponding setting parameter range, and obtain corresponding operating parameters within the first corresponding setting parameter range, which are recorded as first filtered operating parameters. The specific functions and processing of the control unit 104 are also shown in step S510.

[0200] The control unit 104 is further configured to filter the unit operating parameters in the second data group according to the second corresponding setting parameter range, and obtain corresponding operating parameters within the second corresponding setting parameter range, which are recorded as second filtered operating parameters. The specific functions and processing of the control unit 104 are also shown in step S520.

[0201] The control unit 104 is further configured to use the input and output of the initial control parameter prediction model, the first screening operation parameter and the second screening operation parameter as a screening data set. The specific functions and processing of the control unit 104 are also shown in step S530.

[0202] The control unit 104 is further configured to construct an input and output data pair for the screening data set based on the input and output requirement relationship of each training model in the two or more preset training models, and obtain the second sample data under each training model. The specific functions and processing of the control unit 104 are also shown in step S540.

[0203] To achieve the best effect of the control parameter prediction model, such as Figure 7As shown, the current operating data of the unit system is obtained in real time through the 5G module of the unit system controller through the remote PC host computer, where each data includes the current state of the unit, the operating condition of the unit, the cooling water flow switch state, the chilled water flow switch state, the throttle valve conduction opening, the unit load, the chilled water inlet pressure, the chilled water outlet pressure, the current cooling capacity, the condensing pressure, the evaporating pressure, the magnetic levitation compressor current, the magnetic levitation compressor motor operating frequency, the magnetic levitation controller bus voltage, the magnetic levitation bearing displacement accuracy, the magnetic levitation bearing current, the magnetic bearing control parameters K1, K2, K3, a total of 19 feature categories.

[0204] Among them, the current state of the unit refers to the state of the unit being shut down or running, and the operating condition of the unit refers to the cooling water flow switch state, the chilled water flow switch state, the throttle valve conduction opening, the unit load, the chilled water inlet pressure, the chilled water outlet pressure, the current cooling capacity, the condensing pressure, the evaporating pressure, the magnetic levitation compressor current, the magnetic levitation compressor motor operating frequency, the magnetic levitation controller bus voltage, the magnetic levitation bearing displacement accuracy, the magnetic levitation bearing current, and the set values ​​of the magnetic bearing control parameters K1, K2, and K3 under each condition.

[0205] The data of the magnetic bearing unit during operation is collected, and each piece of data contains the 19 feature categories of the above unit operation data. The obtained data is preprocessed as mentioned above, for example: abnormal data is deleted, including data that is obviously incorrect and inconsistent with reality, data with artificially found data records, and data with some missing information.

[0206] In the scheme of the present invention, based on each of more than two preset training models, the second data group obtained based on the initial control parameter model is preprocessed so that the second sample data obtained under each training model is more accurate, and then, the control parameter prediction model is further trained to achieve optimal control.

[0207] The control unit 104 is further configured to train and test each training model using the second sample data under each training model to obtain an updated control parameter prediction model under each training model; thereby, an updated control parameter prediction model under each training model in more than two preset training models is obtained. The specific functions and processing of the control unit 104 are also referred to step S430.

[0208] The control unit 104 is further configured to perform weighted average processing on the updated control parameter prediction model under each training model in the two or more preset training models to obtain the weighted average updated control parameter prediction model of the two or more preset training models as the updated control parameter prediction model of the magnetic levitation compressor. The specific functions and processing of the control unit 104 are also shown in step S440.

[0209] See also Figure 7 In the example shown, the main controller of the magnetic levitation compressor system communicates with the magnetic levitation compressor system through the 5G communication module: receiving the real-time operating current and displacement accuracy data of the magnetic levitation bearing transmitted by the magnetic levitation compressor system, sending the current operating conditions and pressure difference of the compressor system to the bearing controller, and updating the latest control parameter prediction model to the main control chip of the bearing. The main controller of the magnetic levitation compressor system interacts with the remote PC host computer (including the unit operation data monitoring system and the control parameter prediction model training system): the remote PC host computer transmits the latest control parameter prediction model to the main controller of the magnetic levitation compressor system, and the main controller of the magnetic levitation compressor system feeds back the pressure difference, working conditions and received magnetic bearing data of the magnetic levitation compressor system to the remote PC host computer.

