Aluminum rotor motor dynamic energy efficiency monitoring method and system based on PINN

Abstract

The embodiment of the application relates to a PINN-based dynamic energy efficiency monitoring method and system for an aluminum rotor motor, comprising: acquiring electrical signal time sequence data on the stator side of an aluminum rotor induction motor; constructing a physical information neural network, wherein a composite loss function comprises a motor electromagnetic dynamics equation residual term, an aluminum rotor thermal dynamics equation residual term, and a boundary regularization term based on the physical properties of aluminum material; based on the composite loss function, using an alternating optimization strategy, training the physical information neural network using historical offline data in the electrical signal time sequence data; using the trained physical information neural network, identifying internal state parameters of the aluminum rotor induction motor that cannot be directly measured; decoupling calculation of transient loss of the aluminum rotor induction motor to obtain dynamic energy efficiency of the aluminum rotor induction motor. The application can solve the problems of unmeasurable energy efficiency and difficult-to-identify parameter drift of the aluminum rotor induction motor under dynamic working conditions.

Description

Technical Field

[0001] This invention relates to the field of energy efficiency monitoring technology for aluminum rotor induction motors, and particularly to a dynamic energy efficiency monitoring method and system for aluminum rotor motors based on PINN. Background Technology

[0002] Cast aluminum rotor induction motors are widely used in small and medium-sized industrial drives due to their low density, ease of die casting, and overall cost advantages. In actual production, motors often operate under dynamic conditions such as non-rated loads, voltage fluctuations, and frequent start-stop cycles. Accurate monitoring of their instantaneous energy efficiency has become a key technological bottleneck for industrial energy conservation optimization. Existing energy efficiency monitoring methods mainly include the equivalent circuit method based on steady-state assumptions, the air gap torque method, and data-driven artificial intelligence methods. The equivalent circuit method relies on no-load and stall tests to obtain parameters, making it unsuitable for dynamic on-site conditions. While the air gap torque method is non-invasive, it requires precise stator resistance data and struggles to dynamically separate stray losses. Although data-driven methods based on neural networks can fit the nonlinear mapping between input power and output efficiency, they rely on massive amounts of labeled training data across all operating conditions, and the models lack physical interpretability, limiting their generalization ability. Furthermore, for aluminum rotor induction motors, the rotor resistance drifts significantly under dynamic loads due to the high temperature coefficient of resistance in aluminum. In addition, the closed rotation of the squirrel-cage rotor structure makes direct rotor temperature measurement impossible, making it difficult for traditional methods to track the nonlinear changes in thermo-electric coupling in real time.

[0003] Therefore, the main problems with existing technologies are: the rotor resistance of aluminum rotor induction motors under dynamic operating conditions drifts in real time with temperature and is unmeasurable, causing energy efficiency monitoring methods based on constant parameters or steady-state assumptions to fail; pure data-driven models lack physical constraints, making it difficult to guarantee the physical consistency and generalization ability of dynamic energy efficiency calculations; and traditional physical models cannot simultaneously characterize the multi-timescale coupling characteristics of electromagnetic transients and thermal transients, resulting in inaccurate or even unusable energy efficiency monitoring results. These technical defects collectively constitute the technical obstacles of unmeasurable energy efficiency and difficult-to-identify parameter drift of aluminum rotor induction motors under dynamic operating conditions, restricting the energy-saving optimization and energy efficiency management level improvement of industrial motor systems. Summary of the Invention

[0004] The purpose of this invention is to provide at least a method and system for monitoring the dynamic energy efficiency of aluminum rotor motors based on PINN, which can solve the problems of unmeasurable energy efficiency and difficult identification of parameter drift in aluminum rotor induction motors under dynamic operating conditions.

[0005] To address the aforementioned technical problems, at least one embodiment of this application provides a dynamic energy efficiency monitoring method for aluminum rotor motors based on PINN, comprising:

[0006] Acquire electrical signal timing data on the stator side of the aluminum rotor induction motor. The electrical signal timing data consists of time-series measurement data of the three-phase voltage and three-phase current on the stator side.

[0007] A physical information neural network is constructed. The composite loss function of the physical information neural network includes residual terms of the motor electromagnetic dynamics equation, residual terms of the aluminum rotor thermodynamics equation, and boundary regularization terms based on the physical properties of aluminum material.

[0008] Based on the composite loss function, an alternating optimization strategy is adopted, and historical offline data in the time series data of electrical signals is used to train the physical information neural network.

[0009] Using a trained physical information neural network, the internal state parameters of an aluminum rotor induction motor that cannot be directly measured are identified.

[0010] Based on real-time online data and internal state parameters in electrical signal time series data, the transient loss of aluminum rotor induction motor is decoupled and calculated to obtain the dynamic energy efficiency of aluminum rotor induction motor.

[0011] In some optional embodiments, acquiring electrical signal timing data on the stator side of the aluminum rotor induction motor includes:

[0012] Real-time measurement data of three-phase voltage and three-phase current on the stator side of the aluminum rotor induction motor were collected;

[0013] Synchronous rotating coordinate system transformation is performed on real-time measurement data of three-phase voltage and three-phase current.

[0014] The transformed electrical signal is standardized to generate electrical signal timing data.

[0015] In some alternative embodiments, a physical information neural network is constructed, including:

[0016] A physical information neural network with a dual time scale parallel structure is constructed. The physical information neural network contains a first branch and a second branch that are independent of each other. The first branch is used to estimate rotor flux, rotor current and rotor speed parameters, and the second branch is used to estimate rotor resistance and rotor temperature parameters.

[0017] Replace the rotor resistance in the residual term of the motor electromagnetic dynamics equation with the rotor resistance output from the second branch, and replace the rotor current in the residual term of the aluminum rotor thermodynamics equation with the rotor current calculated based on the rotor flux linkage output from the first branch.

[0018] The residual terms of the motor electromagnetic dynamics equation and the aluminum rotor thermodynamics equation are weighted and then added to the composite loss function.

[0019] In some optional embodiments, constructing a physical information neural network further includes:

[0020] In the composite loss function, a first dynamic weighting coefficient corresponding to the residual term of the motor electromagnetic dynamics equation and a second dynamic weighting coefficient corresponding to the residual term of the aluminum rotor thermodynamics equation are respectively set.

[0021] The values ​​of the first dynamic weighting coefficient and the second dynamic weighting coefficient are adaptively adjusted based on the rate of change of current in the electrical signal timing data.

[0022] In some optional embodiments, the values ​​of the first dynamic weighting coefficient and the second dynamic weighting coefficient are adaptively adjusted according to the rate of change of current in the electrical signal timing data, including:

[0023] Obtain the rate of change of current in the timing data of electrical signals;

[0024] When the rate of change of current exceeds a preset threshold, the value of the first dynamic weighting coefficient is increased and the value of the second dynamic weighting coefficient is decreased.

[0025] When the rate of change of current is lower than a preset threshold, the value of the first dynamic weighting coefficient is reduced and the value of the second dynamic weighting coefficient is increased.

[0026] In some optional embodiments, an alternating optimization strategy is employed to train the physical information neural network using historical offline data from electrical signal time-series data, including:

[0027] The training process is divided into alternating electromagnetic-dominated training phases and thermal-dominated training phases.

[0028] During the electromagnetic-dominated training phase, the network parameters of the second branch are frozen, and the network parameters related to the rapidly changing electromagnetic state are updated based on the adjusted first dynamic weight coefficients.

[0029] During the hot-dominated training phase, the network parameters of the first branch are frozen, and the network parameters related to the slowly changing hot parameters are updated based on the adjusted second dynamic weight coefficients.

[0030] Repeat the alternating update process described above until the composite loss function converges to the preset threshold.

[0031] In some alternative embodiments, identifying non-measurable internal state parameters of the aluminum rotor induction motor includes:

[0032] Real-time electrical signal timing data is input into a physical information neural network. Through forward propagation and automatic differential calculation, the internal state parameters that satisfy the physical constraints of the motor electromagnetic dynamics equation and the aluminum rotor thermodynamics equation are output.

[0033] The internal state parameters include at least one of rotor flux linkage, rotor resistance, rotor temperature, and rotor speed, wherein the trajectories of the changes in rotor resistance and rotor temperature are constrained by the residual terms of the aluminum rotor thermodynamic equation.

[0034] In some optional embodiments, the transient losses of the aluminum rotor induction motor are decoupled and calculated to obtain the dynamic energy efficiency of the aluminum rotor induction motor, including:

[0035] Calculate the instantaneous input active power of the motor based on electrical signal timing data;

[0036] The stator copper loss, rotor copper loss, dynamic iron loss, stray loss, and mechanical loss are calculated based on the internal state parameters.

