Method and system for wide-temperature starting of industrial internet brushless motor

Abstract

The invention relates to the technical field of industrial automation and motor control, discloses a wide-temperature starting method and system for an industrial internet brushless motor, and aims to solve the problems of starting parameter mismatch, stalling or torque oscillation caused by physical delay of a temperature sensor and a reactive control strategy in the prior art. The method comprises the following steps: collecting thermal state and electric state data of a motor in real time, and constructing a multi-dimensional time sequence characteristic matrix; predicting a future temperature trajectory through an eight-bit integer quantization LSTM model deployed at an edge end; a global thermal environment early warning instruction issued by the cloud is fused, and a corrected temperature track is generated; based on the track, calculating an optimal starting control parameter by using a motor physical parameterization model and executing starting; and after starting is completed, uploading a data log for iteration of the cloud model. According to the application, the starting reliability and the control performance of the brushless motor in a wide temperature range are remarkably improved through an edge cloud collaborative zero-lag temperature prediction and prospective parameter generation mechanism.

Description

Technical Field

[0001] This invention belongs to the field of industrial automation and motor control technology, specifically relating to a method and system for wide-temperature start-up of a brushless motor in the industrial internet. Background Technology

[0002] With the deepening application of the Industrial Internet in the field of intelligent manufacturing, higher requirements are placed on the environmental adaptability and starting reliability of core drive units such as brushless DC motors. Especially under wide temperature range conditions, the resistance of the motor windings, the magnetic flux characteristics of the permanent magnet, and the friction state of mechanical parts will all change significantly with temperature. If the starting control parameters fail to match these changes in real time, problems such as starting failure, overcurrent, or torque fluctuations can easily occur.

[0003] Current common control schemes mainly rely on real-time feedback from temperature sensors for adjustment. However, these sensors have an inherent thermal response delay, causing the temperature information acquired by the system to lag behind the actual thermal state of the motor, resulting in "sensing lag." Reactive control strategies based on this often only adjust after parameter mismatch occurs, making it difficult to adapt to rapid temperature changes.

[0004] To address the lag issue, some studies have attempted to introduce prediction mechanisms. However, simple prediction methods lack accuracy under drastic temperature changes, while high-precision prediction models are computationally complex and cannot meet the stringent real-time requirements of embedded systems. Furthermore, existing control systems often operate in isolation, lacking the ability to collaborate with industrial internet platforms. They cannot comprehensively utilize multi-source information for local compensation, nor can they leverage cloud data for global optimization, thus limiting the overall robustness of the system in complex industrial environments.

[0005] Therefore, existing technologies generally suffer from problems such as temperature sensing lag, the contradiction between the real-time performance and accuracy of prediction models, and the dynamic matching of control parameters caused by system isolation. These have become the main technical bottlenecks restricting the high-reliability starting of brushless motors in a wide temperature environment. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for wide-temperature start-up of brushless motors in the industrial internet, which can effectively solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for wide-temperature start-up of a brushless motor in the industrial internet includes the following specific steps: Real-time synchronous acquisition of multi-modal operation data of brushless motors, and preprocessing of the acquired data to construct a multi-dimensional feature matrix containing time-series features of temperature, current and voltage; The multidimensional feature matrix is ​​input into the integer quantization recurrent neural network prediction model deployed in the edge computing control module, and forward inference is performed to output the local predicted temperature trajectory of the brushless motor within a future preset time window. Receive a global thermal environment early warning command issued by the cloud server, and perform fusion correction on the local predicted temperature trajectory based on the early warning command to generate the final temperature trajectory; Based on the final temperature trajectory, a preset motor physical parameterization model is invoked to perform calculations and dynamically generate an optimal start-up control parameter sequence that matches the current thermodynamic state. Based on the optimal start-up control parameter sequence, the brushless motor is controlled to execute a three-stage start-up process.

[0008] Furthermore, constructing a multidimensional feature matrix that includes time-series features of temperature, current, and voltage specifically includes: Simultaneously acquire analog signals of stator winding temperature, bearing temperature, three-phase current and DC bus voltage by using a multi-modal sensor array embedded in the motor body. Using an analog-to-digital converter driven by a unified sampling clock, synchronous sampling is performed on all the analog signal channels, and the converted digital data is written to a buffer through a direct memory access controller. A data window is extracted from the buffer at a fixed period, the data within the window is zero-point calibrated, and converted into a quantity value with physical units. The converted physical quantities are normalized, and the three-phase currents are subjected to coordinate transformation to calculate the characteristics of the effective current value. The normalized winding average temperature, bearing temperature, effective current value, bus voltage, and a reserved dimension of data are arranged in chronological order to construct the multidimensional feature matrix.

[0009] Furthermore, the integer-quantized recurrent neural network prediction model is an eight-bit integer-quantized long short-term memory network model, and the forward inference specifically includes: The multidimensional feature matrix is ​​flattened into a one-dimensional vector and then input into the model. Reliable start-up industrial brushless DC motor.

