WO2023050050A1 - 基于神经网络的同步电机的参数辨识的方法 - Google Patents

基于神经网络的同步电机的参数辨识的方法 Download PDF

Info

Publication number
WO2023050050A1
WO2023050050A1 PCT/CN2021/121272 CN2021121272W WO2023050050A1 WO 2023050050 A1 WO2023050050 A1 WO 2023050050A1 CN 2021121272 W CN2021121272 W CN 2021121272W WO 2023050050 A1 WO2023050050 A1 WO 2023050050A1
Authority
WO
WIPO (PCT)
Prior art keywords
neural network
network model
synchronous motor
parameters
identified
Prior art date
Application number
PCT/CN2021/121272
Other languages
English (en)
French (fr)
Inventor
陈宇
童思雨
臧晓云
Original Assignee
罗伯特·博世有限公司
陈宇
童思雨
臧晓云
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 罗伯特·博世有限公司, 陈宇, 童思雨, 臧晓云 filed Critical 罗伯特·博世有限公司
Priority to DE112021007423.7T priority Critical patent/DE112021007423T5/de
Priority to CN202180102774.3A priority patent/CN118077134A/zh
Priority to PCT/CN2021/121272 priority patent/WO2023050050A1/zh
Publication of WO2023050050A1 publication Critical patent/WO2023050050A1/zh

Links

Images

Classifications

    • H—ELECTRICITY
    • H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
    • H02P—CONTROL OR REGULATION OF ELECTRIC MOTORS, ELECTRIC GENERATORS OR DYNAMO-ELECTRIC CONVERTERS; CONTROLLING TRANSFORMERS, REACTORS OR CHOKE COILS
    • H02P21/00—Arrangements or methods for the control of electric machines by vector control, e.g. by control of field orientation
    • H02P21/0003—Control strategies in general, e.g. linear type, e.g. P, PI, PID, using robust control
    • H02P21/0014—Control strategies in general, e.g. linear type, e.g. P, PI, PID, using robust control using neural networks
    • H—ELECTRICITY
    • H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
    • H02P—CONTROL OR REGULATION OF ELECTRIC MOTORS, ELECTRIC GENERATORS OR DYNAMO-ELECTRIC CONVERTERS; CONTROLLING TRANSFORMERS, REACTORS OR CHOKE COILS
    • H02P21/00—Arrangements or methods for the control of electric machines by vector control, e.g. by control of field orientation
    • H02P21/14—Estimation or adaptation of machine parameters, e.g. flux, current or voltage