[0037] The dynamic energy efficiency of the aluminum rotor induction motor is obtained by subtracting stator copper loss, rotor copper loss, dynamic iron loss, stray loss, and mechanical loss from the instantaneous input active power.

[0038] In some optional embodiments, calculating dynamic iron loss includes:

[0039] The time derivative of the rotor flux linkage is calculated using the automatic differentiation function of the physical information neural network.

[0040] The dynamic iron loss under variable frequency operation is calculated based on the time derivative of the rotor flux linkage and the synchronous electric angular velocity.

[0041] At least one embodiment of this application also provides a dynamic energy efficiency monitoring system for aluminum rotor motors based on PINN, including:

[0042] The data acquisition module is used to acquire the electrical signal timing data of the stator side of the aluminum rotor induction motor. The electrical signal timing data is the time series measurement data of the three-phase voltage and three-phase current on the stator side.

[0043] The network construction module is used to construct a physical information neural network. The composite loss function of the physical information neural network includes residual terms of the motor electromagnetic dynamics equation, residual terms of the aluminum rotor thermodynamics equation, and boundary regularization terms based on the physical properties of aluminum material.

[0044] The model training module is used to train the physical information neural network based on a composite loss function and an alternating optimization strategy, using historical offline data from electrical signal time series data.

[0045] The state identification module is used to identify the internal state parameters of the aluminum rotor induction motor that cannot be directly measured by utilizing the trained physical information neural network.

[0046] The energy efficiency calculation module is used to decouple the transient losses of the aluminum rotor induction motor based on real-time online data and internal state parameters in the electrical signal time series data, so as to obtain the dynamic energy efficiency of the aluminum rotor induction motor.

[0047] The PINN-based dynamic energy efficiency monitoring method for aluminum rotor motors provided in this application constructs a composite loss function that embeds the motor's electromagnetic dynamics equation, the aluminum rotor's thermodynamics equation, and the physical boundary regularization term of the aluminum material. This subjectes the physical information neural network to rigid constraints from both electromagnetic and thermal transient physical laws during training, enabling real-time identification of internal state parameters, including rotor resistance and temperature, using only stator-side electrical signal time-series data. Since the changes in rotor resistance and temperature follow the thermodynamic properties of aluminum, this application can accurately track the real-time drift of rotor resistance under dynamic loads without the need for a rotor temperature sensor. Furthermore, an alternating optimization strategy effectively decouples the multi-timescale coupling problem between the millisecond-level electromagnetic response and the minute-level thermal response, ensuring the convergence accuracy of both high-frequency electromagnetic state reconstruction and low-frequency thermal parameter identification after training. Finally, by decoupling transient losses based on electrical signal time-series data and the identified internal state parameters, stator copper losses, rotor copper losses, dynamic iron losses, and stray mechanical losses can be dynamically separated, directly obtaining dynamic energy efficiency that conforms to the law of conservation of energy. Therefore, this invention solves the problem of inaccurate energy efficiency monitoring caused by the unmeasurable and real-time drift of rotor resistance in aluminum rotor induction motors under dynamic operating conditions, and realizes non-invasive, physically consistent, and dynamic real-time energy efficiency calculation without the need for full-condition tag data. Attached Figure Description

[0048] One or more embodiments are illustrated by way of example with reference to the accompanying drawings, and these illustrative descriptions do not constitute a limitation on the embodiments.

[0049] Figure 1 This is a flowchart of a dynamic energy efficiency monitoring method for an aluminum rotor motor based on PINN, provided in one embodiment of this application;

[0050] Figure 2 This is a flowchart of a method for constructing a physical information neural network according to an embodiment of this application;

[0051] Figure 3 This is a flowchart of a method for adjusting the dynamic weight coefficients of a physical information neural network according to an embodiment of this application;

[0052] Figure 4 This is a flowchart of a method for training a physical information neural network using electrical signal timing data, provided in one embodiment of this application.

[0053] Figure 5This is a flowchart of a method for decoupling the calculation of transient losses of an aluminum rotor induction motor provided in one embodiment of this application;

[0054] Figure 6 This is a schematic diagram of a dynamic energy efficiency monitoring system for an aluminum rotor motor based on PINN, provided in one embodiment of this application;

[0055] Figure 7 This is a schematic diagram of the structure of an electronic device provided in one embodiment of this application. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the various embodiments of this application will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been provided in the various embodiments of this application to help readers better understand this application. However, the technical solutions claimed in this application can be implemented even without these technical details and various changes and modifications based on the following embodiments. The division of the various embodiments below is for the convenience of description and should not constitute any limitation on the specific implementation of this application. The various embodiments can be combined with and referenced by each other without contradiction.

[0057] In this invention, Physics-Informed Neural Networks (PINN) is a scientific machine learning model that embeds physical equations as constraints into the loss function of a neural network. Its training process is driven by both measured data and physical laws, enabling it to infer unmeasurable internal state variables from measurable port signals. This invention directly embeds the partial differential equations or ordinary differential equations controlling motor operation into the loss function of the neural network. The training process of the neural network is constrained by physical laws, allowing it to learn not only from measured data but also to infer unmeasurable hidden states using physical equations. Currently, most PINN research in the field of motors focuses on permanent magnet synchronous motors. This invention aims to solve the problems of unmeasurable energy efficiency and difficult parameter drift identification in aluminum rotor induction motors under dynamic operating conditions. Using the voltage and current at the motor stator terminals as the sole physical input, a Physics-Informed Neural Network is constructed. This network satisfies the electromagnetic-thermodynamic equations of the motor as its core constraint. By minimizing the residuals of the physical equations, it learns internal state variables that conform to physical laws, and then, combined with a loss separation model, calculates the dynamic efficiency of the motor in real time. This invention provides a dynamic energy efficiency monitoring method for aluminum rotor motors based on PINN. The implementation details of this embodiment are described in detail below with reference to the accompanying drawings. The following content is only for the convenience of understanding the implementation details and is not necessary for the implementation scheme.

[0058] Example 1:

[0059] The PINN-based dynamic energy efficiency monitoring method for aluminum rotor motors in this embodiment can be applied to electronic devices with communication, computing, and data storage capabilities. Its specific process can be as follows: Figure 1 As shown, the specific steps include 101 to 105:

[0060] Step 101: Obtain the electrical signal timing data of the stator side of the aluminum rotor induction motor. The electrical signal timing data is the time series measurement data of the three-phase voltage and three-phase current on the stator side.

[0061] In some embodiments, real-time measurement data of three-phase voltage and three-phase current on the stator side of the aluminum rotor induction motor are continuously acquired via voltage and current transformers installed at the motor control cabinet. The acquired real-time measurement data of three-phase voltage and current are transformed into synchronous rotating coordinate systems along the d-axis and q-axis, converting the AC signal in the three-phase stationary coordinate system into a DC signal in the synchronous rotating coordinate system. The transformed electrical signal is then standardized with zero mean and unit variance to map the data to a uniform numerical range, generating electrical signal time-series data suitable for neural network input.

[0062] In some examples, the generated electrical signal time-series data is subjected to multi-scale embedding processing. The data flowing to the subsequent first branch maintains the original sampling rate, while the data flowing to the subsequent second branch is processed by moving average to extract power envelope features in order to adapt to the time-scale characteristics of different physical processes.

[0063] In some embodiments, step 101 specifically includes:

[0064] Real-time measurement data of three-phase voltage and three-phase current on the stator side of an aluminum rotor induction motor are collected. Three-phase voltage signals at the input terminals of the motor stator windings are obtained through voltage transformers and current transformers, respectively. , , and three-phase current signal , , The sampling frequency is set to 10 kHz to fully capture the high-frequency components of the electromagnetic transient process. The acquired analog signal is converted into a digital signal by an analog-to-digital converter and transmitted to a digital signal processor or embedded edge computing unit.

[0065] A synchronous rotating coordinate system transformation is performed on the real-time measurement data of three-phase voltage and current. The preprocessed raw three-phase voltage and current data undergo a Park transformation, mapping the AC signals from the three-phase stationary coordinate system to the synchronous rotating coordinate system, obtaining the d-axis and q-axis components. The synchronous electrical angular velocity is calculated from the motor's power supply frequency. For variable frequency drive motors, the synchronous electrical angular velocity can be directly obtained from the inverter's communication interface or extracted by processing the stator voltage signal using a phase-locked loop. Through coordinate system transformation, the sinusoidal AC signal during steady-state operation is converted into a constant DC signal, eliminating the time-varying characteristics of the AC signal and reducing the learning difficulty and computational complexity of the physical information neural network.

