Method and device for parameter identification of complex system using parallel neural network

Abstract

The present invention relates to the field of neural network application technology, and more specifically, to a method and device for parameter identification of complex systems using parallel neural networks. The present invention proposes a method for parameter identification of complex systems using parallel neural networks; the method of the present invention uses a self-designed parallel neural network, which uses multiple influencing factors as the input of the input layer I1, connects the input layer I1 and the system parameter layer P3 through the deep neural network of the intermediate layer M2, identifies the system parameters through the system parameter layer P3, and uses the target parameters as the output of the output layer O4, which can achieve accurate identification of system parameters under variable working conditions and provide a basis for state observation and precise control of complex systems. The present invention solves the problem of parameter identification of complex systems affected by multiple factors.

Description

Technical Field

[0001] The present invention relates to the field of neural network application technology, and in particular to: 1. a method for performing parameter identification on a complex system using a parallel neural network; 2. a device for performing parameter identification on a complex system using a parallel neural network; wherein the complex system is affected by multiple influencing factors. Background Art

[0002] In complex industrial automation systems such as automobiles and aerospace, in order to achieve accurate observation and control of system state quantities, it is necessary to obtain the specific values ​​of multiple target parameters of the system in advance. However, these target parameters cannot be directly measured by sensors, so the inventors considered using parameter identification, that is, identifying the system parameters associated with the target parameters, thereby indirectly obtaining the specific values ​​of the target parameters. However, after research, it was found that: due to the wide operating range of most complex systems, the corresponding multiple system parameters will change in real time with the working conditions, and the influencing factors and change patterns of each system parameter are different, which poses a huge challenge to parameter identification.

[0003] To facilitate understanding, the rotor temperature of an induction motor in a new energy vehicle is used as an example to illustrate the difficulties of identifying the parameters of the aforementioned complex system:

[0004] In induction motors, the rotor rotates at high speeds, making the corresponding rotor temperature, a target parameter, difficult to obtain directly through sensors. However, this target parameter is crucial for precise motor control and safe performance. Therefore, in engineering applications, the rotor temperature is typically indirectly obtained by establishing a rotor temperature thermal network system. This requires pre-identification of system parameters such as rotor losses and rotor thermal resistance. Combined with the measurable temperatures of other components, the rotor temperature is calculated and used as the observed value.

[0005] However, automotive motors have a wide operating range, and the various system parameters (such as thermal resistance and losses) in their rotor temperature thermal network system are different under different motor operating conditions. In addition, the influencing factors of each system parameter are also different. For example, the thermal resistance between the stator and rotor is mainly affected by the motor speed, while the loss of the motor rotor is affected by multiple factors such as speed, stator current, and rotor temperature. This makes it extremely difficult to accurately identify the system parameters under different operating conditions.

[0006] In general, existing parameter identification lacks effective tools and cannot achieve accurate identification of system parameters for complex systems.

[0007] Therefore, the inventors designed a parallel neural network and applied it to the target complex system, achieving accurate identification of system parameters under variable working conditions and also enabling data acquisition of target parameters. Summary of the Invention

[0008] Based on this, it is necessary to provide a method and device for parameter identification of complex systems using parallel neural networks to address the problem that existing technologies cannot accurately identify system parameters of complex systems affected by multiple influencing factors.

[0009] The present invention is achieved by adopting the following technical solutions:

[0010] In a first aspect, the present invention provides a method for identifying parameters of a complex system using a parallel neural network, which is used to identify system parameters of a target complex system that is affected by multiple influencing factors.

[0011] Methods for parameter identification of complex systems using parallel neural networks include:

[0012] S1, using the sample data set of the target complex system to train the pre-built parallel neural network, obtain the parallel neural network with the best performance, and use it as the identification network model;

[0013] The parallel neural network includes: input layer I1, middle layer M2, system parameter layer P3, and output layer O4.

[0014] The input layer I1 is used to input several influencing factors into the middle layer M2; the middle layer M2 is several deep neural networks, used to connect the input layer I1 and the system parameter layer P3; the system parameter layer P3 is used to output several system parameters obtained by the middle layer M2; the output layer O4 has two types of inputs, one type of input is the system parameters obtained by the system parameter layer P3, and the other type of input is several system state quantities; the output layer O4 is used to combine the system parameters and system state quantities and calculate several target parameters based on the corresponding calculation relationship.