[0210] See also Figure 7 In the example shown, the training architectures of the three models are updated while preprocessing the acquired data, such as re-introducing the newly acquired data into the model for training:

[0211] a. The number of nodes in the hidden layer of the extreme learning machine is 10.

[0212] b. The number of random features of random forest is mtry=4.

[0213] c. The support vector machine inputs 16 values ​​and outputs 3 values. Among them, the input 16 values, such as: the current state of the unit, the cooling water flow switch state under each working condition, the chilled water flow switch state, the throttle valve guide vane opening, the unit load, the chilled water inlet pressure, the chilled water outlet pressure, the current cooling capacity, the condensing pressure, the evaporation pressure, the magnetic levitation compressor current, the magnetic levitation compressor motor operating frequency, the magnetic levitation controller bus voltage, the magnetic levitation bearing displacement accuracy, the magnetic levitation bearing current; output 3 values, such as: magnetic bearing control parameters K1, K2, K3.

[0214] The data is distinguished into training set and test set, and the previous generation of control parameter prediction model is further optimized and iterated. This cycle is repeated until each magnetic levitation compressor can achieve the expected operating effect under different working conditions, and then the iteration is stopped. When the components of the unit are damaged and the parameters need to be re-matched, the control parameter prediction model can be iterated again.

[0215] The optimal parameter adaptive control scheme and control parameter training device of the magnetic levitation compressor constructed by the scheme of the present invention finds the optimal parameters within a certain parameter range through optimization and model training, enhances the reliability of control, improves the bearing control performance, and greatly improves the stability of the magnetic levitation compressor system. The magnetic bearing control parameter prediction method provided by the scheme of the present invention can be used not only in the magnetic levitation bearing system, but also in other control occasions.

[0216] In some embodiments, the two or more preset training models include: at least two of an extreme learning machine model, a support vector machine model, and a random forest model. Of course, preferably, the two or more preset training models include: an extreme learning machine model, a support vector machine model, and a random forest model.

[0217] Figure 8 This is a schematic diagram of the implementation flow of the extreme learning machine model structure. Figure 8 As shown in Figure 1, the implementation process of the extreme learning machine model structure is as follows: Extreme Learning Machine (ELM) is also a neural network, which belongs to a feedforward model with a single hidden layer. ELM randomly initializes the weights of the input layer and the hidden layer and the bias of the nodes, and finally calculates the weights of the hidden layer and the output layer.

[0218] Assume there are n sample points (x j ,t j ), then the extreme learning machine model can be written as follows:

[0219]

[0220] Among them, g(x) is the activation function, W i =[W i,1 ,W i,2 ,..,W i,n ] T is the input weight, β i is the weight of the output layer (beta), b i is the bias of the i-th hidden node. i *X j W i and X j The inner product of O. j Indicates output, X j represents input, and N represents the number of neural network layers.

[0221] The learning goal of the extreme learning machine is to minimize the error between the predicted value and the actual value of the output:

[0222]

[0223] Among them, t j Indicates actual value.

[0224] That is, there is β i , W i and b i , so that:

[0225]

[0226] It can be expressed as a matrix:

[0227] H*β=T (4).

[0228] Among them, H is the output of the hidden layer node, β is the weight of the connection between the hidden layer and the output layer, and T is the output label in the training data pair, which is specifically expressed as follows:

[0229]

[0230]

[0231] In order to be able to train a single hidden layer neural network, we want to get b i and β i , so that:

[0232]

[0233] Among them, H represents the output of the hidden layer node, β represents the weight of the connection between the hidden layer and the output layer, T represents the output label in the training data pair, and b i represents the bias of the i-th hidden node, W i =[W i,1 ,W i,2 ,..,W i,n ] T represents the input weight. i=1,...,L, which is equivalent to minimizing the loss function E:

[0234]

[0235] In the ELM algorithm, when we randomly determine the input weights of the input layer and the hidden layer, W i and bias, b i , the output matrix H of the hidden layer can be uniquely determined. The solution to the single hidden layer neural network model can be transformed into solving a linear equation:

[0236] H*β=T (8).