[0066] The transformed electrical signals are standardized to generate electrical signal time-series data. Zero-mean, unit-variance standardization is applied to the d-axis and q-axis voltage and current signals after the synchronous rotating coordinate system transformation. The mean and standard deviation of each signal within a sliding time window are calculated. The value of each sampling point is subtracted from the mean of the corresponding signal and then divided by the standard deviation, mapping all input signals to a unified numerical range with a mean of 0 and a variance of 1. Standardization eliminates dimensional differences between different physical quantities, improving the convergence speed and stability of the physical information neural network training process. After the above processing, electrical signal time-series data containing d-axis voltage, d-axis current, q-axis voltage, and q-axis current are generated as input to the physical information neural network.

[0067] Step 102: Construct a physical information neural network. The composite loss function of the physical information neural network includes residual terms of the motor electromagnetic dynamics equation, residual terms of the aluminum rotor thermodynamics equation, and boundary regularization terms based on the physical properties of aluminum materials.

[0068] In some embodiments, a physical information neural network with a dual-timescale parallel structure is constructed. This network comprises two independent branches: a first branch and a slowly varying thermal parameter branch. The first branch is used to reconstruct high-frequency electromagnetic state variables such as rotor flux linkage and stator current. Its hidden layers employ 4 to 6 fully connected layers, with 64 to 128 neurons per layer. The activation function is either the Swish function or the hyperbolic tangent function, which possess smooth higher-order derivative characteristics, to ensure the accuracy of automatic differentiation calculation of the time derivative. The slowly varying thermal parameter branch is used to identify rotor resistance and rotor temperature that drift with temperature. Its hidden layers employ 2 to 3 shallow layers, with 16 to 32 neurons per layer. The output layer uses the Softplus activation function to force the output of non-negative resistance and temperature values.

[0069] In some cases, the composite loss function also includes a data loss term, which is the mean square error between the predicted stator current and the measured stator current. A boundary regularization term based on the physical properties of aluminum is used to limit the rotor resistance change rate output by the physical information neural network to no more than the physical limit determined by the thermophysical properties of aluminum, effectively suppressing the interference of high-frequency electromagnetic noise in industrial environments on parameter identification results.

[0070] Step 103: Based on the composite loss function, an alternating optimization strategy is adopted to train the physical information neural network using historical offline data in the electrical signal time series data.

[0071] In some embodiments, the training process is divided into alternating electromagnetic-dominated training phases and thermal-dominated training phases, alternating by training rounds, with even-numbered rounds being the electromagnetic-dominated training phase and odd-numbered rounds being the thermal-dominated training phase. During the electromagnetic-dominated training phase, network parameters related to slowly varying thermal parameters are frozen, and only network parameters related to rapidly varying electromagnetic states are updated, focusing on optimizing data loss terms and residual terms of the motor electromagnetic dynamics equations. During the thermal-dominated training phase, network parameters related to rapidly varying electromagnetic states are frozen, and only network parameters related to slowly varying thermal parameters are updated, focusing on optimizing residual terms of the aluminum rotor thermodynamics equations. This alternating training process is repeated until the composite loss function converges to a threshold preset according to the monitoring accuracy requirements.

[0072] In some embodiments, a first dynamic weighting coefficient corresponding to the residual term of the motor electromagnetic dynamics equation and a second dynamic weighting coefficient corresponding to the residual term of the aluminum rotor thermodynamics equation are respectively set in the composite loss function. During training, the values ​​of the first and second dynamic weighting coefficients are adjusted in real time according to the current change rate in the electrical signal time-series data. When the current change rate is detected to exceed a preset threshold, the value of the first dynamic weighting coefficient is increased and the value of the second dynamic weighting coefficient is decreased. When the current change rate is detected to be lower than the preset threshold, the value of the first dynamic weighting coefficient is decreased and the value of the second dynamic weighting coefficient is increased.

[0073] Step 104: Using the trained physical information neural network, identify the internal state parameters of the aluminum rotor induction motor that cannot be directly measured.

[0074] In some embodiments, real-time acquired electrical signal timing data is input into a trained physical information neural network. Through the network's forward propagation process, the internal state parameters of the aluminum rotor induction motor are identified. The identified internal state parameters include at least one of rotor flux linkage, rotor resistance, rotor temperature, and rotor speed.

[0075] In some cases, the trajectories of rotor resistance and rotor temperature changes are strictly constrained by the residual terms of the aluminum rotor thermodynamic equation, ensuring that their variation patterns conform to the thermodynamic properties of aluminum and eliminating spurious parameter jumps caused by measurement noise. The entire identification process does not require the installation of any temperature or speed sensors on the motor rotor side; accurate inversion of the unmeasurable internal state can be achieved solely through electrical signals from the stator side.

[0076] Step 105: Based on the real-time online data and internal state parameters in the electrical signal timing data, the transient loss of the aluminum rotor induction motor is decoupled and calculated to obtain the dynamic energy efficiency of the aluminum rotor induction motor.

[0077] In some embodiments, the instantaneous input active power of the motor is calculated based on electrical signal timing data. Stator copper losses, rotor copper losses, dynamic iron losses, stray losses, and mechanical losses are calculated based on the identified internal state parameters. The dynamic energy efficiency of the aluminum rotor induction motor is obtained by subtracting these transient losses from the instantaneous input active power.

[0078] In some examples, when calculating dynamic iron losses, the automatic differentiation function of the physical information neural network is used to calculate the time derivative of the rotor flux linkage. Based on the time derivative of the rotor flux linkage and the synchronous electrical angular velocity, the dynamic iron losses under variable frequency operation are calculated. When calculating rotor copper losses, the dynamic rotor resistance obtained in real time is used.

[0079] As described above, this embodiment achieves completely non-invasive monitoring by acquiring only the stator-side electrical signals, without altering the original motor structure. By constructing a dual-timescale physical information neural network that integrates multi-physics constraints, the modeling rigidity problem caused by the time-scale difference between electromagnetic and thermal processes is solved. The use of a physical condition-driven alternating optimization strategy and a dynamic weight adjustment mechanism improves the convergence and robustness of model training. Through hard constraints of physical laws, accurate identification of unmeasurable parameters such as rotor resistance is achieved. Based on this, precise decoupling of various transient losses under dynamic conditions is completed, effectively solving the problems of unmeasurable energy efficiency and difficult parameter drift identification in aluminum rotor induction motors under dynamic conditions, providing technical support for refined energy efficiency management of industrial motor systems.

[0080] Example 2:

[0081] The second embodiment of the present invention provides a dynamic energy efficiency monitoring method for aluminum rotor motors based on PINN. This embodiment is a supplement to the first embodiment described above, and the supplement is that the method for constructing the physical information neural network in step 102 is explained in detail. See [link to relevant documentation]. Figure 2 As shown, the specific steps include 201-203:

[0082] Step 201: Construct a physical information neural network with a dual time-scale parallel structure. The network contains a first branch and a second branch that are independent of each other. The first branch is used to estimate the rotor flux linkage, rotor current and rotor speed parameters, and the second branch is used to estimate the rotor resistance and rotor temperature parameters.

[0083] In some embodiments, a residual fully connected network is employed in the first branch network structure to reconstruct high-frequency electromagnetic state variables. The input layer of this branch receives preprocessed and coordinate-transformed electrical signal timing data, including direct-axis voltage components. quadrature axis voltage component Direct-axis current component and cross-axis current components and timestamp The hidden layers are set to 4 to 6 fully connected layers, with 64 to 128 neurons in each layer, to capture the nonlinear mapping relationships in electromagnetic transient processes. These fully connected layers are stacked in the form of residual blocks to form a residual fully connected network. The residual blocks directly transmit the input features to the output through skip connections, which can avoid the gradient vanishing problem that occurs when high-frequency electromagnetic transient signals are transmitted in deep networks, ensuring the complete extraction of electromagnetic temporal features and adapting to the rapid changes in electromagnetic states at the millisecond level.

[0084] In some cases, to meet the high-order smoothness requirements of subsequent automatic differentiation calculations of the time derivative, the hidden layer activation function is chosen to be the Swish function or the hyperbolic tangent function, and the output vector of this branch is denoted as... It includes at least the direct-axis component of the rotor flux linkage. Rotor flux cross-axis component And the predicted value of stator current.

[0085] In the forward computation flow with physical property constraints, the first branch aims to reconstruct the high-frequency changing electromagnetic state inside the motor in real time without magnetic flux sensors, and force it to conform to Maxwell's electromagnetic equations. The network receives timestamps and voltage and current vectors in a synchronously rotating coordinate system. A residual network structure is used for layer-by-layer propagation, expressed as:

[0086] ;

[0087] In the formula, For the first The output feature vector of the layer, and The first Layer weight matrix and bias vector, The activation function is used; the high-frequency transient prediction vector output by the output layer is represented as:

[0088] ;

[0089] In the formula, and These are the predicted values ​​for the rotor flux linkages of the direct-axis and quadrature-axis rotors, respectively. and These are the predicted values ​​for the direct-axis and quadrature-axis stator currents, respectively.