Definitions

  • the invention relates to a method for identifying parameters of a synchronous motor based on a neural network, a device for identifying parameters of a synchronous motor based on a neural network, and a computer program product.
  • Synchronous motors have attracted increasing attention in industrial applications due to their low cost, high efficiency, and high stability.
  • the traditional linear model can no longer accurately reflect the real dynamic characteristics of the motor due to reasons such as magnetic circuit saturation, temperature drift, and cross-coupling, and many parameters of the motor show nonlinear characteristics.
  • parameter identification of the synchronous motor is required to achieve better control performance.
  • motor parameters are usually modeled in the form of a look-up table, but this requires traversing the full range of operating points of the motor, which is time-consuming and expensive.
  • methods for fitting nonlinear inductive surfaces through mathematical models or special formulas are also known, but these mathematical models or formulas are often very complex, their coefficients are also parameterized, and the curve drawing algorithm also requires corresponding database support.
  • the object of the present invention is to provide a method for parameter identification of a synchronous motor based on a neural network, a device for identifying a parameter of a synchronous motor based on a neural network, and a computer program product to at least solve some of the problems in the prior art .
  • a method for parameter identification of a synchronous motor based on a neural network comprising the following steps:
  • the present invention especially includes the following technical idea: use the neural network to establish the parameter model of the motor, and use the self-learning ability and fuzzy logic of the artificial neural network to optimize the model parameters when there are only a small number of operating points and therefore only limited observation data Estimation improves the efficiency and accuracy of parameter identification.
  • a two-stage model training process is proposed to complete parameter estimation, so that different task emphases can be assigned to the two training processes, and different characteristics of unknown parameters can be well fitted in a more targeted manner.
  • the intermediate results include:
  • the first estimated value of the second parameter among the parameters to be identified In the case of performing linear approximation on the first parameter, the first estimated value of the second parameter among the parameters to be identified.
  • the final result includes:
  • a second estimated value of a second parameter among the parameters to be identified after nonlinear compensation is performed on the linear approximation of the first parameter.
  • different training data are used for the training of the first neural network model and the training of the second neural network model.
  • the observed data in the case of unsaturated magnetic circuits can be selected empirically to fit the linear characteristics of unknown parameters, and the observed data in the case of magnetic circuit saturation can be selected for nonlinear compensation, thereby speeding up the network convergence process .
  • the parameters to be identified of the synchronous motor include: stator resistance, d-axis inductance, q-axis inductance and permanent magnet flux linkage of the synchronous motor.
  • the scheme proposed here can realize the synchronous identification of multiple motor parameters, so there is no need to pre-detect some fixed value parameters with the help of additional sensors, saving time overhead.
  • the step S1 includes: establishing the input-output relationship of the first neural network model based on the static voltage state equation of the synchronous motor in the d, q-axis coordinate system, wherein, the q-axis current and the d-axis current and the synchronous The electrical angular velocity of the motor is used as the input of the first neural network model, the q-axis voltage and the d-axis voltage are used as the output of the first neural network model, and the weight of the first neural network model is at least partly represented by the parameters to be identified.
  • the artificial neural network can have a good input-output mapping relationship, can generate reasonable outputs for new inputs, and continuously optimize its own performance during the learning process. Fully exploiting existing data patterns in very limited cases speeds up the parameter identification process.
  • step S1 includes:
  • the known observation data including the d-axis current, q-axis current, d-axis voltage, q-axis voltage and electrical angular velocity of the synchronous motor that have been measured at different operating points;
  • the intermediate result of the parameter to be identified of the synchronous motor is determined based on the weight of the first neural network model when the loss function satisfies the first preset condition.
  • the following technical advantage is achieved in particular: During this time, the unknown parameters of the system are reflected in the internal weights, so that a direct identification of the unknown parameters of the plant is avoided.
  • adjusting the weights of the first neural network model includes: iteratively updating the weights of the first neural network model by means of a gradient descent algorithm, especially a backpropagation algorithm.
  • the above algorithm can automatically extract "reasonable” solving rules based on the already mastered input-output relationship, thereby automatically realizing the determination of internal weights, resolving the complex mapping problem of the internal mechanism, and enabling
  • the network has good generalization ability and fault tolerance.
  • the step S2 includes: establishing a third neural network model as the compensation module, using the q-axis current and the d-axis current as the input of the third neural network model, and using the nonlinear characteristic part of the parameter to be identified as the first The output of the three neural network models.
  • the following technical advantages are realized: by introducing the compensation module, on the one hand, the nonlinear factors included in the parameter linear assumption can be reasonably compensated; on the other hand, a dedicated mathematical model can be well created for the nonlinear characteristics of the parameters, so that This makes the final parameter identification results more reliable.
  • the step S2 includes: adding the compensation module to the first neural network model in the following manner: the input layer of the compensation module shares part of the input layer of the first neural network model, and the output layer of the compensation module is connected to to the output layer of the first neural network model.
  • the compensation module by embedding the compensation module into the first neural network model in a manner of sharing input and output, the compensation module is “driven” to perform passive training during the overall training process of the second neural network model. In this way, the unknown nonlinear characteristics of the parameters are converted into the internal parameters of the model for solution. In the absence of direct training data, an input-output mapping relationship is indirectly established for the compensation module through the backpropagation capability of the entire network.
  • step S3 includes:
  • the second neural network model is trained with known observation data, wherein the weights of the trained first neural network model and the weights of the compensation module are adjusted until the loss function satisfies the second preset condition;
  • the final result of the parameters to be identified of the synchronous motor is determined based on the weight of the first neural network model and the weight of the compensation module when the loss function satisfies the second preset condition.
  • the training of the second neural network model is carried out based on the training results of the first stage, and it can be improved on the basis of the intermediate results of the parameters to be identified, which improves the time efficiency and speeds up the process. Convergence process.
  • the following technical advantage is achieved: Since the parameter estimation in the first stage is done under a linear approximation of a specific parameter, in order to distinguish it well from the nonlinear characteristic part in the compensation stage, this linear characteristic part can continue to be considered as The constant exists and thus does not participate in the further training process. However, since this linear approximation is not accurate enough, it is possible that the intermediate results for some fixed parameters incorrectly include the nonlinear contribution of certain parameters. In order to compensate the parts that do not belong to these fixed value parameters, it is necessary to adjust these fixed value parameters in the second stage so as to eliminate the interference items.
  • the loss function is expressed by the following formula:
  • N is the number of known observations used
  • U sd is the d-axis voltage of the synchronous motor in the known observations
  • U sq is the q-axis voltage of the synchronous motor in the known observation data
  • the error between the model output and the actual output can be effectively considered and utilized, so that the model can continuously self-learn toward the direction approaching the real value, providing reliable evaluation indicators and termination conditions for the training process.
  • the method also includes the following steps:
  • the following technical advantage is achieved:
  • the parameter generation process can be kept offline, thus ensuring low computational effort and hardware for the controller of the synchronous motor requirements, a cost-effective control concept is achieved overall.
  • the controller can update the parameter identification results in real time according to different working conditions, and can realize more reliable motor control.
  • a device for parameter identification of a synchronous motor based on a neural network the device is used to perform the method according to the first aspect of the present invention, and the device includes:
  • the first identification module is configured to be able to establish and train a first neural network model and determine an intermediate result of parameters to be identified of the synchronous motor by means of the trained first neural network model;
  • a modification module configured to form a second neural network model by adding a compensation module to the trained first neural network model
  • the second identification module is configured to train the second neural network model based on the intermediate results of the parameters to be identified, and determine the final result of the parameters to be identified by means of the trained second neural network model.
  • a computer program product comprising a computer program for implementing the method according to the first aspect of the present invention when executed by a computer.
  • Fig. 1 shows a flow chart of a method for parameter identification of a synchronous motor based on a neural network according to an exemplary embodiment of the present invention
  • Fig. 2 shows a flowchart of a method step of the method shown in Fig. 1;
  • FIG 3 shows a flowchart of two method steps of the method shown in Figure 1;
  • FIG. 4 shows a schematic diagram of a device for parameter identification of a synchronous motor based on a neural network according to an exemplary embodiment of the present invention
  • Fig. 5 shows a schematic diagram of operating points that need to be traversed for parameter identification by means of a lookup table method
  • Fig. 6 shows the schematic diagram of the principle of the nonlinear characteristics of the d-axis flux linkage of the synchronous motor
  • Fig. 7 shows an exemplary schematic diagram of the first neural network model used in the method according to the present invention.
  • Figure 8 shows an exemplary schematic diagram of a compensation module used in the method according to the present invention.
  • Figure 9 shows an exemplary schematic diagram of a second neural network model used in the method according to the present invention.
  • 10a-10b show schematic diagrams of the nonlinear characteristic part of the d-axis inductance and the nonlinear characteristic part of the q-axis inductance of a synchronous motor determined by the method according to the present invention as a function of current.
  • Fig. 1 shows a flowchart of a method for parameter identification of a synchronous motor based on a neural network according to an exemplary embodiment of the present invention.
  • step S1 a first neural network model is established and trained, and intermediate results of parameters to be identified of the synchronous motor are determined by means of the trained first neural network model.
  • a synchronous machine includes, for example, a synchronous reluctance machine, a permanent magnet synchronous machine, an induction machine or other types of synchronous machines with parametric non-linear properties.
  • the parameters to be identified of a synchronous machine include, for example: the stator resistance R of the synchronous machine, the d-axis inductance L d , the q-axis inductance L q and/or the flux linkage of the permanent magnets
  • the stator resistance R and the flux linkage of the permanent magnet can be considered to be is an intrinsic parameter of the synchronous motor and is therefore invariant.
  • the d-axis inductance L d and q-axis inductance L q of the synchronous motor will change with the change of current. If they are treated as constants, it will definitely affect the control accuracy of the motor and stability.
  • the intermediate result of the parameters to be identified includes, for example, a linear approximation of a first parameter of the parameters to be identified (for example the d-axis inductance L d and the q-axis inductance L q ).
  • the intermediate result also includes: in the case of linear approximation to the first parameter, the second parameter among the parameters to be identified (such as stator resistance R and permanent magnet flux linkage ) of the first estimate.
  • the input-output relationship of the first neural network model is established based on the static voltage state equation of the synchronous motor in the d, q axis coordinate system.
  • the static voltage state equation used is as follows:
  • Usd is the d-axis voltage of the synchronous motor
  • R is the stator resistance of the synchronous motor
  • ⁇ 1 is the electrical angular velocity of the synchronous motor
  • Lq is the q-axis inductance of the synchronous motor
  • i sq is the q-axis current of the synchronous motor
  • U sq is the q-axis voltage
  • isd is the d-axis current of the synchronous motor
  • L d is the d-axis inductance of the synchronous motor.
  • the first neural network model for example, use the q-axis current and the d-axis current and the angular velocity of the synchronous motor as the input of the first neural network model, and use the q-axis voltage and the d-axis voltage as the output of the first neural network model to synchronize Motor parameters to be identified R, L d , L q , Modifiable weights characterizing the first neural network model.
  • a second neural network model is formed by adding a compensation module to the trained first neural network model.
  • the compensation module to describe the nonlinear characteristics of the parameters to be identified, for example, the third neural network model is established as the compensation module, the q-axis current and the d-axis current are used as the input of the third neural network model, and the d-axis inductance and q The nonlinear characteristic part of the shaft inductance is used as the output of the third neural network model.
  • step S3 the second neural network model is trained based on the intermediate results of the parameters to be identified, and the final result of the parameters to be identified is determined by means of the trained second neural network model.
  • the final result represents the result of the parameter to be identified that can characterize its actual value or actual functional relationship.
  • the final result of the parameters to be identified includes, for example: the linear approximation value of the first parameter (such as d-axis inductance and q-axis inductance) among the parameters to be identified is nonlinearly compensated.
  • the final result also includes: after nonlinear compensation of the linear approximation of the first parameter, the second parameter among the parameters to be identified (such as stator resistance R and permanent magnet flux linkage ), compared with the previously determined first estimated value, the second estimated value of the second parameter substantially no longer has or only has less nonlinear component contribution of the first parameter. By excluding this part, the second estimate of the parameter is closer to its true value.