[0237] Among them, formula (8) is the same as formula (4).

[0238] And the output weight ③β can be determined:

[0239]

[0240] Among them, H + is the generalized inverse of the matrix H. And it can be proved that the solution Its norm is the smallest and unique; β represents the weight connecting the hidden layer and the output layer, and T represents the output label in the training data pair.

[0241] The implementation steps of the ELM algorithm are as follows:

[0242] (1) Select the number of neurons in the hidden layer. The number of neurons is calculated according to the formula where m is the number of neurons in the hidden layer, n is the number of input nodes, l is the number of output nodes, and α is a constant between 1 and 10.

[0243] In the solution of the present invention, the number of nodes in the input layer of the initial control parameter prediction model is 5, and the number of nodes in the output layer is 3. After repeated adjustment, the number of nodes in the hidden layer is finally selected as 8.

[0244] (2) Randomly set the connection weights weight (0 < weight < 1) between the input layer and the hidden layer and the biases bias (0 < bias < 1) of the neurons.

[0245] (3) Select an activation function. Due to the good performance of the sigmoid function, the function H = 1 / (1 + exp(-H)) is selected in the solution of the invention. Calculate the matrix H of the hidden layer output layer, and calculate the weight β of the output layer using the formula.

[0246] Through multiple calculations and tests, the number of neurons is finally selected as 8. Import 3800 groups of data after data processing, and solve the prediction model of the extreme learning machine.

[0247] Set the model parameters well. Use 200 groups of test data for testing. If the accuracy rate reaches 90%, integrate it into the host computer.

[0248] Fig. 9 It is a schematic flow chart for solving the random forest control parameter prediction model. As Fig. 9 shown, the implementation steps of the random forest regression prediction model structure are as follows:

[0249] ① There are ntree trees in the random forest. Each regression tree corresponds to a training set. To build ntree regression trees, the corresponding number of training sets need to be generated. In the process of generating the random forest algorithm, the bagging sampling technique is mainly used to generate ntree training subsets from the original training set. Each sampling is random and with replacement.

[0250] ②Then train a decision tree base learner for each new data set in turn according to the random attribute selection method. Assuming that the number of independent variables in the original sample is p, a subset of mtry independent variables is randomly selected from the independent variable set of each decision tree, and then an optimal independent variable is selected as the partition attribute. In the random forest regression algorithm, mtry = log 2 p.

[0251] ③ Thus, ntree regression decision trees are obtained. Each tree starts to recursively generate child nodes from top to bottom, and determines when the decision tree will terminate the split by setting the minimum threshold of the number of samples contained in the leaf node (the regression tree can be 5).

[0252] ④ From ntree decision trees, ntree prediction values ​​can be obtained. The weighted average of ntree prediction values ​​is taken using the ensemble learning idea as the regression prediction result, which will be the final output result of the algorithm.

[0253] The regression tree mathematical model is:

[0254]

[0255] Where: M is the number of samples in the subset of the regression tree model; c m is the mean response value of the data samples in each subset; R m are the subsets divided; I(x∈Rm) is the characteristic function, when x∈R m The value is 1 when it is set, otherwise the value is 0.

[0256] From the original training sample set of 3800 random forests (RF), 3800 training samples were randomly selected with replacement to obtain 1 sub-sample set, denoted as M=3800, and one regression tree was constructed for one sample; the experiment was repeated 300 times to obtain 500 sub-sample sets (the 500 sub-sample sets were independent of each other and the samples could be repeated), and 500 regression decision trees were constructed, denoted as ntree=300.

[0257] Since the selected independent variable of the initial control parameter prediction model in the solution of the present invention is p=5, a regression tree learner is trained by randomly selecting feature numbers. Generally, the random feature number mtry=log 2 p, so in the solution of the present invention, mtry=2. 500 regression decision trees are trained, and finally the results of the 500 regression subtrees are averaged as the prediction result.