[0090] When the network outputs the predicted magnet link , Upon arrival, automatic differential calculation is immediately triggered at the current computation graph node: using the chain rule, reverse chain differentiation is performed along the forward-propagating computation graph to analytically solve for the predicted magnet link to the input timestamp. partial derivatives , This allows us to obtain analytical solutions for the data, avoiding the truncation error and high-frequency noise amplification caused by traditional numerical difference methods.

[0091] Then, the output and partial derivatives are substituted into the motor voltage equation to calculate the electromagnetic dynamic residual. The automatic differential calculation is performed by the automatic differential engine built into the physical information neural network. The calculation process is completed based on the computation graph generated by forward propagation, without the need for an additional differential calculation module. This is the core implementation method of sensorless electromagnetic state identification in this embodiment.

[0092] In some examples, taking the voltage residual of a direct-axis rotor as an example, the expression is:

[0093] ;

[0094] In the formula, The voltage residual of the direct-axis rotor is dimensionless. The rotor resistance referred to the stator side varies with rotor temperature; For rotor self-inductance; This is the predicted value for the flux linkage of a straight-axis rotor; This is the first-order partial derivative of the direct-axis rotor flux linkage with respect to time. Synchronous electric angular velocity; The rotor's electric angular velocity; This is the predicted value for the quadrature shaft rotor flux linkage; Mutual induction; This is the predicted value of the direct-axis stator current. In this embodiment, the flux linkage itself does not have a truth label; its calculation aims to use Maxwell's law to set convergence conditions in a multidimensional computational space, infinitely approaching the physical truth value, thus ensuring the physical residual in the above formula. Approaching zero.

[0095] In some embodiments, the second branch network structure employs a shallow long short-term memory network or a small multilayer perceptron to identify thermal parameters that drift with temperature. The input to this branch contains only timestamps. To reflect the slow-changing characteristics of the thermal process, high-frequency information such as voltage and current is not directly input. Two to three hidden layers are used, each with 16 to 32 neurons to prevent overfitting. The output layer uses the Softplus activation function to force non-negative output values. The output vector of this branch is denoted as... It includes at least the real-time rotor resistance. and rotor temperature In some cases, mutual inductance is also involved. .

[0096] In the forward computation flow with polarity constraints, the input layer of the second branch receives the global timeline sequence. Alternatively, the low-frequency power envelope feature can be obtained through smoothing and downsampling. The hidden layer extracts low-frequency heat accumulation features through a shallow long short-term memory network. In reality, resistance and absolute temperature are never negative. Conventional linear outputs are prone to outputting brief negative values ​​during the gradient descent oscillation period, causing subsequent thermodynamic equation calculations to fail. Therefore, this embodiment uses the Softplus function in this branch, expressed as:

[0097] ;

[0098] In the formula, This is the predicted value of the rotor resistance; This is the output layer weight matrix; This is the feature vector output by the last hidden layer; This is the output layer bias vector. By blocking non-physical phenomena that might cause the network to output negative resistance or negative temperature, the final output is a slowly varying parameter. ,in This is the predicted value of the rotor temperature. The output value of the soft-applied activation function is always non-negative, matching the inherent property that rotor resistance and rotor temperature are both non-negative physical quantities. This eliminates non-physical outputs at the network structure level, ensuring the validity of subsequent calculations of the thermodynamic equations.

[0099] Because strong electromagnetic interference in industrial environments can cause conventional computing networks to mistakenly interpret resistance values ​​as undergoing drastic changes within milliseconds, when the network attempts to output such unreasonable drastic changes, the absolute value of the rate of change of resistance will be misinterpreted. It must be greater than the preset physical limit. For the predicted resistance The rate of change of resistance is obtained by automatic differentiation again. The system retrieves preset physical limit parameters. , The maximum possible rate of change of resistance, calculated from the specific heat capacity of aluminum and the maximum peak current of the motor, represents the physical limit rate of temperature rise of the aluminum material. Then, the boundary regularization term is calculated:

[0100] ;

[0101] In the formula, For linear rectifier functions, defined as follows: When the measurement system introduces high-frequency pulse width modulation noise, causing the network to tend to predict rapidly changing resistance, the absolute value of the rate of change of resistance is greater than... When the ReLU function is activated, a huge penalty value is added to the total loss; conversely, under normal physical heating conditions, the difference is negative, the ReLU output is zero, and the penalty does not intervene.

[0102] Step 202: Replace the rotor resistance in the residual term of the motor electromagnetic dynamics equation with the rotor resistance output by the second branch, and replace the rotor current in the residual term of the aluminum rotor thermodynamics equation with the rotor current calculated based on the rotor flux linkage output by the first branch.

[0103] In some embodiments, to embed physical constraints, the electromagnetic dynamics equations of the induction motor and the thermodynamics equations of the aluminum rotor are rewritten in residual form and used as the composite loss function of the hard constraint embedding network. The specific implementation forms of each constraint equation are as follows.

[0104] It should be noted that in the following residual equation, the stator voltage , and stator current , All values ​​are actual measurements, rotor flux linkage , The predicted value of the first branch output is the rotor resistance. This is the predicted value output by the second branch. The stator voltage residual equation describes the quantitative relationship between the stator side voltage and current and the internal electromagnetic state, and its expression is:

[0105] ;

[0106] ;

[0107] In the above two formulas, and These are the residuals of the stator voltage equations on the direct axis and quadrature axis, respectively; , These are the components of the stator voltage on the direct axis and quadrature axis, respectively; , These are the components of the stator current on the direct axis and quadrature axis, respectively; Stator resistance; The leakage flux coefficient is defined as follows: ; For the self-sensing of the stator; For rotor self-inductance; , These are the first derivatives of the direct-axis and quadrature-axis components of the stator current with respect to time, respectively. , These are the first derivatives of the rotor flux linkage direct-axis and quadrature-axis components with respect to time, respectively. Synchronous electric angular velocity; , The rotor flux linkage components along the direct and quadrature axes are internal state parameters that cannot be directly measured and are obtained through physical constraint inversion by a physical information neural network. The time derivative of the stator current is obtained by automatically differentiating the measured stator current. This process is consistent with the calculation of the rotor flux linkage derivative and is completed using the automatic differentiation engine of the physical information neural network.

[0108] For a squirrel-cage rotor, the rotor voltage residual equation describes the rotor short-circuit characteristics, and its expression is:

[0109] ;

[0110] ;

[0111] in , These are the residuals of the rotor voltage equation on the direct axis and quadrature axis, respectively; The rotor resistance, which is referred to the stator side, is the core identification target, and its value dynamically drifts with the rotor temperature. The rotor's electric angular velocity; ω is the slip angular velocity.

[0112] The thermodynamic constraint equation for the aluminum rotor established in this embodiment quantitatively maps the rotor resistance change to the thermodynamic processes of the aluminum material:

[0113] ;

[0114] In the formula, The instantaneous rate of change of rotor resistance over time; is the temperature coefficient of resistance of aluminum; The cold-state reference resistance of the rotor at the reference temperature; For rotor heat capacity; This refers to the instantaneous power loss on the rotor side. This refers to the instantaneous temperature of the rotor. The ambient temperature; This represents the thermal resistance of the rotor to the environment. This equation forces the change trajectory of the rotor resistance to conform to the thermodynamic laws of aluminum, transforming the resistance identification problem into a temperature estimation problem.

[0115] The residual terms of the aluminum rotor thermodynamic equations include the residuals of the dynamic thermal equilibrium differential equations:

[0116] ;

[0117] in For rotor heat capacity; This is the derivative of the rotor temperature with respect to time; , These are the components of the rotor current along the direct axis and quadrature axis, respectively. The real-time rotor resistance varies with time; This refers to the instantaneous temperature of the rotor. The ambient temperature; This represents the thermal resistance for heat dissipation. The left side of the equation represents the heat absorbed by the rotor's heat capacity per unit time, and the first term on the right side... The second item on the right represents the heat generated by rotor copper losses. This represents the heat dissipated by the rotor into the environment. The rotor temperature change rate is shown on the left. By controlling the rotor temperature output from the second branch Obtained through automatic differentiation. The rotor current component in the heating power on the right. , The following can be derived by inversely calculating the rotor flux output from the first branch and the measured stator current:

[0118] ;

[0119] Step 203: Weight the residual terms of the motor electromagnetic dynamics equation and the aluminum rotor thermodynamics equation and then add them to the composite loss function.