[0015] The input layer I1, the middle layer M2, and the system parameter layer P3 constitute multiple parallel networks; the number of parallel networks is the same as the number of system parameters, and one parallel network has a deep neural network; each parallel network is independent and there is no connection relationship between them; the specific types of influencing factors, system parameters, target parameters, and system state quantities are obtained based on the target complex system; for any parallel network, the influencing factors of its input directly affect the system parameters of its output; the calculation relationship between system parameters, system state quantities and target parameters is constructed based on the physical model of the target complex system.

[0016] S2, input the specific value of the influencing factor under a certain working condition to the input layer I1 of the identification network model, and the system parameter layer P3 of the identification network model outputs the specific value corresponding to the system parameter, which is the identification value of the system parameter.

[0017] This method of using a parallel neural network to perform parameter identification on a complex system implements the method or process according to an embodiment of the present disclosure.

[0018] In a second aspect, the present invention discloses a device for performing parameter identification on a complex system using a parallel neural network, which uses the method for performing parameter identification on a complex system using a parallel neural network disclosed in the first aspect.

[0019] The device for performing parameter identification on a complex system by using a parallel neural network comprises: an identification network acquisition module and a complex system processing module.

[0020] The identification network acquisition module is used to train the pre-built parallel neural network using the sample data set of the target complex system to obtain the parallel neural network with the best performance as the identification network model.

[0021] The complex system processing module is used to input the specific values ​​of the influencing factors under a certain working condition into the input layer I1 of the identification network model, and the system parameter layer P3 of the identification network model outputs the specific values ​​corresponding to the system parameters, which are used as the identification values ​​of the system parameters.

[0022] This device for performing parameter identification on a complex system using a parallel neural network implements the method or process according to an embodiment of the present disclosure.

[0023] In a third aspect, the present invention discloses a computer program product, comprising a computer program, which, when executed by a processor, implements the steps of the method for parameter identification of a complex system using a parallel neural network disclosed in the first aspect.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] 1. The present invention aims at a complex system whose system parameters change with multiple influencing factors, and proposes a method for parameter identification of a complex system using a parallel neural network. The method of the present invention adopts a self-designed parallel neural network, which takes multiple influencing factors as the input of the input layer I1, connects the input layer I1 and the system parameter layer P3 through the deep neural network of the middle layer M2, identifies the system parameters through the system parameter layer P3, and takes the target parameters as the output of the output layer O4. It can realize the accurate identification of system parameters under variable working conditions, and provide a basis for state observation and precise control of complex systems.

[0026] 2. The parallel neural network used in the present invention has multiple parallel networks consisting of the input layer I1, the middle layer M2, and the system parameter layer P3. The strong fitting ability of the neural network can be used to identify the dynamically changing system parameters of complex nonlinear systems. At the same time, the physical mechanism between influencing factors and key parameters can be used to improve the interpretability and generalization ability of the neural network model.

[0027] 3. The method of the present invention uses parallel neural networks to identify parameters of complex systems. There are methods suitable for embedded design and methods suitable for high-performance hardware use. The method can be selected according to actual conditions. It has strong practicality and broad application prospects.

[0028] 4. The present invention conducts simulation comparison and verification on the rotor temperature thermal network system of the induction motor of new energy vehicles to verify the effectiveness, accuracy, and applicability of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0030] Figure 1 A structural diagram of a parallel neural network provided in Example 1 of the present invention;

[0031] Figure 2 This is a structural diagram of a parallel neural network in a rotor temperature thermal network system for an induction motor of a new energy vehicle provided by Example 4 of the present invention;

[0032] Figure 3 This is a simplified model diagram of the rotor temperature thermal network system of the induction motor of a new energy vehicle in Example 4 of the present invention;

[0033] Figure 4 This is a graph showing the curve changes of some parameters in the sample data set used for training the parallel neural network in Example 5 of the present invention;

[0034] Figure 5 This is a graph showing changes in model loss when training a parallel neural network in Example 5 of the present invention;

[0035] Figure 6 is the rotor loss P in Example 5 of the present invention r The identification results of Figure 1 ;

[0036] Figure 7 is the rotor loss P in Example 5 of the present invention r The identification results of Figure 2 ;

[0037] Figure 8 is the rotor loss P in Example 5 of the present invention r The identification results of Figure 3 ;

[0038] Figure 9 is the rotor temperature T in Example 5 of the present invention r The comparison chart of the obtained results. DETAILED DESCRIPTION

[0039] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0040] It should be noted that when a component is referred to as being "mounted on" another component, it may be directly on the other component or there may be a central component. When a component is considered to be "set on" another component, it may be directly set on the other component or there may be a central component. When a component is considered to be "fixed to" another component, it may be directly fixed to the other component or there may be a central component.