  • the intermediate results of the determined parameters to be identified are updated by means of the second neural network model, and the identification equations of the unknown parameters can be determined according to the updated results.
  • FIG. 2 shows a flowchart of one method step of the method shown in FIG. 1 .
  • method step S1 in FIG. 1 exemplarily includes steps S11-S18.
  • step S11 observation data is acquired.
  • the permanent magnet synchronous motor as an example, here, for example, firstly, the permanent magnet synchronous motor is controlled to run stably. Then by changing the reference current, load torque, motor speed and other conditions, N groups of different operating points can be traversed in the widest possible working range, and the d-axis voltage U sd and q-axis voltage U sq of the motor can be recorded at these operating points , d-axis current i sd , q-axis current i sq and electrical angular velocity ⁇ 1 .
  • the value of N can be freely adjusted according to the time overhead and the expected model accuracy, and usually N is in the range of 15-30.
  • step S12 d-axis current i sd , q-axis current i sq and electrical angular velocity ⁇ 1 in a set of data are extracted from the observation data, and input into the first neural network model in a forward propagation manner.
  • step S13 the output of the first neural network model is calculated according to the initial estimated value of the weights between neurons in each layer.
  • the weights of the first neural network model at least partly describe the parameters R, L d , L q .
  • the values of the weights are taken without considering the nonlinear characteristics of the motor.
  • the initial estimated values of these weights can be obtained by engineering experts based on experience, or can be any value of external random input that conforms to the mathematical and physical meanings.
  • step S14 an output error is calculated.
  • the d-axis voltage output by the first neural network model is calculated respectively and the deviation between the d-axis voltage U sd in the observed data extracted in step S12, and calculate the q-axis voltage output by the first neural network model The deviation from the q-axis voltage U sq in the extracted observation data.
  • step S15 based on the obtained error, the connection weight is adjusted by means of a backpropagation algorithm.
  • the values are updated layer by layer via backpropagation to the first layer via error updates, and all weights are updated together at the end of backpropagation.
  • the learning rate used in the backpropagation algorithm can be freely adjusted depending on the required model accuracy and time expenditure.
  • step S16 it is checked whether the observed data for training the model is exhausted.
  • the number of observation data traversed is greater than or equal to N.
  • step S12 If not exhausted, return to step S12 from step S16, so as to continue to extract the next training sample.
  • step S17 calculate the total loss function for all training samples (ie, all observation data) and judge whether the output result of the loss function satisfies the first preset condition. Here, for example, it is judged whether the output result of the loss function is smaller than the first limit value.
  • step S17 If the output result of the loss function is greater than the first limit value, it means that the current performance of the model has not reached the standard, and it needs to continue training, so jump back from step S17 to step S12 to start a new iterative cycle.
  • step S18 the current weights in the first neural network model can be output, and based on this, the linear approximation part of the parameters to be identified of the synchronous motor can be determined. It is worth noting that since the nonlinear factors of the synchronous motor are not considered at this stage, the parameters determined here exist in the form of constants.
  • the result determined at this stage can be regarded as the average value of this parameter on all observed data, so it is not absolutely accurate, and needs It is corrected by the subsequent nonlinear compensation process.
  • the results determined at this stage also take into account a part of the nonlinear contribution of the first parameter.
  • FIG. 3 shows a flowchart of two method steps of the method shown in FIG. 1 .
  • method step S3 in FIG. 1 exemplarily includes steps S31-S38.
  • step S2 a compensation module in the form of a third neural network model is added to the trained first neural network model, thereby forming a second neural network model.
  • step S31 d-axis current i sd , q-axis current i sq and electrical angular velocity ⁇ 1 in a set of data are extracted from the acquired observation data, and input into the second neural network model in a forward propagation manner.
  • the d-axis current i sd , q-axis current i sq and electrical angular velocity ⁇ 1 are input into the trained first neural network model, and the d-axis current isd and q-axis current i sq are input into the compensation module middle.
  • step S32 the output of the second neural network model is calculated.
  • the corresponding voltage is calculated according to the weight relationship between neurons in each layer
  • step S33 the error between the voltage output by the second neural network model and the voltage in the observed data is calculated.
  • step S34 based on the error, the weight of the trained first neural network model is adjusted by means of a backpropagation algorithm.
  • the weights of the first neural network model still at least partially describe the motor parameters R, L d , L q , However, since the intermediate results L d0 , L q0 of L d , L q have already represented the linear characteristic part, they will not be involved in the further training process. However, R, A part of the non-linear contribution of L d , L q is erroneously included in the value of , so it needs to be compensated. Thus, when adjusting the weights of the trained first neural network model, only R,
  • step S35 the network weight of the third neural network model (ie, the compensation module) is adjusted.
  • the total output error of the second neural network model reversely affects the output value of the third neural network model, thereby indirectly changing the internal weight of the third neural network model.
  • step S36 it is checked whether the observed data for training the model is exhausted.
  • step S31 If not exhausted, return to step S31 from step S36, so as to continue to extract the next training sample.
  • step S37 calculate the total loss function for all training samples (i.e. all observation data) and judge whether the output result of the loss function is less than the second preset condition, that is, for example, whether it is less than the second limit value.
  • step S37 If the output result of the loss function is greater than the second limit value, jump back from step S37 to step S31, so as to start a new iterative cycle.
  • the weight of the second neural network model can be output at this time in step S38, and based on this, R, the final value of .
  • the third neural network model can be used to output the fitted nonlinear characteristic part ⁇ L d0 , ⁇ L q0 of L d , L q , and in combination with the linear characteristic part L of L d , L q determined by the first neural network model In the case of d0 , L q0 , the function relationship between L d , L q and current can be finally determined.
  • Fig. 4 shows a schematic diagram of a device for parameter identification of a synchronous motor based on a neural network according to an exemplary embodiment of the present invention.
  • the device 1 includes a first recognition module 10 , a modification module 20 and a second recognition module 30 .
  • the first identification module 10 and the second identification module 30 are both used to identify the unknown parameters of the synchronous motor, they have different task priorities.
  • the first identification module 10 is mainly used to determine the linear characteristic part of the parameter to be identified Or the parameter value in the case of linear approximation of the parameter
  • the second identification module 20 is mainly used to determine the result of the parameter to be identified considering nonlinear factors, or to determine the result after nonlinear compensation.
  • the first identification module 10 is, for example, used to establish and train a first neural network model and determine intermediate results of parameters to be identified of the synchronous motor with the help of the trained first neural network model.
  • the modifying module 20 is used, for example, to form a second neural network model by adding a compensation module to the trained first neural network model.
  • the second identification module 20 is, for example, configured to train the second neural network model based on the intermediate results of the parameters to be identified, and determine the final result of the parameters to be identified by means of the trained second neural network model.
  • FIG. 5 shows a schematic diagram of operating points that need to be traversed for parameter identification by means of a lookup table method.
  • Fig. 6 shows a schematic schematic diagram of the nonlinear characteristics of the d-axis flux linkage of a synchronous motor.
  • the d-axis flux linkage should vary linearly with the corresponding current and thus should form a plane.
  • Figure 6 due to the saturation of the d-axis magnetic circuit, the relationship between the d-axis flux linkage and the d and q-axis currents is nonlinear, which is intuitively expressed as the irregular surface shown in Figure 6. This nonlinearity further causes the d-axis inductance to change with the current. If it continues to be treated as a constant, it will definitely affect the control accuracy and stability of the motor.
  • Fig. 7 shows an exemplary schematic diagram of a first neural network model used in the method according to the present invention.
  • the first neural network model 60 exemplarily includes a two-layer neuron structure, and each layer of neural network realizes full connection. Its input and output relationship is established by the following static voltage state equation of the synchronous motor in the d, q axis coordinate system:
  • Usd is the d-axis voltage of the synchronous motor
  • R is the stator resistance of the synchronous motor
  • ⁇ 1 is the electrical angular velocity of the synchronous motor
  • Lq is the q-axis inductance of the synchronous motor
  • i sq is the q-axis current of the synchronous motor
  • U sq is the q-axis voltage
  • isd is the d-axis current of the synchronous motor
  • L d is the d-axis inductance of the synchronous motor.
  • the q-axis current i sq and the d-axis current i sd and the electrical angular velocity ⁇ 1 of the synchronous motor are used as the input of the first neural network model 60 , taking the q-axis voltage U sq and the d-axis voltage U sd as the output of the first neural network model 60 .
  • the weights between the first input node and the two output nodes are R and ⁇ 1 respectively, and the weights between the second input node and the two output nodes are - ⁇ 1 and R , the weights between the third input node and the two output nodes are respectively and
  • ⁇ is 0.
  • R, L d , L q and The values of the modifiable weights belonging to the first neural network model 60 can be iteratively updated in back error propagation.
  • the first neural network model 60 will output the estimated value of the q-axis voltage and the d-axis voltage, where the following loss function can be constructed based on the difference between the measured voltage and the estimated value:
  • N is the number of known observations used
  • U sd is the d-axis voltage of the synchronous motor in the known observations
  • U sq is the q-axis voltage of the synchronous motor in the known observation data
  • the weight of the network is at least partially modified by means of the backpropagation algorithm until the value of the loss function J satisfies the first preset condition.
  • the parameters R, L d , L to be identified can be determined based on the internal weights of the first neural network model 60 q and intermediate results.
  • FIG. 8 shows an exemplary schematic diagram of a compensation module used in the method according to the invention.
  • the compensation module 70 in the form of a third neural network model exemplarily includes an input layer, an output layer and at least one hidden layer. Depending on the final desired accuracy and the amount of observed data, the number of hidden layers and the number of neurons included can be adjusted.
  • L q and L d are composed of two parts respectively, and their expressions are as follows:
  • L q , L d are q-axis inductance and d-axis inductance respectively
  • L d0 , L q0 represent the linear characteristic part of q-axis inductance and d-axis inductance respectively
  • ⁇ L d0 , ⁇ L q0 represent q-axis inductance and d-axis inductance respectively
  • the nonlinear characteristic part of . L d0 , L q0 have already been determined in the form of intermediate results, for example, by means of the first neural network model. Therefore, it is intended to determine ⁇ L d0 , ⁇ L q0 by means of the non-linear compensation module 70 .
  • the relationship between i sd , i sq and ⁇ L d0 , ⁇ L q0 can be directly established through the compensation module, namely , taking the q-axis current and the d-axis current isd and i sq as the input of the third neural network model, and taking the nonlinear characteristic parts ⁇ L d0 and ⁇ L q0 of the parameters to be identified as the output of the third neural network model.
  • Fig. 9 shows an exemplary schematic diagram of a second neural network model used in the method according to the present invention.
  • the compensation module 70 shown in FIG. 8 is added to the first neural network model 60 shown in FIG. 7 in such a way that the compensation module 70 shares part of the input layer with the first neural network model 60, that is, shares “d, q axis current isd , isq ′′, and the output layer of the compensation module 70 is fully connected to the output layer of the first neural network model 60 , thereby forming the second neural network model 80 as a whole.
  • the input of the second neural network model 80 is still the q-axis current i sq and the d-axis current isd and the electrical angular velocity ⁇ 1 of the synchronous motor, and the output of the second neural network model 80 is still the q-axis voltage U sq and d-axis voltage U sd .
  • the input-output relationship of the second neural network model 80 is currently established based on the following static voltage state equation:
  • Usd is the d-axis voltage of the synchronous motor
  • R is the stator resistance of the synchronous motor
  • ⁇ 1 is the electrical angular velocity of the synchronous motor
  • L q0 is the linear characteristic part of the q-axis inductance of the synchronous motor
  • ⁇ L q0 is the q
  • i sq is the q-axis current of the synchronous motor
  • ⁇ is the zero angle of the synchronous motor
  • U sq is the q-axis voltage
  • i sd is the d-axis current
  • L d0 is the linear characteristic part of the d-axis inductance
  • ⁇ L d0 is d Non-linear characteristic part of shaft inductance.
  • the second neural network model 80 When training the second neural network model 80 , for the first neural network model 60 , starting from the weight value of the first neural network model 60 at the end of the training, it is readjusted.
  • the d and q-axis inductance linear characteristic parts L d0 and L q0 are considered to be fixed values, so that the weights represented by these two will not participate in the second iteration training. Therefore, only R, value of .
  • the preset value of the internal weight of the network is changed, so that the output ⁇ L d0 and ⁇ L q0 can be changed.
  • N is the number of known observations used
  • U sd is the d-axis voltage of the synchronous motor in the known observations
  • U sq is the q-axis voltage of the synchronous motor in the known observation data
  • the internal weights corresponding to the first neural network model 60 can be used to determine the parameters to be identified R, of the final result.
  • the nonlinear characteristic parts ⁇ L d0 , ⁇ L q0 of the fitted L d , L q are output by means of the compensation module 70 , and the linear characteristic parts L d0 , L q0 in the intermediate result combining L d , L q
  • the functional relationship between L d , L q and d, q axis current can be finally determined.
  • Figures 10a-10b show schematic diagrams of the nonlinear characteristic part of the d-axis inductance and the nonlinear characteristic part of the q-axis inductance of a synchronous motor determined by the method according to the present invention as a function of current.