[0120] In some embodiments, the composite loss function It is constructed as a weighted sum of data loss, electromagnetic physical loss, thermophysical loss, and boundary regularization loss:

[0121] ;

[0122] In the formula, Total loss; This refers to data loss, specifically the mean square error between the predicted stator current and the actual measured stator current. The first dynamic weighting coefficient is dimensionless. Electromagnetic physical loss is defined as residual. , , , The sum of squares; The second dynamic weighting coefficient is dimensionless. The thermophysical loss is defined as the square of the residual in the dynamic heat balance equation; is the weight coefficient of the regularization term, which is dimensionless; This is the boundary regularization term.

[0123] It should be noted that the two branches are not directly connected in the computational structure, but are cross-coupled through the physical constraint equations in the loss function. The electromagnetic physical loss includes the rotor resistance output by the second branch. If the resistance value is inaccurate, the electromagnetic residual will increase, and the weight parameters of the second branch will be updated during gradient backpropagation. Similarly, the thermophysical loss includes the rotor current information output by the first branch. If the current is inaccurate, the thermal residual will increase, and the weight parameters of the first branch will be updated during gradient backpropagation. .

[0124] In a two-branch collaborative computing process, the time constants of electromagnetic changes are on the order of milliseconds, while those of temperature changes are on the order of minutes, resulting in a significant difference in their gradients. This embodiment employs an alternating optimization timing method:

[0125] In the first time-series state, i.e., the electromagnetic state, freeze the weight parameters of the second branch. Forward propagation is performed only on the first branch to calculate the total loss (data loss plus electromagnetic physical loss). The gradient update during backpropagation only applies to the weights of the first branch. In the second time-series state, i.e., the thermal resistance state, the weight parameters of the first branch are frozen. Using the high-frequency electromagnetic data locked in the first branch as environmental conditions, forward propagation is performed on the second branch to calculate the total loss (thermophysical loss plus boundary regularization loss), and backpropagation is used to update the weights of the second branch. The above-described bi-branch alternating optimization process terminates when the composite loss function converges, and the number of alternation rounds is adaptively adjusted according to the motor model and operating conditions.

[0126] The logic behind real-time intervention and weight correction of physical constraints lies in: physical constraint terms , As a component of the loss function, it directly participates in the calculation of the gradient. Its gradient expression is:

[0127] ;

[0128] In the formula, For the loss function with respect to network parameters The gradient; This is the partial derivative of the data loss term with respect to the parameters; The weights for the physical loss term; This represents the partial derivative of the physical constraint loss term with respect to the parameters. When the network outputs a mathematically well-fitting result in a forward iteration that violates the law of energy conservation, although the data loss... The difference is small, but the physical residual increases dramatically. This causes the gradient vector calculated during backpropagation to deviate from the purely data-driven direction, forcing the optimizer to pull the network model back into the true solution space that conforms to physical laws when updating the weight parameters.

[0129] During the training phase, step 101 is performed to acquire electrical signal time-series data as a training sample set. The training sample set contains historical offline data across multiple time windows, with each time window containing continuously sampled direct-axis voltages. quadrature axis voltage Direct-axis current quadrature axis current and corresponding timestamp The data from each time window are organized into feature vectors, and the measured stator current within the same time window is used as the supervision label. The feature vectors are then input into the first and second branches respectively, and training is performed according to the above alternating optimization process until the composite loss function converges.

[0130] During the usage phase, i.e., after training is complete, in online monitoring, backpropagation and parameter updates are no longer performed. Real-time online data from the acquired electrical signal time-series data is input into the trained physical information neural network, and only forward propagation calculations are performed. The first branch outputs the rotor flux linkage based on the current voltage, current, and timestamp. , and stator current prediction value , The second branch outputs the rotor resistance based on the current timestamp. and rotor temperature These internal state parameters constitute the identification results, which are then fed into step 105 for transient loss decoupling calculation, ultimately outputting the dynamic energy efficiency. During the usage phase, real-time collected voltage and current data are input to the network point-by-point as individual sampling points, with timestamps using the cumulative time relative to the training start time or a standardized value of the system clock. The network's forward propagation no longer relies on a time window, directly outputting the current rotor flux linkage, predicted current values, rotor resistance, and temperature. The single-point sampling input method used in the usage phase performs real-time inference based on the features learned through the time window during the training phase, preserving the physical constraints of network training while meeting the millisecond-level real-time monitoring response requirements of industrial sites.

[0131] It should be noted that in this embodiment, the training phase and the usage phase are separated: the training phase utilizes historical offline data and is completed on the cloud or high-performance computing platform, resulting in a relatively long training cycle; the usage phase deploys the trained network model to edge computing devices, performing only forward inference to meet the real-time requirements of industrial sites. The electrical signal time-series data used in the two phases come from different sources: the training data is historical data collected offline, while the usage data is real-time online data.

[0132] As described above, this embodiment constructs a physical information neural network with a dual-timescale parallel structure, deeply coupling the first and second branches at the physical loss function level to achieve collaborative modeling of electromagnetic millisecond-level and thermal minute-level responses. This scheme enables the neural network to simultaneously and accurately reconstruct the high-frequency electromagnetic state and identify rotor resistance and rotor temperature that drift with temperature, using only stator terminal electrical signals, without any temperature sensors. The two branches are mutually constrained and corrected through a cross-coupling mechanism of physical residuals, effectively avoiding gradient imbalance and convergence difficulties that occur when a single network learns multi-timescale physical processes. This improves the accuracy and robustness of parameter identification under dynamic operating conditions, providing a model foundation for realizing non-invasive, high-dynamic-response energy efficiency monitoring of aluminum rotor motors.

[0133] Example 3:

[0134] The third embodiment of this invention provides a dynamic energy efficiency monitoring method for aluminum rotor motors based on PINN. This embodiment supplements the aforementioned second embodiment, with the supplement focusing on a detailed explanation of the dynamic weight coefficient adjustment method for the physical information neural network in step 102. See [link to relevant documentation]. Figure 3 As shown, the specific steps include 301-302:

[0135] Step 301: In the composite loss function, set the first dynamic weighting coefficient corresponding to the residual term of the motor electromagnetic dynamics equation and the second dynamic weighting coefficient corresponding to the residual term of the aluminum rotor thermodynamics equation.

[0136] In some embodiments, the composite loss function The expression is:

[0137] ;

[0138] in: This represents the total loss during neural network training. The mean square error between the predicted stator current and the measured stator current is due to data loss. These are the residual terms of the electromagnetic dynamics equations of the electric motor; This is the residual term in the thermodynamic equation for the aluminum rotor; This is a boundary regularization term based on the physical properties of aluminum materials; These are the weight coefficients for the regularization term; Defined as the first dynamic weighting coefficient, corresponding to the residual term in the motor electromagnetic dynamics equation; Defined as the second dynamic weighting coefficient, corresponding to the residual term of the aluminum rotor thermodynamic equation. These two weighting coefficients are not fixed hyperparameters, but rather dynamic variables that change with time or operating conditions; their specific values ​​are determined in step 302. Dynamic weighting coefficients and Adjust the learning priority of balancing electromagnetic and thermal processes based on the real-time operating conditions of the motor; total loss The optimization objective of PINN training is to minimize the total loss so that the network can simultaneously meet the fitting requirements of measurement data and the constraints of physical laws, thereby achieving physics-guided data-driven learning and ensuring the model's generalization ability and physical interpretability.

[0139] In some examples, the expression for the boundary regularization term is:

[0140] ;

[0141] in, The maximum possible rate of change of resistance is determined by the specific heat capacity of aluminum and the maximum current of the motor. The activation function is a linear rectification function. This regularization term, based on the thermophysical properties of aluminum, limits the maximum rate of change of rotor resistance, effectively suppresses spurious rotor resistance jumps caused by high-frequency PWM noise from the power grid, and improves the robustness of the model under noisy data in industrial settings.

[0142] In some cases, the first dynamic weight coefficient The initial value is set to 1.0, and the second dynamic weighting coefficient... The initial value is set to 0.1 to reflect the dominance of electromagnetic transient loss during the initial training phase. Boundary regularization term weight coefficients. Set to 0.01 to provide adequate material physical constraints without masking the main loss terms.

[0143] Step 302: The values ​​of the first dynamic weighting coefficient and the second dynamic weighting coefficient are adaptively adjusted according to the rate of change of current in the electrical signal timing data.

[0144] In some embodiments, the stator current component in the electrical signal timing data is acquired in real time, and the current change rate is calculated. Specifically, for the direct-axis current component... and cross-axis current components Calculate their first derivatives with respect to time, and take the root mean square of both as the combined rate of change of current. ,Right now The rate of change is precisely calculated at each sampling time by the automatic differentiation function of the physical information neural network.