[0041] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this invention pertains. The terms used herein in the specification of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The term "or / and" as used herein includes any and all combinations of one or more of the associated listed items.

[0042] Example 1

[0043] See Figure 1 , which is a structural diagram of a parallel neural network provided in this embodiment 1, which is applied in a certain target complex network (such as the rotor temperature thermal network system of the new energy vehicle induction motor mentioned in the background technology, or other complex systems - such as the new energy vehicle battery SOC estimation system, the new energy vehicle SOH estimation system, etc.).

[0044] like Figure 1 As shown, the parallel neural network includes: input layer I1, middle layer M2, system parameter layer P3, and output layer O4.

[0045] In other words, the parallel neural network can be divided into 4 layers, with connections only between adjacent layers. For ease of representation, the input layer I1, the middle layer M2, the system parameter layer P3, and the output layer O4 are arranged in order from top to bottom.

[0046] Among them, the input layer I1 is used to input several influencing factors (i.e. Figure 1I1, I2, I3, … in ).

[0047] The middle layer M2 consists of several deep neural networks, which connect the input layer I1 and the system parameter layer P3. The types of deep neural networks in the middle layer M2 include, but are not limited to, feedforward fully connected neural networks, convolutional neural networks, and Transformer neural networks, and the choice is based on the actual situation.

[0048] The system parameter layer P3 is used to output several system parameters obtained by the middle layer M2 (i.e. Figure 1 P1, P2, P3, ... in ).

[0049] The output layer O4 has two types of inputs, one of which is the system parameters obtained by the system parameter layer P3 (i.e. Figure 1 P1, P2, P3, ...), the other type of input is a number of system state quantities (i.e. Figure 1 V1, V2, ...)

[0050] The output layer O4 is used to combine the system parameters and system state quantities and calculate several target parameters based on the corresponding calculation relationship (i.e. Figure 1 O1, O2, ... in ).

[0051] Among them, the input layer I1, the middle layer M2, and the system parameter layer P3 constitute multiple parallel networks. The number of parallel networks is the same as the number of system parameters. For example, Figure 1 Here, P1 corresponds to parallel network-1, P2 corresponds to parallel network-2, P3 corresponds to parallel network-3, and so on.

[0052] A parallel network has a deep neural network. It should be noted that different parallel networks can use the same type of deep neural network or different types of deep neural networks.

[0053] Each parallel network is independent and has no connection relationship with each other; that is, the parallel networks do not interfere with each other.

[0054] As mentioned above, when the parallel neural network is applied to a target complex system, the specific types of influencing factors, system parameters, target parameters, and system state quantities are obtained based on the target complex system.

[0055] For any parallel network, the factors affecting its input directly affect the system parameters of its output. That is, P1 has a direct functional relationship with I1, P2 has a direct functional relationship with I1 and I2, P3 has a direct functional relationship with I1, I2, and I3, and so on.

[0056] Also, see Figure 1The influencing factors of a single parallel network input can be single or multiple; the influencing factors of different parallel network inputs can be repeated in type, for example: Figure 1 The influencing factors of the three parallel networks shown are all I1, and the influencing factors of parallel network-2 and parallel network-3 are both I1 and I2.

[0057] In addition, the calculation relationship between system parameters, system state quantities and target parameters is constructed based on the physical model of the target complex system: Figure 1 Taking the figure as an example, assuming that the system parameters are only P1, P2, and P3, the system state quantities are only V1 and V2, and the target parameters are only O1 and O2, then the calculation relationship between these three types of parameters is established through the weight coefficient group [w1, w2, w3, w4, w5, w6], which can be expressed as: O1=P1w1+P2w2+P3w3, O2=P1w4+P2w5+P3w6, which is the mathematical formula expression of the physical model of the target complex system.

[0058] The parallel neural network proposed in this embodiment 1 can be built into electronic devices and other devices for specific applications.

[0059] The parallel neural network proposed in this embodiment 1 can mainly realize two functions: 1. Identify the system parameters of the complex system; 2. Further obtain the numerical values ​​of the target parameters of the target complex system.