[0258] The accuracy of the established random forest regression model is verified by using the remaining 200 groups of validation samples. If the accuracy rate exceeds 90%, the random forest model is integrated into the host computer.

[0259] Fig.10 This is a schematic diagram of the support vector machine regression solution process. Fig.10 As shown in Figure 1, the implementation steps of the support vector machine structure are as follows: Support Vector Machine (SVM) is a binary classification model that maps the feature vector of an instance (taking two dimensions as an example) to some points in space, such as Fig.10 The red and black points in the figure belong to two different categories. The purpose of SVM is to draw a line that "best" distinguishes the two categories of points, so that if new points appear in the future, this line can also make a good classification. There are countless lines that can be drawn, and the difference lies in the effect. Each line can be called a dividing hyperplane. The best line we hope to find is the "partitioning hyperplane with the largest margin". From Fig.10 It can be seen that the distances from the points on the dotted line to the dividing hyperplane are the same. In fact, only these points jointly determine the position of the hyperplane, so they are called "support vectors."

[0260] The partitioning hyperplane can be defined as a linear equation:

[0261] ω T x+b=y;

[0262] Where ω={ω 1 ;ω 2 ;...;ω d} is a normal vector that determines the direction of the hyperplane; d is the number of eigenvalues; x is the training sample; b is the displacement term that determines the distance between the hyperplane and the origin; as long as the normal vector ω and the displacement b are determined, a partitioning hyperplane can be uniquely determined. x represents the input of the training sample, y represents the output of the training sample, and the superscript T represents the transposition symbol.

[0263] Since the SVM regression model (SVR) is used in the solution of the present invention, the hyperplane decision boundary obtained when the support vector is used for classification is the regression model when the support vector is used for regression:

[0264] y i =ω T x i +b (11).

[0265] The distance between any point on the partitioning hyperplane and the marginal hyperplanes on both sides is So the minimum margin is

[0266] The constraint condition in SVR regression is to make the points in each training set as concentrated as possible near the support vector; in SVR, in order to obtain the distance from all feature sample points to the hyperplane, the sample points closest to the hyperplane are found at the same time.

[0267] The function interval is defined: for the selected training data set T and hyperplane (ω, b), the hyperplane and the feature sample point (x i ,y i ) is:

[0268] r i =y i (ω T x i +b) (12).

[0269] In order to find the support vector, we can calculate the minimum distance r of all sample points. min The one that:

[0270] r min =min(r) (13).

[0271] There is a serious problem with the function interval, that is, when ω and b increase or decrease at a given ratio, the hyperplane equation does not change, but the function interval becomes a multiple of the original ratio. In order to avoid this situation, a concept of geometric interval is introduced, namely:

[0272] r i =y i ((ω T / ||ω||)x i +b / ||ω||) (14).

[0273] This ensures that if ω and b change proportionally, the hyperplane will also change according to the given proportional multiple. Since the learning purpose of SVM is to maximize the corresponding geometric interval, the goal is to maximize the geometric interval, and the expression is as follows:

[0274]

[0275] In terms of optimization objective function, SVR and SVM are consistent. However, under the constraint condition, SVM makes each point in the training set as far away from the support vector of its own category (sign) as possible; for the regression model, our indicator is to make each feature sample point (x i ,y i ), try to fit a linear model as much as possible:

[0276] f(X)=ω T X+b.

[0277] In ε-SVR, if the sample point is close enough to the regression model and falls within the interval of the regression model, the loss of the sample point will not be calculated. The loss function is defined as follows:

[0278]

[0279] Therefore, the objective function of SVR can be expressed as follows:

[0280]

[0281] By introducing slack variables and Lagrange multipliers, the problem is transformed into a dual problem. The coefficient ω of the support vector machine model is solved using the Karush-Kuhn-Tucker conditions (KKT) and the sequential minimal optimization algorithm (SMO). T , b, determine the input 5 values ​​of the support vector machine and output 3 values. Input 5 values, such as: unit load, magnetic levitation compressor current, magnetic levitation compressor motor operating frequency, magnetic levitation bearing displacement accuracy, magnetic levitation bearing current; output 3 values, such as: magnetic bearing control parameters K1, K2, K3.