Landscapes

  • Engineering & Computer Science (AREA)
  • Power Engineering (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Control Of Electric Motors In General (AREA)
  • Control Of Ac Motors In General (AREA)

Abstract

一种基于神经网络的同步电机的参数辨识的方法，该方法包括以下步骤：（S1）建立并训练第一神经网络模型，借助经训练的第一神经网络模型确定同步电机的待辨识参数的中间结果；（S2）通过在经训练的第一神经网络模型中添加补偿模块形成第二神经网络模型；（S3）基于待辨识参数的中间结果对第二神经网络模型进行训练，借助经训练的第二神经网络模型确定待辨识参数的最终结果。该技术方案还涉及一种基于神经网络的同步电机的参数辨识的设备和一种计算机程序产品。

Description

基于神经网络的同步电机的参数辨识的方法 技术领域
本发明涉及一种基于神经网络的同步电机的参数辨识的方法、一种基于神经网络的同步电机的参数辨识的设备和一种计算机程序产品。
背景技术
同步电机由于其低成本、高效率和高稳定性等特性而在工业应用中越来越受到关注。然而，当同步电机带负载工作时，随着磁路饱和、温度漂移、交叉耦合等原因，使得传统线性模型不再能准确反映电机的真实动态特性，电机的诸多参数表现出非线性特性。在这种情况下，需要对同步电机进行参数辨识，以实现更好的控制性能。
在现有技术中，通常以查找表的形式对电机参数进行建模，但这需要遍历电机的全范围操作点，耗时且成本高。此外，还已知通过数学模型或特殊公式来拟合非线性电感表面的方法，然而这些数学模型或公式往往非常复杂，它们的系数也是参数化的，而且曲线绘制算法也需要相应的数据库支持。
在这种背景下，期待提供一种替代的电机参数辨识方法，以在有限操作点的情况下实现良好的参数拟合效果。
发明内容
本发明的目的在于提供一种基于神经网络的同步电机的参数辨识的方法、一种基于神经网络的同步电机的参数辨识的设备和一种计算机程序产品，以至少解决现有技术中的部分问题。
根据本发明的第一方面，提供一种基于神经网络的同步电机的参数辨识的方法，所述方法包括以下步骤：
S1)建立并训练第一神经网络模型，借助经训练的第一神经网络模型确定同步电机的待辨识参数的中间结果；
S2)通过在经训练的第一神经网络模型中添加补偿模块形成第二神经网络模型；以及
S3)基于待辨识参数的中间结果对第二神经网络模型进行训练，借助经训练的第二神经网络模型确定待辨识参数的最终结果。
本发明尤其包括以下技术构思：利用神经网络建立了电机的参数模型，在只有少量操作点并因此只有有限观测数据的情况下，利用人工神经网络的自学习能力和模糊逻辑对模型参数进行寻优估计，提高了参数辨识的效率和准确性。此外，提出基于两阶段的模型训练过程来完成参数估计，由此能够为这两个训练过程分配不同的任务侧重点，能够更加有针对性地对未知参数的不同特性进行良好拟合。
可选地，所述中间结果包括：
待辨识参数中的第一参数的线性近似值；以及
在对第一参数进行线性近似的情况下，待辨识参数中的第二参数的第一估计值。
可选地，所述最终结果包括：
待辨识参数中的第一参数的线性近似值被非线性补偿后的值；以及
在对第一参数的线性近似值进行非线性补偿之后，待辨识参数中的第二参数的第二估计值。
在此，尤其实现以下技术优点：通过借助第一神经网络模型和第二神经网络模型的参数辨识过程，实现了电机未知参数的线性特征和非线性特征的解耦分析，简化了辨识原理。
可选地，针对第一神经网络模型的训练和第二神经网络模型的训练使用不同的训练数据。
在此，尤其实现以下优点：可以根据经验选取磁路未饱和情况下的观测数据拟合未知参数的线性特性，并且选取磁路饱和情况下的观测数据进行非线性补偿，从而能够加速网络收敛过程。
可选地，所述同步电机的待辨识参数包括：同步电机的定子电阻、d轴电感、q轴电感和永磁体磁链。
在此，尤其实现以下技术优点：区别于借助查找表的参数识别方式，在此提出的方案能够实现多个电机参数的同步辨识，因此不需要借助附加 传感器对一些定值参数进行预先检测，节省了时间开销。
可选地，所述步骤S1包括：基于同步电机在d、q轴坐标系下的静态电压状态方程建立第一神经网络模型的输入-输出关系，其中，以q轴电流和d轴电流以及同步电机的电角速度作为第一神经网络模型的输入，以q轴电压和d轴电压作为第一神经网络模型的输出，至少部分地通过所述待辨识参数表征第一神经网络模型的权重。
在此，尤其实现以下技术优点：通过监督学习，人工神经网络能够具有很好的输入-输出映射关系，能够为新的输入产生合理的输出并在学习过程中不断对自身性能进行优化，在样本非常有限的情况下充分利用现有数据规律加快了参数识别过程。
可选地，所述步骤S1包括：
获取同步电机的已知观测数据，所述已知观测数据包括同步电机的已经在不同操作点处测得的d轴电流、q轴电流、d轴电压、q轴电压和电角速度；
借助已知观测数据对第一神经网络模型进行训练，其中，调整第一神经网络模型的权重，直至损失函数满足第一预设条件；以及
基于损失函数满足第一预设条件时的第一神经网络模型的权重确定同步电机的待辨识参数的中间结果。
在此，尤其实现以下技术优点：在此期间，系统的未知参数反映在内部权重上，从而避免了对被控对象的未知参数的直接辨识。
可选地，调整第一神经网络模型的权重包括：借助梯度下降算法、尤其反向传播算法对第一神经网络模型的权重进行迭代更新。
在此，尤其实现以下技术优点：通过上述算法能够基于已经掌握的输入-输出关系自动提取“合理的”求解规则，从而自动实现内部权重的确定，化解了内部机制复杂的映射问题，且能够使网络具有良好的泛化能力和容错能力。
可选地，所述步骤S2包括：建立第三神经网络模型作为所述补偿模块，以q轴电流和d轴电流为第三神经网络模型的输入，以待辨识参数的非线性特性部分作为第三神经网络模型的输出。
在此，尤其实现以下技术优点：通过引入补偿模块，一方面能够对参 数线性假设情况下纳入的非线性因素进行合理补偿，另一方面能够良好地为参数的非线性特性创建专属数学模型，从而使最终的参数辨识结果更加可靠。
可选地，所述步骤S2包括：通过如下方式将补偿模块添加到第一神经网络模型中：使补偿模块的输入层共享第一神经网络模型的部分输入层，并且使补偿模块的输出层连接到第一神经网络模型的输出层。
在此，尤其实现以下技术优点：通过以共享输入输出的方式将补偿模块嵌入到第一神经网络模型中，在第二神经网络模型的整体训练过程中，“带动”补偿模块进行被动训练。由此，使参数的未知非线性特征转换为模型的内部参数进行求解，在没有直接训练数据可用的情况下，通过网络整体的反向传播能力为补偿模块间接建立了输入-输出映射关系。
可选地，所述步骤S3包括：
借助已知观测数据对第二神经网络模型进行训练，其中，调整经训练的第一神经网络模型的权重以及补偿模块的权重，直至损失函数满足第二预设条件；以及
基于损失函数满足第二预设条件时的第一神经网络模型的权重以及补偿模块的权重确定同步电机的待辨识参数的最终结果。
在此，尤其实现以下技术优点：使第二神经网络模型的训练基于第一阶段的训练结果来进行，可以在待辨识参数的中间结果的基础上对其进行完善，提高了时间效率，加快了收敛进程。
可选地，在调整经训练的第一神经网络模型的权重时，保持待辨识参数中的第一参数的中间结果不变，并改变待辨识参数中的第二参数的中间结果。
在此，尤其实现以下技术优点：由于第一阶段的参数估计是在特定参数的线性近似下完成的，为了在补偿阶段将其与非线性特性部分良好区分开，可以继续认为该线性特性部分以常数存在并因此不参与进一步训练过程。但是，由于这种线性近似是不够准确的，所以有可能的是：有些定值参数的中间结果中错误地包含了特定参数的非线性贡献。为了对原本不属于这些定值参数的部分进行补偿，需要在第二阶段中对这些定值参数进行调整，以便能够剔除干扰项。
可选地，所述损失函数通过以下公式表示：
Figure PCTCN2021121272-appb-000001
其中，N是所使用的已知观测数据的数量，U sd是已知观测数据中的同步电机的d轴电压，
Figure PCTCN2021121272-appb-000002
是第一或第二神经网络模型输出的d轴电压，U sq是已知观测数据中的同步电机的q轴电压，
Figure PCTCN2021121272-appb-000003
是第一或第二神经网络模型输出的q轴电压。
在此，尤其实现以下技术优点：可以有效考虑并利用模型输出和实际输出之间的误差，使模型朝着逼近真实值的方向不断自学习，为训练过程提供了可靠评价指标和终止条件。
可选地，所述方法还包括以下步骤：
将待辨识参数的最终结果发送给同步电机的控制器，使控制器基于最终结果控制同步电机的运行；和/或
将训练完成的第二神经网络模型的权重发送给同步电机的控制器的内部神经网络，使内部神经网络在应用所述权重的情况下实时地根据同步电机的运行参数确定待辨识参数的在线最终结果，使控制器基于在线最终结果控制同步电机的运行。
在此，尤其实现以下技术优点：通过直接将电机参数发送给同步电机的控制器，可将参数生成过程保留在线下，因此，对于同步电机的控制器而言，确保了低的计算开销和硬件要求，在总体上实现了成本有利的控制方案。通过将内部神经网络集成在控制器中，并将训练好的权重值提供给该内部神经网络，可以使控制器根据不同工况实时地更新参数辨识结果，能够实现更可靠地电机控制。
根据本发明的第二方面，提供一种基于神经网络的同步电机的参数辨识的设备，所述设备用于执行根据本发明的第一方面所述的方法，所述设备包括：
第一辨识模块，其配置成能够建立并训练第一神经网络模型并借助经训练的第一神经网络模型确定同步电机的待辨识参数的中间结果；
修改模块，其配置成能够通过在经训练的第一神经网络模型中添加补偿模块形成第二神经网络模型；以及
第二辨识模块，其配置成能够基于待辨识参数的中间结果对第二神经网络模型进行训练，借助经训练的第二神经网络模型确定待辨识参数的最终结果。
根据本发明的第三方面，提供一种计算机程序产品，其中，所述计算机程序产品包括计算机程序，所述计算机程序用于在被计算机执行时实施根据本发明的第一方面所述的方法。
附图说明
下面，通过参看附图更详细地描述本发明，可以更好地理解本发明的原理、特点和优点。附图包括：