[0145] In some cases, a transient threshold is set. Its value is pre-calibrated based on the motor's rated current and typical load change rate. When the combined current change rate... When the threshold is exceeded, it is determined that the current operating condition is transient, such as motor startup, sudden load change, or variable frequency speed regulation. At this time, the first dynamic weighting coefficient is automatically adjusted. Increase it to nearly 1.0, while simultaneously adjusting the second dynamic weighting coefficient. Reduce it to near 0. The physical significance of this setting is that, under transient conditions, the electromagnetic response is dominant, and the network should concentrate its computing power to fit the rapidly changing electromagnetic states such as flux linkage and current, while treating the slowly changing temperature as an approximate constant, thereby avoiding the interference of thermal residual gradient on the convergence of electromagnetic parameters.

[0146] When the overall current change rate When the value falls below a preset threshold, the system is determined to be in a steady-state condition. At this point, the system smoothly reduces the first dynamic weight coefficient. And significantly increase the second dynamic weighting coefficient. Typical values ​​are It dropped to 0.1. The value is increased to 1.0. The physical significance of this is that, under steady-state conditions, the electromagnetic state tends to be constant, and the network should utilize the accumulated temperature rise over a long period to precisely calibrate the rotor resistance. and rotor temperature This ensures the accuracy of thermal parameter identification.

[0147] In some cases, the adjustment of the weighting coefficients uses a smooth transition rather than a step jump. Specifically, the weighting coefficients are defined as the rate of change of current. Continuous functions, for example ,in For the threshold, To adjust the slope, continuous adjustments are made to avoid training instability caused by sudden weight changes. Simultaneously, the calculation window length for the current rate of change can be dynamically adjusted according to the actual sampling rate. For example, the variance of the current within the sliding time window can be used instead of the instantaneous derivative to enhance noise immunity.

[0148] This embodiment addresses the convergence problem caused by the imbalance of electromagnetic and thermal residual gradients in multi-physics coupled training by employing a physical condition feedforward weight scheduling mechanism based on the rate of change of current. It achieves a dynamic balance in the learning priorities of electromagnetic and thermal processes. Under transient conditions, priority is given to ensuring the accuracy of electromagnetic state identification, while under steady-state conditions, the focus is on calibrating rotor thermal parameters, thus improving the convergence speed of model training and the parameter identification accuracy under dynamic conditions. This method has extremely low computational cost, requires no additional sensor data, and can be directly deployed in edge computing devices in industrial settings, demonstrating strong engineering practicality.

[0149] Example 4:

[0150] The fourth embodiment of the present invention provides a dynamic energy efficiency monitoring method for aluminum rotor motors based on PINN. This embodiment is a supplement to the first embodiment described above. The supplement is that the method of using an alternating optimization strategy and training the physical information neural network with electrical signal time-series data in step 103 is explained in detail. See [link to relevant documentation]. Figure 4 As shown, the specific steps include 401 to 404:

[0151] Step 401: Divide the training process into an alternating electromagnetic-dominated training phase and a thermal-dominated training phase.

[0152] In some embodiments, the training is divided into phases according to the number of training rounds, with even-numbered training rounds being the electromagnetic-dominated training phase and odd-numbered training rounds being the thermal-dominated training phase. Each training round includes a complete forward and backward propagation process, and a network parameter update is completed using a batch of electrical signal timing data.

[0153] In some examples, the training batch size is set to 32 to 128, adjusted according to the computing power of the edge computing device. The initial training phase consists of 10 warm-up rounds, during which all network parameters are updated simultaneously to allow the network to converge quickly to an initial feasible solution, after which the alternating training phase begins.

[0154] Step 402: During the electromagnetic-dominated training phase, freeze the network parameters of the second branch and update the network parameters related to the rapidly changing electromagnetic state based on the adjusted first dynamic weight coefficients.

[0155] In some embodiments, at the start of the electromagnetically dominated training phase, all network weight parameters of the slowly varying thermal parameter branch are... Set to non-trainable, i.e., frozen state. In this state, the gradients from backpropagation will not flow into the slowly varying branch; only the weight parameters of the rapidly varying electromagnetic state branch will be updated. The first dynamic weighting coefficient in the composite loss function. Set to a larger value, the second dynamic weighting coefficient It is set to a smaller value. Specifically, according to the weighting coefficient determined by the rate of change of current in Example 3, under transient conditions... Approaching 1.0, Approaching 0, the optimization objective is to minimize data loss. and electromagnetic physical loss The weighted sum of thermophysical losses The contribution was significantly suppressed.

[0156] In some examples, an adaptive moment estimation optimizer is used during the electromagnetic-dominated training phase, with an initial learning rate of 0.001, which decays to 0.9 times the original rate every 500 epochs. During training, the input is electrical signal timing data, and the fast-variable branch outputs the rotor flux linkage. , And the predicted stator current value, substituted into the stator voltage residual. , and rotor voltage residual , Calculate the electromagnetic physical loss. This stage continues for a fixed number of iterations, or until the rate of change of the electromagnetic physical loss falls below a set threshold.

[0157] Step 403: During the heat-dominated training phase, freeze the network parameters of the first branch and update the network parameters related to the slowly varying heat parameters based on the adjusted second dynamic weight coefficients;

[0158] In some embodiments, at the start of the heat-dominated training phase, all network weight parameters of the rapidly changing electromagnetic state branch are... Freeze, only update the weight parameters of the slowly varying thermal parameter branch. The second dynamic weighting coefficient in the composite loss function Set to a larger value, first dynamic weight coefficient It is set to a smaller value. Under steady-state conditions, Approaching 1.0, Approaching 0. The optimization objective focuses on minimizing thermophysical losses. This means bringing the residuals of the dynamic thermal equilibrium differential equations close to zero. At this point, the calculation of thermophysical losses requires using the rotor flux linkage output from the rapidly changing branch to inversely deduce the rotor current. and However, since the parameters of the fast branch are frozen, its output is used as a fixed value in the calculation of heat loss, and the gradient is only propagated back to the slow branch through the thermal residual.

[0159] In some examples, the thermally-dominated training phase also employs an adaptive moment estimation optimizer with a learning rate set to 0.0005, lower than the initial learning rate in the electromagnetic phase, to reflect the slow-changing nature of thermal parameters. This phase continues with a fixed number of iterations, or until the rate of change of the thermophysical loss falls below a set threshold. During training, the composite loss function is monitored in real time. If the value fluctuates or diverges, the learning rate should be appropriately reduced or the number of training rounds per stage should be shortened.

[0160] Step 404: Repeat the above alternating update process until the composite loss function converges to the preset threshold.

[0161] In some embodiments, completing one electromagnetic-dominated training phase and one thermal-dominated training phase constitutes an alternating cycle. At the end of each cycle, the composite loss function value on the validation set is evaluated. If the composite loss function value continues to decrease and falls below a preset convergence threshold, or if the loss value no longer decreases significantly over multiple consecutive cycles, training is considered converged, and alternating optimization is stopped. Otherwise, the next alternating cycle continues.

[0162] In some cases, the convergence threshold is set to one percent of the initial loss value, or to an absolute threshold, for example... To prevent overfitting, an early stopping strategy is employed during training, meaning training is terminated when the validation set loss no longer decreases after 10 consecutive alternating cycles. After training, the final weight parameters of the rapidly changing electromagnetic state branch and the slowly changing thermal parameter branch are saved. and This is used for subsequent real-time inference.

[0163] This embodiment effectively addresses the rigidity issue in network training caused by the significant timescale difference between electromagnetic and thermal processes by employing a parameter update strategy that alternates between training rounds. It also avoids gradient imbalance when a single network simultaneously learns both fast and slow processes. This strategy simulates the numerical analysis approach of first solving the electromagnetic steady-state solution and then correcting the thermal parameters, significantly improving the convergence speed and stability of the model training while ensuring the accuracy of electromagnetic state and thermal parameter identification. It is suitable for complex dynamic working conditions with significant noise in industrial environments.

[0164] Example 5:

[0165] The fifth embodiment of this invention provides a dynamic energy efficiency monitoring method for an aluminum rotor motor based on PINN. This embodiment supplements the first embodiment described above, with the supplement focusing on a more detailed explanation of the decoupling calculation method for the transient losses of the aluminum rotor induction motor in step 105. See also Figure 5 As shown, the specific steps include 501 to 503:

[0166] Step 501: Calculate the instantaneous input active power of the motor based on the electrical signal timing data.

[0167] In some embodiments, instantaneous input active power Calculated in a synchronously rotating coordinate system, based on the direct-axis voltage component after Park transformation. quadrature axis voltage component Direct-axis current component and cross-axis current components The expression for instantaneous input active power is:

[0168] ;

[0169] in: This refers to the real-time input active power of the motor. This represents the component of the stator voltage along the d-axis. This represents the q-axis component of the stator voltage; This represents the component of the stator current along the d-axis. This represents the component of the stator current along the q-axis. For time; coefficient It is derived from the three-phase power conversion under the equal power Park transformation; this calculation is performed in each sampling period, is not affected by transient reactive power exchange, and directly reflects the instantaneous active power input to the motor from the electrical port.