[0282] The 3800 groups of original training sample sets are divided to train the SVR model, and the remaining 200 groups are used as validation samples. If the validation accuracy reaches 90%, they are integrated into the host computer as a regression prediction model.

[0283] Fig.11 A flow chart of the process of solving machine learning problems. Fig.11 As shown in the figure, the schematic diagram of heterogeneous (i.e. different types of training models) integrated learning solution integrates the extreme learning machine, support vector machine and random forest regressor established above into the host computer data monitoring system. Input the real-time data into the three models trained previously to obtain three prediction results. According to the prediction accuracy of the previous model, the weights are assigned and the average is taken as the control parameter prediction output; among them, the weighted average prediction has better effect. For example: 0.1*prediction value of extreme learning machine + 0.5*prediction value of support vector machine + 0.4*prediction value of random forest regressor = final output value.

[0284] In the scheme of the present invention, at least two, preferably three, of the extreme learning machine model, the support vector machine model and the random forest model are used for training and testing, and the three prediction results corresponding to the three models trained previously are input with real-time data and weighted averaged to be used as the control parameter prediction output, so that the training is more comprehensive, thereby making the output more accurate and stable.

[0285] Since the processing and functions implemented by the device of this embodiment basically correspond to the embodiments, principles and examples of the aforementioned method, for the details not fully described in this embodiment, please refer to the relevant descriptions in the aforementioned embodiments, and no further elaboration will be made here.

[0286] According to an embodiment of the present invention, a magnetic levitation compressor corresponding to the control device of the magnetic levitation compressor is also provided. The magnetic levitation compressor may include: the control device of the magnetic levitation compressor described above.

[0287] Since the processing and functions implemented by the magnetic levitation compressor of this embodiment basically correspond to the embodiments, principles and examples of the aforementioned device, for the details not fully described in this embodiment, please refer to the relevant descriptions in the aforementioned embodiments, and no further elaboration will be made here.

[0288] According to an embodiment of the present invention, a computer program product corresponding to a magnetic levitation compressor is also provided, including a computer program, and when the computer program is executed by a processor, the steps of the control method of the magnetic levitation compressor described above are implemented.

[0289] Since the processing and functions implemented by the product of this embodiment basically correspond to the embodiments, principles and examples of the aforementioned magnetic levitation compressor, for the details not fully described in this embodiment, please refer to the relevant descriptions in the aforementioned embodiments, and no further elaboration will be made here.

[0290] According to an embodiment of the present invention, a storage medium corresponding to the control method of the magnetic levitation compressor is also provided, wherein the storage medium includes a stored program, wherein when the program is running, the device where the storage medium is located is controlled to execute the steps of the control method of the magnetic levitation compressor described above.

[0291] Since the processing and functions implemented by the storage medium of this embodiment basically correspond to the embodiments, principles and examples of the aforementioned method, for the details not fully described in this embodiment, please refer to the relevant descriptions in the aforementioned embodiments, and no further elaboration will be made here.

[0292] In summary, it is easy for those skilled in the art to understand that, under the premise of no conflict, the above-mentioned advantageous methods can be freely combined and superimposed.

[0293] The above description is only an embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention shall be included in the scope of the claims of the present invention.