[0060] In order to distinguish the descriptions, the subsequent embodiment 2 introduces the process of parameter identification of a complex system based on a parallel neural network, and the subsequent embodiment 3 introduces the process of obtaining target parameter values ​​of a complex system based on a parallel neural network.

[0061] Example 2

[0062] Based on the parallel neural network proposed in Example 1, this embodiment 2 provides two methods for parameter identification of complex systems using parallel neural networks, which are used to identify system parameters of a target complex system, wherein the target complex system is affected by multiple influencing factors.

[0063] 2.1, the first method includes the following steps:

[0064] S201, using a sample data set of the target complex system to train a pre-built parallel neural network, obtaining a parallel neural network with optimal performance and using it as an identification network model;

[0065] S202, inputting specific values ​​of influencing factors under a certain working condition into the input layer I1 of the identification network model, and the system parameter layer P3 of the identification network model outputting specific values ​​corresponding to the system parameters, i.e., the identification values ​​of the system parameters;

[0066] S203, traversing and inputting the values ​​of the influencing factors under different working conditions, obtaining the specific values ​​corresponding to the system parameters, and then constructing a numerical mapping table between the influencing factors and the system parameters;

[0067] S204 , according to a specific value of the influencing factor, look up a specific value corresponding to the system parameter in a value mapping table and use it as an identification value of the system parameter.

[0068] It should be noted that the more refined the numerical mapping table is, the closer its accuracy is to the identification network model.

[0069] The method in 2.1 is applicable to low-performance controllers (such as embedded MCUs)—the value mapping table obtained in S203 can be stored in the low-performance controller to complete the identification of system parameters by looking up the table.

[0070] 2.2, the second method includes the following steps:

[0071] S201, using a sample data set of the target complex system to train a pre-built parallel neural network, obtaining a parallel neural network with optimal performance and using it as an identification network model;

[0072] S202, input the specific value of the influencing factor under a certain working condition to the input layer I1 of the identification network model, and the system parameter layer P3 of the identification network model outputs the specific value corresponding to the system parameter, that is, the identification value of the system parameter.

[0073] The method in 2.2 is applicable to high-performance controllers—the identification network model obtained in S201 can be stored in the high-performance controller, so that the identification network model can be directly used to complete the identification of system parameters.

[0074] Methods 2.1 and 2.2 both involve training parallel neural networks. Specifically, the training methods are as follows:

[0075] S101: Obtain a sample data set obtained based on a target complex system experiment. The sample data set includes label data of influencing factors and target parameters, which can be used as true values.

[0076] S102: Establish the loss function required for model training based on the target parameters.

[0077] S103: Set the network hyperparameters required for model training based on the size of the sample data set and the performance of the computing platform.

[0078] S104, based on the sample data set and loss function, the parallel neural network is trained for multiple rounds, and the network parameters representing the various connection relationships of the model are iteratively updated (including: ① the connection relationship between the input layer I1, the middle layer M2, and the system parameter layer M3; ② the connection relationship within the middle layer M2) to obtain the network parameter combination with the best performance (i.e., the best accuracy and generalization performance), corresponding to the parallel neural network with the best performance, which is used as the identification network model.

[0079] This embodiment 2 also simultaneously discloses a device for performing parameter identification on a complex system using a parallel neural network, using the method 2.1 or 2.2 above.

[0080] The device for performing parameter identification on a complex system by using a parallel neural network comprises: an identification network acquisition module and a complex system processing module.

[0081] The identification network acquisition module is configured as follows: using the sample data set of the target complex system to train the pre-built parallel neural network, obtaining the parallel neural network with the best performance and using it as the identification network model;

[0082] The complex system processing module is configured as follows: the specific values ​​of the influencing factors under a certain working condition are input into the input layer I1 of the identification network model, and the system parameter layer P3 of the identification network model outputs the specific values ​​corresponding to the system parameters, which are used as the identification values ​​of the system parameters.

[0083] Of course, the device for performing parameter identification on a complex system using a parallel neural network may further include: a numerical mapping table construction module and a numerical query module.

[0084] The numerical mapping table construction module is configured as follows: traversing the numerical values ​​of the influencing factors under different working conditions, obtaining the specific numerical values ​​corresponding to the system parameters, and then constructing a numerical mapping table of the influencing factors and the system parameters.