[0170] In some cases, power calculations are performed with a fixed sampling period, consistent with the electrical signal acquisition period. During the calculation, the voltage and current signals are subjected to moving average filtering to remove high-frequency noise interference with the power calculation results.

[0171] Step 502: Calculate stator copper loss, rotor copper loss, dynamic iron loss, stray loss and mechanical loss based on internal state parameters.

[0172] In some embodiments, the stator copper loss is determined by the stator resistance and the square of the stator current amplitude, as expressed by:

[0173] ;

[0174] in: For transient stator copper loss; Stator resistance;

[0175] In some cases, rotor copper losses are calculated based on rotor current and real-time identified dynamic rotor resistance. The rotor current is derived by inversely calculating the rotor flux linkage and stator current, and its expression is as follows:

[0176] ;

[0177] ;

[0178] in: This represents the component of the rotor current along the d-axis. This represents the q-axis component of the rotor current. This represents the component of the rotor flux linkage along the d-axis. This represents the component of the rotor flux linkage along the q-axis. Mutual induction; For rotor self-inductance; Transient rotor copper loss; The rotor resistance is identified in real time. Rotor copper loss is one of the main losses in aluminum rotor motors. This embodiment uses the real-time identified, thermodynamically constrained rotor resistance to calculate rotor copper loss, which solves the problem of underestimating copper loss and overestimating efficiency under heavy load caused by the traditional method using cold resistance.

[0179] In some cases, dynamic iron loss The calculation utilizes the automatic differentiation function of a physical information neural network. The direct-axis component of the rotor flux linkage output by the fast-changing branch is... Calculate its first derivative with respect to time. Combined with synchronous electric angular velocity The dynamic iron loss expression is:

[0180] ;

[0181] in: Transient iron loss under dynamic frequency conversion conditions; The equivalent iron loss resistance is a known constant. The time derivative of the rotor flux linkage; The first term of the equation corresponds to the dynamic eddy current loss caused by rapid changes in the magnetic flux linkage, and the second term corresponds to the hysteresis and eddy current loss of the steady-state rotating magnetic field. It can accurately capture the additional iron loss during frequency conversion or load change.

[0182] In some cases, stray losses and mechanical losses are adaptively separated through the air gap torque and the residuals of the mechanodynamic equations. The expression for the air gap torque is:

[0183] ;

[0184] in: This refers to the air gap torque. This represents the number of pole pairs of the motor.

[0185] In some examples, the automatic differentiation function of the physical information neural network is utilized to directly calculate the time derivative of the rotor flux linkage without the need for additional numerical difference calculations, thus ensuring the accuracy and real-time performance of the derivative calculation. This method can accurately capture the dynamic iron loss generated by rapid changes in flux linkage under variable frequency operating conditions, overcoming the limitation of traditional steady-state iron loss calculation formulas failing in transient processes.

[0186] Step 503: Subtract stator copper loss, rotor copper loss, dynamic iron loss, stray loss and mechanical loss from the instantaneous input active power to obtain the dynamic energy efficiency of the aluminum rotor induction motor.

[0187] In some embodiments, dynamic energy efficiency Defined as the ratio of output mechanical power to input active power, since output mechanical power is difficult to measure directly, it is indirectly obtained by subtracting all losses from the input active power. The expression is:

[0188] ;

[0189] in: This represents the true transient efficiency of the motor. Stray loss refers to the additional losses caused by harmonics, leakage flux, etc. Mechanical losses include frictional losses and wind losses; this formula calculates the real-time transient efficiency of the motor under dynamic operating conditions, and can capture efficiency changes within milliseconds of load changes and the long-term efficiency decay trend caused by rotor temperature rise. It is the final output indicator of the dynamic energy efficiency monitoring of this invention.

[0190] In some embodiments, mechanical loss and stray loss The residuals are obtained by combining pre-calibrated characteristic curves with online residual correction, and the specific determination method is as follows.

[0191] Mechanical loss This includes bearing friction loss and rotor wind resistance loss, the values ​​of which depend on the rotor speed. The relationship between mechanical loss and speed is pre-calibrated through offline no-load testing: Under no-load conditions, the power supply frequency is changed to adjust the speed, and the input active power and stator copper loss are recorded at each stable speed. The remaining power after subtracting the stator copper loss from the input active power is taken as the mechanical loss at that speed, thus obtaining the discrete points of mechanical loss as a function of speed. A continuous characteristic curve is then obtained through polynomial fitting.

[0192] Stray loss Stray losses refer to all losses other than stator copper losses, rotor copper losses, iron losses, and mechanical losses, including harmonic losses and additional losses caused by leakage flux. Since stray losses are difficult to model directly, an adaptive residual estimation method can be used to determine their values. Specifically, after calculating stator copper losses, rotor copper losses, dynamic iron losses, and mechanical losses, the air gap power is calculated based on the air gap torque and rotor speed: the air gap torque is calculated using the electromagnetic torque formula, the rotor speed is identified by a physical information neural network, and the air gap power equals the product of the air gap torque and the rotor mechanical angular velocity. Subtracting the rotor copper losses from the air gap power yields the output mechanical power. At this point, the remaining power after subtracting the stator copper losses, dynamic iron losses, mechanical losses, and output mechanical power from the input active power is the stray loss.

[0193] In some cases, to eliminate dynamic errors caused by rotational inertia under dynamic operating conditions, mechanical dynamics equations can be introduced for correction. The net torque of the motor equals the air gap torque minus the load torque and friction torque, and the net torque is proportional to the product of rotational inertia and angular acceleration. By calculating angular acceleration in real time, the load torque and friction loss can be dynamically separated, thereby improving the estimation accuracy of stray losses. Under steady-state operating conditions, stray losses can be approximately considered to be proportional to the square of the stator current, and its proportionality coefficient is obtained through online regression of historical data.

[0194] In some embodiments, a complete loss decoupling and energy efficiency calculation is performed once per sampling period, outputting a dynamic energy efficiency curve with millisecond-level resolution. The calculation results are stored in real time and uploaded to a cloud platform for subsequent energy efficiency analysis and energy-saving optimization control.

[0195] This embodiment performs decoupled transient loss calculations using internal state parameters identified by a physical information neural network, completely independent of any empirical lookup tables or steady-state assumptions. By employing real-time dynamic rotor resistance to calculate rotor copper losses, it accurately reflects the impact of rotor temperature rise on losses, avoiding the problem of inflated efficiency assessments under heavy load conditions caused by traditional methods using cold-state resistance. By utilizing automatic differentiation technology to calculate dynamic iron losses, it achieves precise capture of additional losses under variable frequency operation. This method can accurately obtain the true transient energy efficiency of aluminum rotor induction motors under various dynamic operating conditions, providing reliable data support for refined energy efficiency management and energy-saving optimization of industrial motor systems.

[0196] Example 6:

[0197] Another embodiment of this application relates to a dynamic energy efficiency monitoring system for an aluminum rotor motor based on PINN. The implementation details of this embodiment's dynamic energy efficiency monitoring system based on PINN are described below. The following details are provided for ease of understanding and are not essential for implementing this solution. A schematic diagram of this embodiment's dynamic energy efficiency monitoring system based on PINN can be seen as follows: Figure 6 As shown, it includes a data acquisition module 601, a network construction module 602, a model training module 603, a state identification module 604, and an energy efficiency calculation module 605.

[0198] The data acquisition module 601 is used to acquire the timing data of electrical signals on the stator side of the aluminum rotor induction motor. This timing data consists of time-series measurements of the three-phase voltage and current on the stator side. This module includes a voltage transformer, a current transformer, an analog-to-digital converter (ADC), and a digital signal processor (DSP). The voltage and current transformers are installed on the stator power supply line at the motor control cabinet end, continuously acquiring the three-phase voltage and current signals. The ADC converts the analog signals into digital quantities, and the DSP performs Park transform and normalization on the digital quantities to generate timing data containing direct-axis voltage, quadrature-axis voltage, direct-axis current, and quadrature-axis current components. The processed data is transmitted to the network construction module 602 via a bus.

[0199] The network construction module 602 is used to construct a physical information neural network. The composite loss function of the physical information neural network includes residual terms of the motor electromagnetic dynamics equation, residual terms of the aluminum rotor thermodynamics equation, and boundary regularization terms based on the physical properties of aluminum materials. This module runs on a cloud server or high-performance computing platform, generates a dual-timescale parallel neural network according to a pre-defined network topology, including a fast-changing electromagnetic state branch and a slow-changing thermal parameter branch, and configures a composite loss function that includes residual terms of the motor electromagnetic dynamics equation, residual terms of the aluminum rotor thermodynamics equation, and boundary regularization terms based on the physical properties of aluminum materials. The completed network model is saved as a deployable model file.