Claims

1. A control method for a magnetic levitation compressor, characterized in that: The magnetic suspension compressor has a magnetic bearing, and the control system of the magnetic bearing adopts a displacement current double closed-loop control system; The control method of the magnetic levitation compressor comprises: In the case where the magnetic levitation compressor is running, one or more preset disturbances are introduced into the displacement current double closed-loop control system; and control parameters of the displacement current double closed-loop control system are adjusted more than once; For each of the one or more preset disturbances introduced, when the magnetic bearing operates based on the control parameters of the displacement current double closed-loop control system adjusted each time, the rotor displacement of the magnetic bearing and the stator coil current of the magnetic bearing are obtained; thereby, the control parameters, rotor displacement and stator coil current of the displacement current double closed-loop control system corresponding to all disturbances in the one or more preset disturbances introduced are obtained, which are recorded as a first data group; Based on the first data group, two or more preset training models are used for training and testing to obtain an initial control parameter prediction model of the magnetic levitation compressor, so as to control the operation of the magnetic bearing according to the initial control parameter prediction model of the magnetic levitation compressor; When the displacement current double closed-loop control system operates according to the initial control parameter prediction model, the operating parameters of the magnetic levitation compressor are obtained and recorded as compressor operating parameters; and the operating parameters of the unit in which the magnetic levitation compressor is located are obtained and recorded as unit operating parameters; the input and output of the initial control parameter prediction model, the compressor operating parameters and the unit operating parameters are recorded as a second data group; Based on the second data group, training and testing are performed using two or more preset training models to obtain an updated control parameter prediction model for the magnetic levitation compressor, so as to control the operation of the magnetic bearing according to the updated control parameter prediction model for the magnetic levitation compressor.

2. The control method of the magnetic levitation compressor according to claim 1, characterized in that: in, When the magnetic levitation compressor is running, one or more preset disturbances are introduced into the displacement current double closed-loop control system, including: Adding one or more interference sources to the feedback value of the rotor displacement of the displacement loop of the displacement current double closed-loop control system to introduce one or more preset disturbances into the displacement current double closed-loop control system; wherein the one or more interference sources include: at least one of Gaussian white noise of a set frequency, sinusoidal interference signals of different amplitudes, and random pulse signals of different amplitudes; and / or, Two or more preset training models, including at least two of an extreme learning machine model, a support vector machine model, and a random forest model.

3. The control method of the magnetic levitation compressor according to claim 1, characterized in that: Based on the first data group, two or more preset training models are used for training and testing to obtain an initial control parameter prediction model of the magnetic levitation compressor, including: Based on each of the two or more preset training models, preprocessing the first data group to obtain first sample data under each training model; Using the first sample data under each training model to train and test each training model, a control parameter prediction model under each training model is obtained; thereby, a control parameter prediction model under each training model in more than two preset training models is obtained; The control parameter prediction model under each training model in more than two preset training models is weighted averaged to obtain a weighted average control parameter prediction model of more than two preset training models as the initial control parameter prediction model of the magnetic levitation compressor.

4. The control method of the magnetic levitation compressor according to claim 3, characterized in that: Based on each of the two or more preset training models, preprocessing the first data group to obtain first sample data under each training model includes: Based on the first data group, under each disturbance among the one or more preset disturbances introduced, the control parameters corresponding to the case where the rotor displacement is minimum and the stator coil current is minimum are determined as the optimal control parameters under each disturbance; thereby, the optimal control parameters, the minimum rotor displacement and the minimum stator coil current corresponding to all disturbances among the one or more preset disturbances introduced are obtained, which are recorded as the initial data group; For the initial data group, based on the required relationship between the input and output of each training model in more than two preset training models, a data pair of input and output is constructed to obtain the first sample data under each training model.

5. The control method of the magnetic levitation compressor according to any one of claims 1 to 4, characterized in that: Based on the second data group, two or more preset training models are used for training and testing to obtain an updated control parameter prediction model of the magnetic levitation compressor, including: When the displacement current double closed-loop control system operates according to the initial control parameter prediction model, the operating parameters of the magnetic levitation compressor are obtained and recorded as compressor operating parameters; and the operating parameters of the unit in which the magnetic levitation compressor is located are obtained and recorded as unit operating parameters; the input and output of the initial control parameter prediction model, the compressor operating parameters and the unit operating parameters are recorded as a second data group; Based on each of the two or more preset training models, preprocessing the second data group to obtain second sample data under each training model; Using the second sample data under each training model to train and test each training model, obtaining an updated control parameter prediction model under each training model; thereby obtaining an updated control parameter prediction model under each training model in more than two preset training models; The updated control parameter prediction model under each training model in more than two preset training models is weighted averaged to obtain a weighted average updated control parameter prediction model of more than two preset training models as the updated control parameter prediction model of the magnetic levitation compressor.