[0085] The numerical query module is configured to: according to a specific numerical value of the influencing factor, look up the specific numerical value corresponding to the system parameter in the numerical mapping table and use it as the identification value of the system parameter.

[0086] Example 3

[0087] Based on the parallel neural network proposed in Example 1, this Example 3 provides two methods for parameter identification of complex systems using parallel neural networks (actually, methods for obtaining numerical values ​​of complex systems using parallel neural networks), which are used to obtain numerical values ​​of target parameters of a target complex system. The target complex system is affected by multiple influencing factors.

[0088] 3.1, the first method includes the following steps:

[0089] S301, using a sample data set of the target complex system to train a pre-built parallel neural network, obtaining a parallel neural network with optimal performance, and using it as an identification network model;

[0090] S302, inputting specific values ​​of influencing factors under a certain working condition into the input layer I1 of the identification network model, and the system parameter layer M3 of the identification network model obtains specific values ​​of corresponding system parameters;

[0091] The specific numerical value of the system state quantity under the same working condition as the influencing factor is also input to the output layer O4 of the identification network model, and the output layer O4 of the identification network model outputs the specific numerical value corresponding to the target parameter, which is the obtained value of the target parameter;

[0092] S303, traversing and inputting the values ​​of the influencing factors under different working conditions, obtaining the specific values ​​corresponding to the system parameters, and then constructing a numerical mapping table between the influencing factors and the system parameters;

[0093] S304, according to a specific value of the influencing factor, look up the specific value of the system parameter in the value mapping table and use it as the identification value of the system parameter;

[0094] Furthermore, based on the specific values ​​of the system state quantities under the same working conditions as the influencing factors, and in accordance with the calculation relationship between the influencing factors, the system state quantities and the target parameters, the specific values ​​corresponding to the target parameters are calculated, which are used as the acquired values ​​of the target parameters.

[0095] It should be noted that the more refined the numerical mapping table is, the closer its accuracy is to the identification network model.

[0096] Similar to 2.1, method 3.1 is suitable for low-performance controllers (such as embedded MCUs) - the numerical mapping table obtained in S303 can be stored in the low-performance controller to first complete the identification of system parameters by looking up the table, and then calculate the specific values ​​of the target parameters in combination with the system state quantity, thereby completing the numerical acquisition of the target parameters.

[0097] 3.2, the second method includes the following steps:

[0098] S301, using a sample data set of the target complex system to train a pre-built parallel neural network, obtaining a parallel neural network with optimal performance, and using it as an identification network model;

[0099] S302, inputting specific values ​​of influencing factors under a certain working condition into the input layer I1 of the identification network model, and the system parameter layer M3 of the identification network model obtains specific values ​​of corresponding system parameters;

[0100] The specific numerical value of the system state quantity under the same working conditions as the influencing factors is also input into the output layer O4 of the identification network model, and the output layer O4 of the identification network model outputs the specific numerical value corresponding to the target parameter, which is the acquired value of the target parameter.

[0101] Similar to 2.2, the method in 3.2 is applicable to high-performance controllers. The identification network model obtained in S201 can be stored in the high-performance controller, so that the numerical value of the target parameter can be directly obtained using the identification network model.

[0102] Methods 3.1 and 3.2 both involve training of parallel neural networks, and their processes are the same as those in Example 2 and will not be repeated here.

[0103] This embodiment 3 also simultaneously discloses a device for acquiring numerical values ​​of a complex system using a parallel neural network, using the method 3.1 or 3.2 above.

[0104] The device for acquiring numerical values ​​of a complex system using a parallel neural network includes: an identification network acquisition module and a complex system processing module.

[0105] The identification network acquisition module is configured as follows: using the sample data set of the target complex system to train the pre-built parallel neural network, obtaining the parallel neural network with the best performance and using it as the identification network model;

[0106] The complex system processing module is configured as follows: the specific values ​​of the influencing factors under a certain working condition are input into the input layer I1 of the identification network model, and the system parameter layer M3 of the identification network model obtains the specific values ​​of the corresponding system parameters; the specific values ​​of the system state quantities under the same working conditions as the influencing factors are also input into the output layer O4 of the identification network model, and the output layer O4 of the identification network model outputs the specific values ​​corresponding to the target parameters, which are the acquired values ​​of the target parameters.

[0107] Of course, the device for acquiring numerical values ​​of a complex system using a parallel neural network may also include: a numerical mapping table construction module and a numerical query module.