图1示出了根据本发明的一个示例性实施例的基于神经网络的同步电机的参数辨识的方法的流程图；
图2示出了图1所示方法的一个方法步骤的流程图；
图3示出了图1所示方法的两个方法步骤的流程图；
图4示出了根据本发明的一个示例性实施例的基于神经网络的同步电机的参数辨识的设备的示意图；
图5示出了借助查找表法进行参数辨识需要遍历的操作点的示意图；
图6示出了同步电机的d轴磁链的非线性特性的原理性示意图；
图7示出了在根据本发明的方法中所使用的第一神经网络模型的示例性示意图；
图8示出了在根据本发明的方法中所使用的补偿模块的示例性示意图；
图9示出了在根据本发明的方法中所使用的第二神经网络模型的示例性示意图；以及
图10a-图10b示出了借助根据本发明的方法确定的同步电机的d轴电感的非线性特性部分和q轴电感的非线性特性部分随电流变化的示意图。
具体实施方式
为了使本发明所要解决的技术问题、技术方案以及有益的技术效果更加清楚明白，以下将结合附图以及多个示例性实施例对本发明进行进一步详细说明。应当理解，此处所描述的具体实施例仅用于解释本发明，而不 是用于限定本发明的保护范围。
图1示出了根据本发明的一个示例性实施例的基于神经网络的同步电机的参数辨识的方法的流程图。
在步骤S1中，建立并训练第一神经网络模型，借助经训练的第一神经网络模型确定同步电机的待辨识参数的中间结果。
在本发明的意义上，同步电机例如包括同步磁阻电机、永磁同步电机、感应电机等其他类型的具有参数非线性特性的同步电机。
在本发明的意义上，同步电机的待辨识参数例如包括：同步电机的定子电阻R、d轴电感L d、q轴电感L q和/或永磁体磁链
Figure PCTCN2021121272-appb-000004
在这些参数中，在控制外界温度一定的情况下，可以认为定子电阻R和永磁体磁链
Figure PCTCN2021121272-appb-000005
是同步电机的固有参数并因此是不变的。然而由于无法忽略的铁心饱和现象以及材料因素，同步电机的d轴电感L d和q轴电感L q会随着电流的变化而发生变化，如果将其作为常数对待，必将影响电机的控制精度和稳定度。
在本发明的意义上，待辨识参数的中间结果例如包括：待辨识参数中的第一参数(例如d轴电感L d和q轴电感L q)的线性近似值。此外，该中间结果还包括：在对第一参数进行线性近似的情况下，待辨识参数中的第二参数(例如定子电阻R和永磁体磁链
Figure PCTCN2021121272-appb-000006
)的第一估计值。
在该步骤中，例如基于同步电机在d、q轴坐标系下的静态电压状态方程建立第一神经网络模型的输入-输出关系。以永磁同步电机为例，所使用的静态电压状态方程如下所示：
Figure PCTCN2021121272-appb-000007
Figure PCTCN2021121272-appb-000008
其中，U sd是同步电机的d轴电压，R是同步电机的定子电阻，ω 1是同步电机的电角速度，L q是同步电机的q轴电感，i sq是同步电机的q轴电流，
Figure PCTCN2021121272-appb-000009
是同步电机的永磁体磁链的幅值，θ是同步电机的零位角，U sq是q轴电压，i sd是同步电机的d轴电流，L d是同步电机的d轴电感。
为了建立第一神经网络模型，例如以q轴电流和d轴电流以及同步电机的角速度作为第一神经网络模型的输入，以q轴电压和d轴电压作为第一神经网络模型的输出，以同步电机的待辨识参数R,L d,L q、
Figure PCTCN2021121272-appb-000010
表征第一神 经网络模型的可修改权重。
在步骤S2中，通过在经训练的第一神经网络模型中添加补偿模块形成第二神经网络模型。为了借助补偿模块描述待辨识参数的非线性特性，例如建立第三神经网络模型作为补偿模块，以q轴电流和d轴电流为第三神经网络模型的输入，以待辨识的d轴电感和q轴电感的非线性特性部分作为第三神经网络模型的输出。
在步骤S3中，基于待辨识参数的中间结果对第二神经网络模型进行训练，借助经训练的第二神经网络模型确定待辨识参数的最终结果。本发明的意义上，最终结果表示待辨识参数的能够表征其实际取值或实际函数关系的结果。待辨识参数的最终结果例如包括：待辨识参数中的第一参数(例如d轴电感和q轴电感)的线性近似值被非线性补偿后的值。此外，该最终结果还包括：在对第一参数的线性近似值进行非线性补偿之后，待辨识参数中的第二参数(例如定子电阻R和永磁体磁链
Figure PCTCN2021121272-appb-000011
)的第二估计值，与之前确定的第一估计值相比，第二参数的第二估计值中基本上不再具有或仅较少地具有第一参数的非线性成分贡献。通过排除这一部分，参数的第二估计值更逼近于其真实值。在该步骤中，例如借助第二神经网络模型对已经确定的待辨识参数的各中间结果进行更新，根据更新的结果能够确定各未知参数的辨识方程。
图2示出了图1所示方法的一个方法步骤的流程图。在该实施例中，图1中的方法步骤S1示例性地包括步骤S11-S18。
在步骤S11中，获取观测数据。以永磁同步电机为例，在此例如首先控制永磁同步电机稳定运行。然后通过改变参考电流、负载扭矩、电机转速等条件，以在尽可能宽的工作范围内遍历N组不同的操作点，在这些操作点处记录电机的d轴电压U sd、q轴电压U sq、d轴电流i sd、q轴电流i sq和电角速度ω 1。在此，N的取值可根据时间开销和期望的模型精度自由调整，通常N处于15～30的范围内。
在步骤S12中，从观测数据中提取一组数据中的d轴电流i sd、q轴电流i sq和电角速度ω 1，以正向传播方式输入到第一神经网络模型中。
在步骤S13中，按照各层神经元之间的权重的初始估计值计算第一神经网络模型的输出。在此，第一神经网络模型权重的至少部分地描述电机 的待辨识参数R,L d,L q、
Figure PCTCN2021121272-appb-000012
的在不考虑电机非线性特性情况下的取值，这些权重的初始估计值可由工程专家根据经验得出，也可以是外部随机输入的任一符合数学和物理意义的取值。
在步骤S14中，计算输出误差。在此，例如分别计算第一神经网络模型输出的d轴电压
Figure PCTCN2021121272-appb-000013
与步骤S12中提取的观测数据中的d轴电压U sd之间的偏差，并计算第一神经网络模型输出的q轴电压
Figure PCTCN2021121272-appb-000014
与所提取的观测数据中的q轴电压U sq之间的偏差。
在步骤S15中，基于所得到的误差，借助反向传播算法调整连接权重。在此，通过误差更新值通过反向传播逐层达到第一层，所有权重在反向传播结束时一起更新。在此，视所要求的模型精度和时间开销而定，可自由调整反向传播算法中所使用的学习率。
在步骤S16中，检查用于训练模型的观测数据是否用尽。在此例如检查已经遍历的观测数据的数量是否大于等于N。
如果未用尽，从步骤S16回到步骤S12，以便继续提取下一个训练样本。
如果已经用尽，则在步骤S17中针对所有训练样本(即所有观测数据)计算总的损失函数并判断该损失函数的输出结果是否满足第一预设条件。在此，例如判断损失函数的输出结果是否小于第一极限值。
如果损失函数的输出结果大于第一极限值，则表示模型目前的表现尚未达到标准，需要继续对其进行训练，于是从步骤S17跳转回到步骤S12，以开始新的迭代循环。
如果损失函数的输出结果小于第一极限值，则表示模型在预期结果方面已经有符合的要求的表现。在这种情况下，可以在步骤S18中，将第一神经网络模型中的当前的权重输出，并基于此确定同步电机的待辨识参数的线性近似部分。值得注意的是，由于在该阶段不考虑同步电机的非线性因素，因此在此确定的各参数以常数的形式存在。对于原本应受到非线性影响的第一参数(例如d、q轴电感)而言，现阶段确定的结果可以看作是该参数在所有观测数据上的平均值，因此并不是绝对准确的，需要通过后续的非线性补偿过程对其进行修正。对于原本应是定值的第二参数(例如定子电阻、永磁体磁链)而言，现阶段确定的结果同时还将第一参数的一 部分非线性贡献量考虑在内。
图3示出了图1所示方法的两个方法步骤的流程图。在该实施例中，图1中的方法步骤S3示例性地包括步骤S31-S38。
在步骤S2中，将第三神经网络模型形式的补偿模块添加到经训练的第一神经网络模型中，由此形成了第二神经网络模型。
在步骤S31中，从已获取的观测数据中提取一组数据中的d轴电流i sd、q轴电流i sq和电角速度ω 1，以正向传播方式输入到第二神经网络模型中。在此，例如将d轴电流i sd、q轴电流i sq和电角速度ω 1输入到经训练的第一神经网络模型中，并且将d轴电流i sd、q轴电流i sq输入到补偿模块中。
在步骤S32中，计算第二神经网络模型的输出。在此，例如根据各层神经元之间的权重关系计算相应的电压
Figure PCTCN2021121272-appb-000015
在步骤S33中，计算第二神经网络模型输出的电压与观测数据中的电压之间的误差。
在步骤S34中，基于该误差，借助反向传播算法调整经训练的第一神经网络模型的权重。第一神经网络模型的权重仍至少部分地描述电机参数R,L d,L q、
Figure PCTCN2021121272-appb-000016
然而由于L d,L q的中间结果L d0,L q0已经表征线性特性部分，因此不再使其参与进一步训练过程。但是，R、
Figure PCTCN2021121272-appb-000017
的值中错误地纳入了L d,L q的一部分非线性贡献，因此需要对其进行补偿。于是，在调整经训练的第一神经网络模型的权重时，仅改变R、
Figure PCTCN2021121272-appb-000018
在步骤S35中，调整第三神经网络模型(即补偿模块)的网络权重。在此，例如根据第二神经网络模型的总的输出误差反向影响第三神经网络模型的输出取值，从而间接改变第三神经网络模型的内部权重。
在步骤S36中，检查用于训练模型的观测数据是否用尽。
如果未用尽，在从步骤S36回到步骤S31，以便继续提取下一个训练样本。
如果已经用尽，则在步骤S37中针对所有训练样本(即所有观测数据)计算总的损失函数并判断该损失函数的输出结果是否小于满足第二预设条件，即，例如是否小于第二极限值。
如果损失函数的输出结果大于第二极限值，则从步骤S37跳转回到步骤S31，以便开始新的迭代循环。