[0200] The model training module 603 is used to train the physical information neural network based on a composite loss function and an alternating optimization strategy, utilizing historical offline data from electrical signal time-series data. This module reads the electrical signal time-series data provided by the data acquisition module 601 as training samples and performs backpropagation optimization according to an alternating update strategy of electromagnetic-dominated and thermal-dominated training phases until the composite loss function converges to a preset threshold. The training process can be completed in the cloud, and the trained network weight parameters are fixed and transmitted to the state identification module 604.

[0201] The state identification module 604 is used to identify the non-measurable internal state parameters of the aluminum rotor induction motor using a trained physical information neural network. Deployed in an industrial edge computing gateway, this module receives real-time electrical signal timing data output from the data acquisition module 601, performs forward inference calculations of the network, and outputs at least one internal state parameter from rotor flux linkage, rotor resistance, rotor temperature, and rotor speed. The identification results are updated once per second and transmitted in real-time to the energy efficiency calculation module 605.

[0202] The energy efficiency calculation module 605 is used to decouple and calculate the transient losses of the aluminum rotor induction motor based on real-time online data from the electrical signal timing data and internal state parameters, thereby obtaining the dynamic energy efficiency of the aluminum rotor induction motor. This module receives electrical signal timing data from the data acquisition module 601 and internal state parameters from the state identification module 604. Following the loss decoupling method described in Embodiment 5, it sequentially calculates the instantaneous input active power, stator copper loss, rotor copper loss, dynamic iron loss, stray loss, and mechanical loss, ultimately outputting the dynamic energy efficiency percentage. The calculation results can be uploaded to a monitoring system or cloud platform via an industrial communication protocol for energy efficiency analysis and optimized control.

[0203] It is worth mentioning that all modules involved in this embodiment are logical modules. In practical applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. Furthermore, to highlight the innovative aspects of this application, this embodiment does not introduce units that are not closely related to solving the technical problems proposed in this application; however, this does not mean that other units are absent in this embodiment.

[0204] Example 7:

[0205] Another embodiment of this application relates to an electronic device, such as... Figure 7 As shown, it includes: at least one processor 701; and a memory 702 communicatively connected to the at least one processor 701; wherein the memory 702 stores instructions executable by the at least one processor 701, the instructions being executed by the at least one processor 701 to enable the at least one processor 701 to execute the PINN-based aluminum rotor motor dynamic energy efficiency monitoring method in the above embodiments.

[0206] The memory and processor are connected via a bus, which can include any number of interconnecting buses and bridges, connecting various circuits of one or more processors and memories. The bus can also connect various other circuits, such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and will not be described further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by the processor is transmitted over the wireless medium via an antenna, which further receives data and transmits it to the processor.

[0207] The processor manages the bus and general processing, and also provides various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory, on the other hand, is used to store data used by the processor during operation.

[0208] Example 8:

[0209] Another embodiment of this application relates to a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the method embodiments described above.

[0210] That is, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0211] Those skilled in the art will understand that the above embodiments are specific embodiments for implementing this application, and in practical applications, various changes can be made to them in form and detail without departing from the spirit and scope of this application.

Claims

1. A dynamic energy efficiency monitoring method for aluminum rotor motors based on PINN, characterized in that, include: Acquire electrical signal timing data on the stator side of an aluminum rotor induction motor, wherein the electrical signal timing data is time series measurement data of the three-phase voltage and three-phase current on the stator side; A physical information neural network is constructed, wherein the composite loss function of the physical information neural network includes residual terms of the motor electromagnetic dynamics equation, residual terms of the aluminum rotor thermodynamics equation, and boundary regularization terms based on the physical properties of aluminum material. The construction of the physical information neural network includes: A physical information neural network with a dual time-scale parallel structure is constructed. The physical information neural network includes a first branch and a second branch that are independent of each other. The first branch is used to estimate rotor flux, rotor current and rotor speed parameters, and the second branch is used to estimate rotor resistance and rotor temperature parameters. Replace the rotor resistance in the residual term of the motor electromagnetic dynamics equation with the rotor resistance output by the second branch, and replace the rotor current in the residual term of the aluminum rotor thermodynamics equation with the rotor current calculated based on the rotor flux linkage output by the first branch. The residual terms of the motor electromagnetic dynamics equation and the residual terms of the aluminum rotor thermodynamics equation are weighted and then added to the composite loss function; In the composite loss function, a first dynamic weighting coefficient corresponding to the residual term of the motor electromagnetic dynamics equation and a second dynamic weighting coefficient corresponding to the residual term of the aluminum rotor thermodynamics equation are respectively set; Based on the aforementioned composite loss function, an alternating optimization strategy is employed to train the physical information neural network using historical offline data from the electrical signal time-series data, including: The training process is divided into alternating electromagnetic-dominated training phases and thermal-dominated training phases. During the electromagnetic-dominated training phase, the network parameters of the second branch are frozen, and the network parameters related to the rapidly changing electromagnetic state are updated based on the adjusted first dynamic weight coefficients. During the heat-dominated training phase, the network parameters of the first branch are frozen, and the network parameters related to the slowly changing heat parameters are updated based on the adjusted second dynamic weight coefficients. Repeat the above alternating update process until the composite loss function converges to a preset threshold. Using the trained physical information neural network, the non-measurable internal state parameters of the aluminum rotor induction motor are identified; Based on the real-time online data in the electrical signal timing data and the internal state parameters, the transient loss of the aluminum rotor induction motor is decoupled and calculated to obtain the dynamic energy efficiency of the aluminum rotor induction motor.

2. The method according to claim 1, characterized in that, The acquisition of electrical signal timing data on the stator side of the aluminum rotor induction motor includes: Real-time measurement data of three-phase voltage and three-phase current on the stator side of the aluminum rotor induction motor were collected; The real-time measurement data of the three-phase voltage and three-phase current are subjected to synchronous rotating coordinate system transformation; The transformed electrical signal is standardized to generate the timing data of the electrical signal.

3. The method according to claim 1, characterized in that, The construction of the physical information neural network also includes: The values ​​of the first dynamic weighting coefficient and the second dynamic weighting coefficient are adaptively adjusted according to the rate of change of current in the electrical signal timing data.

4. The method according to claim 3, characterized in that, The values ​​of the first dynamic weighting coefficient and the second dynamic weighting coefficient are adaptively adjusted based on the rate of change of current in the electrical signal time-series data, including: Obtain the rate of change of current in the timing data of the electrical signal; When the rate of change of current exceeds a preset threshold, the value of the first dynamic weighting coefficient is increased and the value of the second dynamic weighting coefficient is decreased. When the rate of change of current is lower than the preset threshold, the value of the first dynamic weight coefficient is reduced and the value of the second dynamic weight coefficient is increased.

5. The method according to claim 1, characterized in that, The decoupling calculation of the transient losses of the aluminum rotor induction motor to obtain the dynamic energy efficiency of the aluminum rotor induction motor includes: The instantaneous input active power of the motor is calculated based on the electrical signal timing data; Based on the aforementioned internal state parameters, the stator copper loss, rotor copper loss, dynamic iron loss, stray loss, and mechanical loss are calculated respectively. The dynamic energy efficiency of the aluminum rotor induction motor is obtained by subtracting the stator copper loss, rotor copper loss, dynamic iron loss, stray loss, and mechanical loss from the instantaneous input active power.

6. The method according to claim 5, characterized in that, The calculation of dynamic iron loss includes: The time derivative of the rotor flux linkage is calculated using the automatic differentiation function of the physical information neural network. Based on the time derivative of the rotor flux linkage and the synchronous electric angular velocity, the dynamic iron loss under variable frequency operation is calculated.

7. A dynamic energy efficiency monitoring system for an aluminum rotor motor based on PINN for use in the method according to claim 1, characterized in that, include: The data acquisition module is used to acquire electrical signal timing data of the stator side of the aluminum rotor induction motor. The electrical signal timing data is the time series measurement data of the three-phase voltage and three-phase current on the stator side. The network construction module is used to construct a physical information neural network. The composite loss function of the physical information neural network includes residual terms of the motor electromagnetic dynamics equation, residual terms of the aluminum rotor thermodynamics equation, and boundary regularization terms based on the physical properties of aluminum material. The model training module is used to train the physical information neural network based on the composite loss function and an alternating optimization strategy, using historical offline data from the electrical signal time series data. The state identification module is used to identify the non-measurable internal state parameters of the aluminum rotor induction motor by utilizing the trained physical information neural network. The energy efficiency calculation module is used to decouple the transient loss of the aluminum rotor induction motor based on the real-time online data in the electrical signal time series data and the internal state parameters, so as to obtain the dynamic energy efficiency of the aluminum rotor induction motor.

Images