[0108] The numerical mapping table construction module is configured as follows: traversing the numerical values ​​of the influencing factors under different working conditions, obtaining the specific numerical values ​​corresponding to the system parameters, and then constructing a numerical mapping table of the influencing factors and the system parameters.

[0109] The numerical query module is configured as follows: based on a specific numerical value of the influencing factor, the specific numerical value of the system parameter is obtained by looking up in the numerical mapping table and used as the identification value of the system parameter; and based on the specific numerical value of the system state quantity under the same working conditions as the influencing factor, according to the calculation relationship between the influencing factor, the system state quantity and the target parameter, the specific numerical value corresponding to the target parameter is calculated and used as the acquisition value of the target parameter.

[0110] Example 4

[0111] This embodiment 4 provides a specific application example of embodiment 1:

[0112] This embodiment 4 provides a parallel neural network for the rotor temperature thermal network system of the induction motor of a new energy vehicle. For the convenience of explanation, the model is named as: rotor temperature parallel neural network, and its structure is as follows: Figure 2 shown.

[0113] See Figure 2 , the composition and parallel neural network of rotor temperature Figure 1 Similar, the specific differences are:

[0114] The number of parallel networks in the parallel neural network is 3.

[0115] For the first parallel network, the influencing factor is the motor speed n r , the system parameter is the thermal conductivity K between the rotor and the stator sr ;

[0116] For the second parallel network, the influencing factor is the motor speed n r , coolant temperature T c , the system parameter is the thermal conductivity K between the rotor and the coolant cr ;

[0117] For the third parallel network, the influencing factor is the motor speed n r , stator current I s , the rotor temperature at the current time t The system parameter is the rotor loss P r .

[0118] System state quantities include: stator temperature T s , adjacent time interval Δt, rotor heat capacity C r .

[0119] The target parameter is the rotor temperature difference between the next time t+1 and the current time t

[0120] The calculation relationship between system parameters, system state quantities, and target parameters is constructed based on the physical model of the rotor temperature thermal network system of the induction motor of a new energy vehicle, specifically:

[0121]

[0122] It should be noted that the above formula is the expression of the physical model of the rotor temperature thermal network system of the induction motor of new energy vehicles.

[0123] Specifically, the construction of the rotor temperature parallel neural network is based on the rotor temperature thermal network system of the induction motor of new energy vehicles:

[0124] According to the heat transfer path and loss of the induction motor, the rotor temperature thermal network system of the induction motor of new energy vehicles can be simplified into a three-node lumped parameter thermal network model, such as Figure 3 shown.

[0125] The three-node lumped parameter thermal network model includes three thermal network temperature nodes: rotor, stator and cooling oil. Three heat transfer paths are set according to the actual heat transfer process between the nodes.

[0126] Then, a calculation relationship describing the rotor temperature change can be established:

[0127]

[0128] In the above calculation relationship, is the target parameter; T s 、T c Can be obtained by measuring with a temperature sensor; are variables that need to be estimated in real time, Δt, C r is a known parameter; K sr , K cr 、P r is the system parameter that needs to be identified; w sr 、w cr and w pr are the weight coefficients obtained by comprehensive calculation of the temperature difference of motor components, sampling time and rotor heat capacity.

[0129] Due to the complex and changeable vehicle operating conditions, the above system parameters change in real time with the motor operating conditions, and the input quantities that cause the changes in each parameter are different:

[0130] P r For example, it is also affected by I s 、 and n r The influence of , which can be approximately expressed as the following function form:

[0131]

[0132] Similarly, by analyzing the thermal conductivity characteristics of the motor system, K cr Mainly affected by T c With n r The influence of K sr The main influence is n r The corresponding functional relationship can be expressed as:

[0133] K cr =fcr (T c ,n r ), K sr =f sr (n r ).

[0134] So, I s 、 n r 、T c As an influencing factor, K sr , K cr 、P r As system parameters, As the target parameter; T s ,Δt,C r As the system state quantity.

[0135] Based on the above relationship, the parallel neural network of Example 1 is adjusted to obtain a rotor temperature parallel neural network.

[0136] Then, referring to the methods of Examples 2 and 3, the rotor temperature parallel neural network can be used to perform K in the rotor temperature thermal network system of the induction motor of the new energy vehicle. sr , K cr 、P r Identification and conduct Get the value of .