如果损失函数的输出结果小于第二极限值，则表示模型在预期结果方面已经有符合的要求的表现。在这种情况下，可以在步骤S38中将第二神经网络模型此时的权重输出，并基于此确定同步电机的待辨识参数中的R、
Figure PCTCN2021121272-appb-000019
的最终取值。此外，还可借助第三神经网络模型输出拟合的L d,L q的非线性特性部分ΔL d0,ΔL q0，在结合借助第一神经网络模型确定的L d,L q的线性特性部分L d0,L q0的情况下，可以最终确定L d,L q与电流的函数关系。
图4示出了根据本发明的一个示例性实施例的基于神经网络的同步电机的参数辨识的设备的示意图。
如图4所示，设备1设备包括第一辨识模块10、修改模块20和第二辨识模块30。在此，第一辨识模块10和第二辨识模块30虽然都用于辨识同步电机的未知参数，但却具有不同的任务重点，例如第一辨识模块10主要用于确定待辨识参数的线性特性部分或者说在对参数进行线性近似情况下的参数值，而第二辨识模块20则主要用于确定待辨识参数的考虑非线性因素情况下的结果，或者说确定非线性补偿之后的结果。
具体地，第一辨识模块10例如用于建立并训练第一神经网络模型并借助经训练的第一神经网络模型确定同步电机的待辨识参数的中间结果。
修改模块20例如用于通过在经训练的第一神经网络模型中添加补偿模块形成第二神经网络模型。
第二辨识模块20例如用于基于待辨识参数的中间结果对第二神经网络模型进行训练，借助经训练的第二神经网络模型确定待辨识参数的最终结果。
图5示出了借助查找表法进行参数辨识需要遍历的操作点的示意图。
为了全面地反映同步电机的特定参数(例如d轴电感和q轴电感)的非线性特性，需要使操作点尽可能全面地覆盖同步电机的整个工作范围，因此从图5可看出，需要采集并测试非常多的操作点。此外，由于磁路饱和现象在小电流处较为显著，因此尤其需要密集地在小电流处进一步提高采样率，这极大地增加了参数辨识的时间开销，而且受限于测量软/硬件的精度，采样数据本身也具有一定不准确性。
图6示出了同步电机的d轴磁链的非线性特性的原理性示意图。
在理想情况下，d轴磁链应当与相应电流成线性变化，因此应当形成平 面。然而从图6可看出，d轴磁路由于出现饱和而导致d轴磁链与d、q轴电流呈现非线性关系，这直观地表现为图6所示的不规则曲面。这种非线性进一步导致d轴电感会随电流变化而发生变化，如果将其继续作为常数对待，必将影响电机的控制精度和稳定度。
图7示出了在根据本发明的方法中所使用的第一神经网络模型的示例性示意图。
第一神经网络模型60示例性地包括两层神经元结构，各层神经网络实现全连接。其输入和输出关系通过同步电机在d、q轴坐标系下的以下静态电压状态方程建立：
Figure PCTCN2021121272-appb-000020
Figure PCTCN2021121272-appb-000021
其中，U sd是同步电机的d轴电压，R是同步电机的定子电阻，ω 1是同步电机的电角速度，L q是同步电机的q轴电感，i sq是同步电机的q轴电流，
Figure PCTCN2021121272-appb-000022
是同步电机的永磁体磁链的幅值，θ是同步电机的零位角，U sq是q轴电压，i sd是同步电机的d轴电流，L d是同步电机的d轴电感。
为了借助该第一神经网络模型60确定同步电机的待辨识参数的线性特性部分，以q轴电流i sq和d轴电流i sd以及同步电机的电角速度ω 1作为第一神经网络模型60的输入，以q轴电压U sq和d轴电压U sd作为第一神经网络模型60的输出。基于静态电压状态方程定义的映射关系，第一输入节点与两个输出节点之间的权重分别为R和ω 1，第二输入节点与两个输出节点之间的权重分别为-ω 1和R，第三输入节点与两个输出节点之间的权重分别为
Figure PCTCN2021121272-appb-000023
和
Figure PCTCN2021121272-appb-000024
在电机的零位角被正确标定的情况下，θ为0。在此，R,L d,L q和
Figure PCTCN2021121272-appb-000025
属于第一神经网络模型60的可修改权重，其值可在反向误差传播中被迭代更新。
针对每个给定输入，第一神经网络模型60会输出对q轴电压和d轴电压的估计值，在此可基于实测电压与估计值之间的差值构造如下损失函数：
Figure PCTCN2021121272-appb-000026
其中，N是所使用的已知观测数据的数量，U sd是已知观测数据中的同 步电机的d轴电压，
Figure PCTCN2021121272-appb-000027
是第一网络模型60输出的d轴电压，U sq是已知观测数据中的同步电机的q轴电压，
Figure PCTCN2021121272-appb-000028
是第一网络模型60输出的q轴电压。
借助反向传播算法至少部分地修改网络的权重，直到损失函数J的值满足第一预设条件，这时可基于第一神经网络模型60的内部权重确定各待辨识参数R,L d,L q和
Figure PCTCN2021121272-appb-000029
的中间结果。
图8示出了在根据本发明的方法中所使用的补偿模块的示例性示意图。
第三神经网络模型形式的补偿模块70示例性地包括输入层、输出层和至少一个隐含层。视最终期望的精度以及观测数据量而定，可对隐含层数量和所包含的神经元数量进行调整。为了拟合q轴电感L q和d轴电感L d的非线性特性，在此例如假设L q、L d分别由两部分组成，其表达式如下所示：
L d＝L d0+ΔL d0
L q＝L q0+ΔL q0
其中，L q、L d分别是q轴电感和d轴电感，L d0、L q0分别表示q轴电感和d轴电感的线性特性部分，ΔL d0、ΔL q0分别表示q轴电感和d轴电感的非线性特性部分。L d0、L q0例如已经借助第一神经网络模型以中间结果的形式确定。因此，旨在借助非线性补偿模块70确定ΔL d0、ΔL q0。由于同步电机的磁路饱和与d、q轴电流i sd、i sq高度相关，基于这一认知，可通过补偿模块直接建立i sd、i sq与ΔL d0、ΔL q0之间的关系，即，以q轴电流和d轴电流i sd、i sq为第三神经网络模型的输入，以待辨识参数的非线性特性部分ΔL d0、ΔL q0作为第三神经网络模型的输出。
图9示出了在根据本发明的方法中所使用的第二神经网络模型的示例性示意图。
在此，图8所示的补偿模块70被如此添加到图7所示的第一神经网络模型60中，使得补偿模块70与第一神经网络模型60共享部分输入层，即共享“d、q轴电流i sd、i sq”，并且补偿模块70的输出层全连接到第一神经网络模型60的输出层，由此在总体上形成第二神经网络模型80。
在总体上，第二神经网络模型80的输入仍是q轴电流i sq和d轴电流i sd以及同步电机的电角速度ω 1，第二神经网络模型80的输出仍是q轴电 压U sq和d轴电压U sd。然而由于考虑了非线性因素，目前基于以下静态电压状态方程建立第二神经网络模型80的输入-输出关系：
Figure PCTCN2021121272-appb-000030
Figure PCTCN2021121272-appb-000031
其中，U sd是同步电机的d轴电压，R是同步电机的定子电阻，ω 1是同步电机的电角速度，L q0是同步电机的q轴电感的线性特性部分，ΔL q0是同步电机的q轴电感的非线性特性部分，i sq是同步电机的q轴电流，
Figure PCTCN2021121272-appb-000032
是同步电机的永磁体磁链的幅值，θ是同步电机的零位角，U sq是q轴电压，i sd是d轴电流，L d0是d轴电感的线性特性部分，ΔL d0是d轴电感的非线性特性部分。
在对第二神经网络模型80进行训练时，针对第一神经网络模型60，从第一神经网络模型60训练结束时的权重取值出发，对其进行再调整。在此，为了便于通过非线性特性部分进行补偿，继续认为d、q轴电感线性特性部分L d0,L q0为固定值，使通过这两者表征的权重不再参与第二次迭代训练。因此，在调整过程中仅改变R、
Figure PCTCN2021121272-appb-000033
的取值。
针对补偿模块70(即第三神经网络模型)，改变该网络内部权重的预设值，从而能够改变输出的ΔL d0、ΔL q0。
持续这种迭代权重更新过程，直至以下损失函数收敛：
Figure PCTCN2021121272-appb-000034
其中，N是所使用的已知观测数据的数量，U sd是已知观测数据中的同步电机的d轴电压，
Figure PCTCN2021121272-appb-000035
是第二神经网络模型输出的d轴电压，U sq是已知观测数据中的同步电机的q轴电压，
Figure PCTCN2021121272-appb-000036
是第二神经网络模型输出的q轴电压。
第二神经网络模型80训练完成时，可通过第一神经网络模型60对应的内部权重确定待辨识参数R、
Figure PCTCN2021121272-appb-000037
的最终结果。此外，在训练完成时借助补偿模块70输出拟合的L d,L q的非线性特性部分ΔL d0、ΔL q0，在结合L d,L q的中间结果中的线性特性部分L d0,L q0的情况下，可以最终确定L d,L q与d、q轴电流的函数关系。
图10a-图10b示出了借助根据本发明的方法确定的同步电机的d轴电 感的非线性特性部分和q轴电感的非线性特性部分随电流变化的示意图。
由图10a-10b可以看出，借助补偿模块最终拟合出d、q轴电感的非线性特性部分ΔL d0、ΔL q0随d、q轴电流i sd、i sq变化形成曲面图。基于所使用的神经网络技术，能够基于有限观测数据点对曲面的未知部分进行智能推理和判断，从而自动插补和重构出残缺数据，以恢复ΔL d0、ΔL q0与电流之间的完整函数关系。由拟合结果可见，d、q轴电感的非线性特性部分并不是常数，而是由于同步电机的磁路饱和等现象受到磁链非线性影响，并因此随着电流的变化而发生变化，尤其在电流i sd、i sq的值较小时，这种非线性特性更为明显。
尽管这里详细描述了本发明的特定实施方式，但它们仅仅是为了解释的目的而给出的，而不应认为它们对本发明的范围构成限制。在不脱离本发明精神和范围的前提下，各种替换、变更和改造可被构想出来。