6. The control method of the magnetic levitation compressor according to claim 5, characterized in that: In the second data group, the compressor operating parameters include at least one of the following: the current of the motor in the compressor, the operating frequency of the motor in the compressor, and the bus voltage of the motor controller in the compressor; the unit in which the magnetic levitation compressor is located includes an air-conditioning unit, and the unit operating parameters include at least one of the following: the shutdown or running state of the air-conditioning unit, the operating condition of the air-conditioning unit, the state of the cooling water flow switch in the air-conditioning unit, the state of the freezing water flow switch in the air-conditioning unit, the opening of the throttle valve in the air-conditioning unit, the load of the air-conditioning unit, the cold water inlet pressure in the air-conditioning unit, the freezing water outlet pressure in the air-conditioning unit, the current heat exchange amount of the air-conditioning unit, the condensing pressure of the condenser in the air-conditioning unit, and the evaporating pressure of the evaporator in the air-conditioning unit; Based on each of the two or more preset training models, preprocessing the second data group to obtain second sample data under each training model includes: The compressor operating parameters in the second data group are screened according to a first corresponding setting parameter range to obtain corresponding operating parameters within the first corresponding setting parameter range, which are recorded as first screened operating parameters; The unit operating parameters in the second data group are screened according to a second corresponding setting parameter range to obtain corresponding operating parameters within the second corresponding setting parameter range, which are recorded as second screened operating parameters; Using the input and output of the initial control parameter prediction model, the first screening operation parameter and the second screening operation parameter as a screening data set; For the screening data group, based on the demand relationship between the input and output of each training model in more than two preset training models, a data pair of input and output is constructed to obtain second sample data under each training model.

7. A control device for a magnetic levitation compressor, characterized in that: The magnetic suspension compressor has a magnetic bearing, and the control system of the magnetic bearing adopts a displacement current double closed-loop control system; The control device of the magnetic suspension compressor comprises: A control unit configured to introduce one or more preset disturbances into the displacement current double closed-loop control system when the magnetic levitation compressor is running; and adjust control parameters of the displacement current double closed-loop control system more than once; an acquisition unit configured to acquire, for each of the one or more preset disturbances introduced, a rotor displacement of the magnetic bearing and a stator coil current of the magnetic bearing when the magnetic bearing operates based on the control parameters of the displacement current double closed-loop control system adjusted each time; thereby, the control parameters, rotor displacement and stator coil current of the displacement current double closed-loop control system corresponding to all disturbances in the one or more preset disturbances introduced are obtained, which are recorded as a first data group; The control unit is further configured to perform training and testing based on the first data group using two or more preset training models to obtain an initial control parameter prediction model of the magnetic levitation compressor, so as to control the operation of the magnetic bearing according to the initial control parameter prediction model of the magnetic levitation compressor; The control unit is further configured to obtain the operating parameters of the magnetic levitation compressor, recorded as compressor operating parameters, when the displacement current double closed-loop control system operates according to the initial control parameter prediction model; and obtain the operating parameters of the unit in which the magnetic levitation compressor is located, recorded as unit operating parameters; and record the input and output of the initial control parameter prediction model, the compressor operating parameters and the unit operating parameters as a second data group; The control unit is also configured to perform training and testing based on the second data group using two or more preset training models to obtain an updated control parameter prediction model of the magnetic levitation compressor, so as to control the operation of the magnetic bearing according to the updated control parameter prediction model of the magnetic levitation compressor.

8. A magnetic levitation compressor, characterized in that: include: The control device for the magnetic levitation compressor as claimed in claim 7.

9. A storage medium, characterized in that: The storage medium includes a stored program, wherein when the program is executed, the device where the storage medium is located is controlled to execute the control method of the magnetic levitation compressor according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the control method of the magnetic levitation compressor according to any one of claims 1 to 6 are implemented.

Images