[0137] Example 5

[0138] This embodiment 5 provides an application example of the rotor temperature parallel neural network in embodiment 4 for system parameter identification and target parameter value acquisition to verify and illustrate the effects of embodiments 1, 2, and 3.

[0139] Specifically, the operations of the application example are as follows:

[0140] 1) Get a sample dataset

[0141] The full working condition is obtained by bench test (by adjusting I s To achieve different rotor torque, motor speed) of each motor component temperature change data (T s 、T c 、T r ).

[0142] like Figure 4 As shown in Figure 1, the temperature changes of various components of the motor at 20N.m\1000rpm are similar to those under other different working conditions.

[0143] 2) Design loss function

[0144] Use the rotor temperature parallel neural network to obtain the estimated value of the rotor temperature change at the i-th moment Compare it with the actual value of rotor temperature change at moment i The error between them (calculated based on the sample data set) is used as the loss function L to evaluate the accuracy of the model reg :

[0145]

[0146] 3) Setting hyperparameters

[0147] The Adam optimizer is selected as the iterative calculation scheme for model parameters, and the model input (i.e., sample data set) is pre-normalized. At the same time, the L2 regularization scheme is introduced to improve the generalization ability of the model and determine hyperparameters such as sample size, learning rate, and learning rate decay coefficient in a single iteration process.

[0148] 4) Perform model training

[0149] The sample data set is divided into training set and test set in appropriate proportions; the training set is used to complete the training of the rotor temperature parallel neural network in a server equipped with an RTX3090 computing platform. The training process is as follows: Figure 5 As shown in the figure, the performance is verified on the test set, and finally the identification network model is obtained.

[0150] 5) Perform system parameter identification

[0151] According to 1.1 of Example 1, K is directly identified by identifying the network model. sr , K cr 、P r .

[0152] P r Take P as an example: r Received s 、 n r The impact of Figures 6 to 8 , respectively When the temperature is 80℃, 100℃ and 120℃, P r Follow me s and n r It can be seen that the change trend basically conforms to the physical law of the real motor, indicating that the identification result is accurate.

[0153] 6) Acquire target parameter data

[0154] According to 2.2 of Example 1, a mapping relationship table (medium precision) is prepared based on the identification network model and stored in the vehicle motor controller; the vehicle motor controller is used to look up the table and calculate the Specific value.

[0155] Then, and Add them together to get the rotor temperature at the next moment t+1 ( (also an estimate).

[0156] See Figure 9 (a)~ Figure 9 (f) shows a comparison between the estimated rotor temperature and the actual temperature (the true value of the sample data set) under certain operating conditions. It can be seen that the temperature error is below 5°C under various operating conditions, verifying the effectiveness, accuracy, and applicability of Examples 1, 2, and 3.

[0157] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0158] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the patent for this invention shall be determined by the appended claims.

Claims

1. A method for parameter identification of a complex system using a parallel neural network, characterized in that: It is used to identify system parameters of a target complex system; wherein the target complex system is affected by multiple influencing factors; The method for performing parameter identification on a complex system using a parallel neural network comprises: S1, using the sample data set of the target complex system to train the pre-built parallel neural network, obtain the parallel neural network with the best performance, and use it as the identification network model; The parallel neural network includes: an input layer I1, an intermediate layer M2, a system parameter layer P3, and an output layer O4; The input layer I1 is used to input several influencing factors into the middle layer M2; the middle layer M2 is a number of deep neural networks, which are used to connect the input layer I1 and the system parameter layer P3; the system parameter layer P3 is used to output several system parameters obtained by the middle layer M2; the output layer O4 has two types of inputs, one of which is the system parameters obtained by the system parameter layer P3, and the other is several system state quantities; the output layer O4 is used to combine the system parameters and system state quantities and calculate several target parameters based on the corresponding calculation relationship; The input layer I1, the middle layer M2, and the system parameter layer P3 constitute multiple parallel networks; the number of parallel networks is the same as the number of system parameters, and each parallel network has a deep neural network; each parallel network is independent and has no connection relationship with each other; the specific types of influencing factors, system parameters, target parameters, and system state quantities are obtained based on the target complex system; for any parallel network, its input influencing factors directly affect its output system parameters; the calculation relationship between system parameters, system state quantities, and target parameters is constructed based on the physical model of the target complex system; S2, input the specific value of the influencing factor under a certain working condition to the input layer I1 of the identification network model, and the system parameter layer P3 of the identification network model outputs the specific value corresponding to the system parameter, which is the identification value of the system parameter.