Claims (15)

  1. 一种基于神经网络的同步电机的参数辨识的方法，所述方法包括以下步骤：
    S1)建立并训练第一神经网络模型(60)，借助经训练的第一神经网络模型(60)确定同步电机的待辨识参数的中间结果；
    S2)通过在经训练的第一神经网络模型(60)中添加补偿模块(70)形成第二神经网络模型(80)；以及
    S3)基于待辨识参数的中间结果对第二神经网络模型(80)进行训练，借助经训练的第二神经网络模型(80)确定待辨识参数的最终结果。
  2. 根据权利要求1所述的方法，其中，所述中间结果包括：
    待辨识参数中的第一参数的线性近似值；以及
    在对第一参数进行线性近似的情况下，待辨识参数中的第二参数的第一估计值。
  3. 根据权利要求1或2所述的方法，其中，所述最终结果包括：
    待辨识参数中的第一参数的线性近似值被非线性补偿后的值；以及
    在对第一参数的线性近似值进行非线性补偿之后，待辨识参数中的第二参数的第二估计值。
  4. 根据权利要求1至3中任一项所述的方法，其中，所述同步电机的待辨识参数包括同步电机的定子电阻、d轴电感、q轴电感和永磁体磁链。
  5. 根据权利要求1至4中任一项所述的方法，其中，所述步骤S1包括：
    基于同步电机在d、q轴坐标系下的静态电压状态方程建立第一神经网络模型(60)的输入-输出关系，其中，以q轴电流和d轴电流以及同步电机的电角速度作为第一神经网络模型(60)的输入，以q轴电压和d轴电压作为第一神经网络模型(60)的输出，至少部分地通过所述待辨识参数 表征第一神经网络模型(60)的权重。
  6. 根据权利要求5所述的方法，其中，所述步骤S1包括：
    获取同步电机的已知观测数据，所述已知观测数据包括同步电机的已经在不同操作点处测得的d轴电流、q轴电流、d轴电压、q轴电压和电角速度；
    借助已知观测数据对第一神经网络模型(60)进行训练，其中，调整第一神经网络模型(60)的权重，直至损失函数满足第一预设条件；以及
    基于损失函数满足第一预设条件时的第一神经网络模型(60)的权重确定同步电机的待辨识参数的中间结果。
  7. 根据权利要求6所述的方法，其中，调整第一神经网络模型(60)的权重包括：借助梯度下降算法、尤其反向传播算法对第一神经网络模型(60)的权重进行迭代更新。
  8. 根据权利要求1至7中任一项所述的方法，其中，所述步骤S2包括：
    建立第三神经网络模型作为所述补偿模块(70)，以q轴电流和d轴电流作为第三神经网络模型的输入，以待辨识参数中的第一参数的非线性特性部分作为第三神经网络模型的输出。
  9. 根据权利要求1至8中任一项所述的方法，其中，所述步骤S2包括：
    通过如下方式将补偿模块(70)添加到第一神经网络模型(60)中：使补偿模块(70)输入层共享第一神经网络模型(60)的部分输入层，并且使补偿模块(70)的输出层连接到第一神经网络模型(60)的输出层。
  10. 根据权利要求1至9中任一项所述的方法，其中，所述步骤S3包括：
    借助已知观测数据对第二神经网络模型(80)进行训练，其中，调整 经训练的第一神经网络模型(60)的权重以及补偿模块(70)的权重，直至损失函数满足第二预设条件；以及
    基于损失函数满足第二预设条件时的第一神经网络模型(60)的权重以及补偿模块(70)的权重确定同步电机的待辨识参数的最终结果。
  11. 根据权利要求10所述的方法，其中，在调整经训练的第一神经网络模型(60)的权重时，保持待辨识参数中的第一参数的中间结果不变，并且改变待辨识参数中的第二参数的中间结果。
  12. 根据权利要求6或11所述的方法，其中，所述损失函数通过以下公式表示：
    Figure PCTCN2021121272-appb-100001
    其中，N是所使用的已知观测数据的数量，U sd是已知观测数据中的同步电机的d轴电压，
    Figure PCTCN2021121272-appb-100002
    是第一神经网络模型(60)或第二神经网络模型(80)输出的d轴电压，U sq是已知观测数据中的同步电机的q轴电压，
    Figure PCTCN2021121272-appb-100003
    是第一神经网络模型(60)或第二神经网络模型(80)输出的q轴电压。
  13. 根据权利要求1至12中任一项所述的方法，其中，所述方法还包括以下步骤：
    将待辨识参数的最终结果发送给同步电机的控制器，使控制器基于所述最终结果控制同步电机的运行；和/或
    将训练完成的第二神经网络模型(80)的权重发送给同步电机的控制器的内部神经网络，使内部神经网络在应用第二神经网络模型(80)的权重的情况下，实时地根据同步电机的运行参数确定待辨识参数的在线最终结果，使控制器基于在线最终结果控制同步电机的运行。
  14. 一种基于神经网络的同步电机的参数辨识的设备(1)，所述设备(1)用于执行根据权利要求1至13中任一项所述的方法，所述设备(1)包括：
    第一辨识模块(10)，其配置成能够建立并训练第一神经网络模型(60) 并借助经训练的第一神经网络模型(60)确定同步电机的待辨识参数的中间结果；
    修改模块(20)，其配置成能够通过在经训练的第一神经网络模型(60)中添加补偿模块(70)形成第二神经网络模型(80)；以及
    第二辨识模块(30)，其配置成能够基于待辨识参数的中间结果对第二神经网络模型(80)进行训练，借助经训练的第二神经网络模型(80)确定待辨识参数的最终结果。
  15. 一种计算机程序产品，其中，所述计算机程序产品包括计算机程序，所述计算机程序用于在被计算机执行时实施根据权利要求1至13中任一项所述的方法。
PCT/CN2021/121272 2021-09-28 2021-09-28 基于神经网络的同步电机的参数辨识的方法 WO2023050050A1 (zh)