2. The method for parameter identification of a complex system using a parallel neural network according to claim 1, characterized in that: The types of the deep neural network of the intermediate layer M2 include but are not limited to: a feedforward fully connected neural network, a convolutional neural network, and a Transformer neural network.

3. The method for parameter identification of a complex system using a parallel neural network according to claim 1, characterized in that: In S2, the specific numerical value of the system state quantity under the same working conditions as the influencing factors is also input into the output layer O4 of the identification network model, and the output layer O4 of the identification network model outputs the specific numerical value corresponding to the target parameter, which is the acquired value of the target parameter.

4. The method for parameter identification of a complex system using a parallel neural network according to claim 1 or 3, characterized in that: Also includes: S3, traversing the input values ​​of the influencing factors under different working conditions, obtaining the specific values ​​corresponding to the system parameters, and then constructing a numerical mapping table between the influencing factors and the system parameters; S4, according to a specific value of the influencing factor, look up the specific value corresponding to the system parameter in the value mapping table and use it as the identification value of the system parameter.

5. According to the method for parameter identification of a complex system using a parallel neural network according to claim 4, in S4, based on the specific numerical value of the system state quantity under the same working conditions as the influencing factor, and in accordance with the calculation relationship between the influencing factor, the system state quantity and the target parameter, the specific numerical value corresponding to the target parameter is calculated, that is, the acquired value of the target parameter.

6. The method for parameter identification of a complex system using a parallel neural network according to claim 1, characterized in that: In S1, the method for training the pre-built parallel neural network includes: S101, obtaining a sample data set obtained based on a target complex system experiment; wherein the sample data set includes label data of influencing factors and target parameters; S102, establishing a loss function required for model training based on the target parameters; S103, setting the network hyperparameters required for model training based on the size of the sample data set and the performance of the computing platform; S104, performing multiple rounds of training on the parallel neural network based on the sample data set and the loss function, iteratively updating the network parameters that characterize the various connection relationships of the model to obtain the network parameter combination with the best performance, and correspondingly obtaining the parallel neural network with the best performance.

7. The method for parameter identification of a complex system using a parallel neural network according to claim 1, characterized in that: The target complex system includes but is not limited to: a rotor temperature thermal network system of an induction motor of a new energy vehicle, a SOC estimation system of a battery of a new energy vehicle, and a SOH estimation system of a battery of a new energy vehicle.

8. The method for parameter identification of a complex system using a parallel neural network according to claim 7, characterized in that: When the target complex system is a rotor temperature thermal network system of an induction motor of a new energy vehicle, the number of parallel networks in the parallel neural network is 3; For the first parallel network, the influencing factor is the motor speed n r , the system parameter is the thermal conductivity K between the rotor and the stator sr ; For the second parallel network, the influencing factor is the motor speed n r , coolant temperature T c , the system parameter is the thermal conductivity K between the rotor and the coolant cr ; For the third parallel network, the influencing factor is the motor speed n r , stator current I s , the rotor temperature at the current time t The system parameter is the rotor loss P r ; System state quantities include: stator temperature T s , adjacent time interval Δt, rotor heat capacity C r ; The target parameter is the rotor temperature difference between the next time t+1 and the current time t The calculation relationship between system parameters, system state quantities and target parameters is constructed based on the physical model of the rotor temperature thermal network system of the induction motor of new energy vehicles; Among them, the expression of the physical model of the rotor temperature thermal network system of the new energy vehicle induction motor is:

9. A device for parameter identification of a complex system using a parallel neural network, characterized in that: It adopts the method for parameter identification of a complex system using a parallel neural network as described in any one of claims 1 to 8; The device for performing parameter identification on a complex system using a parallel neural network comprises: An identification network acquisition module is used to train a pre-built parallel neural network using a sample data set of a target complex system to obtain a parallel neural network with optimal performance as the identification network model; and The complex system processing module is used to input the specific values ​​of the influencing factors under a certain working condition into the input layer I1 of the identification network model, and the system parameter layer P3 of the identification network model outputs the specific values ​​corresponding to the system parameters, which are used as the identification values ​​of the system parameters.

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 method for parameter identification of a complex system using a parallel neural network described in any one of claims 1 to 8 are implemented.

Images