Priority Applications (3)

Application Number Priority Date Filing Date Title
DE112021007423.7T DE112021007423T5 (de) 2021-09-28 2021-09-28 Verfahren zur Parameteridentifizierung eines Synchronmotors basierend auf einem neuronalen Netzwerk
CN202180102774.3A CN118077134A (zh) 2021-09-28 2021-09-28 基于神经网络的同步电机的参数辨识的方法
PCT/CN2021/121272 WO2023050050A1 (zh) 2021-09-28 2021-09-28 基于神经网络的同步电机的参数辨识的方法

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2021/121272 WO2023050050A1 (zh) 2021-09-28 2021-09-28 基于神经网络的同步电机的参数辨识的方法

Publications (1)

Publication Number Publication Date
WO2023050050A1 true WO2023050050A1 (zh) 2023-04-06

Family

ID=85780976

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/121272 WO2023050050A1 (zh) 2021-09-28 2021-09-28 基于神经网络的同步电机的参数辨识的方法

Country Status (3)

Country Link
CN (1) CN118077134A (zh)
DE (1) DE112021007423T5 (zh)
WO (1) WO2023050050A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118278292A (zh) * 2024-06-03 2024-07-02 山东大学 一种pmsg阻抗参数自动辨识方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102497156A (zh) * 2011-12-27 2012-06-13 东南大学 永磁同步电机速度环的神经网络自校正控制方法
US20150039545A1 (en) * 2013-08-05 2015-02-05 City University of London Systems, methods and devices for vector control of permanent magnet synchronous machines using artificial neural networks
CN108228975A (zh) * 2017-12-14 2018-06-29 北京理工大学 电机伺服系统参数辨识方法以及消隙控制方法
CN111313774A (zh) * 2020-02-25 2020-06-19 华南理工大学 一种基于nlms算法的永磁同步电机参数在线辨识方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102497156A (zh) * 2011-12-27 2012-06-13 东南大学 永磁同步电机速度环的神经网络自校正控制方法
US20150039545A1 (en) * 2013-08-05 2015-02-05 City University of London Systems, methods and devices for vector control of permanent magnet synchronous machines using artificial neural networks
CN108228975A (zh) * 2017-12-14 2018-06-29 北京理工大学 电机伺服系统参数辨识方法以及消隙控制方法
CN111313774A (zh) * 2020-02-25 2020-06-19 华南理工大学 一种基于nlms算法的永磁同步电机参数在线辨识方法

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118278292A (zh) * 2024-06-03 2024-07-02 山东大学 一种pmsg阻抗参数自动辨识方法

Also Published As

Publication number Publication date
DE112021007423T5 (de) 2024-01-11
CN118077134A (zh) 2024-05-24

Similar Documents

Publication Publication Date Title
Korzonek et al. A review on MRAS-type speed estimators for reliable and efficient induction motor drives
Wei et al. Neural-network-based adaptive optimal tracking control scheme for discrete-time nonlinear systems with approximation errors
Karanayil et al. Stator and rotor resistance observers for induction motor drive using fuzzy logic and artificial neural networks
CN101938246B (zh) 无速度传感器电机转速的模糊融合辨识方法
Ben Regaya et al. Electric drive control with rotor resistance and rotor speed observers based on fuzzy logic
CN105680746A (zh) 一种利用在线粒子群优化算法来设计永磁同步电动机电流和速度控制器pi参数的方法
WO2023050050A1 (zh) 基于神经网络的同步电机的参数辨识的方法
CN109034390B (zh) 基于bp神经网络三维磁特性测量的相角幅值pid自适应方法
Zhao et al. Neuroadaptive dynamic surface control for induction motors stochastic system based on reduced-order observer
Li et al. Extended State Observer Based Adaptive Back‐Stepping Sliding Mode Control of Electronic Throttle in Transportation Cyber‐Physical Systems
Hussain et al. Neural network observer design for sensorless control of induction motor drive
CN111030535B (zh) 一种异步电机电感参数在线辨识方法
CN110649851B (zh) 异步电机多参数解耦在线辨识方法
Gundogdu et al. Comparison of Si-ANN and extended Kalman filter-based sensorless speed controls of a DC motor
Djelamda et al. Field-oriented control based on adaptive neuro-fuzzy inference system for PMSM dedicated to electric vehicle
Li et al. A flexible current tracking control of sensorless induction motors via adaptive observer
Hozefa et al. Feedback linearization technique for dc series motor with lmi based observer and drem
CN115102444B (zh) 一种永磁同步电机自适应积分滑模预测控制方法
Ge et al. HGO and neural network based integral sliding mode control for PMSMs with uncertainty
Vukadinovic et al. Stator resistance identification based on neural and fuzzy logic principles in an induction motor drive
Sabouni et al. Adaptive backstepping controller design based on neural network for pmsm speed control
Praharsena et al. Stator Flux Estimator Using Feed-Forward Neural Network for Evaluating Hysteresis Loss Curve in Three Phase Induction Motor
Purwanto et al. Application of vector control method for developing ANFIS observer as speed sensor for induction motor speed control in electric vehicle
Viswa Teja et al. Analysis of various speed control methods for PMSM drive-based electric vehicles
Kumar et al. Neuro-fuzzy based estimation of rotor flux for Electric Vehicle operating under partial loading

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21958649

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 112021007423

Country of ref document: DE

ENP Entry into the national phase

Ref document number: 2024519101

Country of ref document: JP

Kind code of ref document: A

WWE Wipo information: entry into national phase

Ref document number: 202180102774.3

Country of ref document: CN