[0010] Furthermore, the cloud server uses a stream processing engine to perform macroscopic anomaly detection on the temperature data of a group of motors in a specific area. When a coordinated temperature change that conforms to statistical laws is detected, the global thermal environment early warning command is generated and issued.

[0011] On the other hand, the industrial internet brushless motor wide-temperature start system disclosed in this application is specifically as follows: A multimodal sensor array module is used to embed inside a brushless motor to simultaneously acquire multiple analog signals of its thermal and electrical states. An edge computing control module, connected to the multimodal sensing array module, and comprising at least: The data preprocessing unit is used to process the multiple analog signals to construct a multi-dimensional time-series feature matrix; The local temperature prediction unit has an embedded integer quantization recurrent neural network model for predicting the local temperature trajectory based on the feature matrix. The prediction result fusion unit is used to fuse the local temperature trajectory with the early warning instructions issued by the cloud to generate the final temperature trajectory; The startup parameter generation unit is used to calculate the optimal startup control parameters based on the final temperature trajectory and the built-in motor physical parameterization model. A drive execution unit is used to generate a drive signal based on the optimal start control parameters to control the motor to start. An industrial internet communication module, integrated into the edge computing control module, is used to establish a secure and reliable two-way communication link with the cloud. A cloud-based intelligent analysis server, which communicates with multiple of the aforementioned industrial internet communication modules, and includes at least: The macroscopic anomaly detection unit is used to analyze group data and generate global thermal environment early warning commands; The global model training unit is used to iteratively optimize the global hot model using aggregated edge data; The model distribution unit is used to distribute the optimized model parameters to the edge side to update the local model.

[0012] In summary, this application includes at least one of the following beneficial technical effects: 1. By deploying a long short-term memory neural network model with multi-bit integer quantization at the edge and combining it with the highly synchronous data acquisition of a multimodal sensor array, high-precision prediction of the temperature change trajectory of key motor components within a preset time period is achieved, with prediction error less than the preset temperature threshold. At the same time, by integrating global thermal environment early warning commands issued from the cloud, the impact of local sensing blind spots and sudden environmental disturbances is effectively compensated, and the sensing lag problem caused by physical delay of traditional temperature sensors is completely eliminated.

[0013] 2. Based on the predicted zero-hysteresis temperature trajectory, the optimal start-up control parameter sequence that strictly matches the current thermodynamic state is analyzed in real time through the built-in mathematical model of motor physical parameters, including key parameters such as current ramp-up slope and prepositioning current amplitude. This mechanism avoids stalling, overcurrent or torque oscillation caused by parameter mismatch in traditional reactive control, and significantly improves the start-up success rate in a wide temperature range.

[0014] 3. This application constructs a complete data closed loop from the edge to the cloud: the edge performs high real-time prediction and control, and the cloud continuously optimizes the global hot model based on massive device data and realizes online updates of the edge model through the model distribution unit; this architecture not only ensures the millisecond-level response performance of single-machine control, but also continuously improves the environmental adaptability and long-term reliability of the system through swarm intelligence. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of an industrial internet method for wide-temperature start-up of brushless motors. Figure 2 This is a schematic diagram illustrating the principle of fusing a multi-digit integer quantized long short-term memory neural network with a cloud-based thermal environment early warning system. Figure 3 This is a flowchart of the functional units within the edge computing control module. Detailed Implementation

[0016] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0017] This embodiment applies to the joint drive system of an industrial robot deployed in an unmanned production line in a high-altitude, cold region. This system needs to achieve highly reliable starting of the brushless DC motor within a wide temperature range of -40℃ to 100℃. Addressing the complex thermo-electric-magnetic coupling effects caused by extreme low temperatures, such as significantly increased winding resistance, enhanced magnetic flux density in permanent magnets, and a dramatic increase in the viscous torque of bearing grease, this embodiment constructs a wide-temperature starting system integrating multimodal perception, edge intelligent prediction, cloud-based collaborative correction, and physical model-driven parameter generation. A complete dynamic starting control process is executed on this system architecture.

[0018] First, from the perspective of system architecture construction, the industrial internet brushless motor wide-temperature start system deployed in this embodiment includes: A multimodal sensor array module is used to embed inside a brushless motor to simultaneously acquire multiple analog signals of its thermal and electrical states. An edge computing control module, connected to the multimodal sensing array module, and comprising at least: The data preprocessing unit is used to process the multiple analog signals to construct a multi-dimensional time-series feature matrix; The local temperature prediction unit has an embedded integer quantization recurrent neural network model for predicting the local temperature trajectory based on the feature matrix. The prediction result fusion unit is used to fuse the local temperature trajectory with the early warning instructions issued by the cloud to generate the final temperature trajectory; The startup parameter generation unit is used to calculate the optimal startup control parameters based on the final temperature trajectory and the built-in motor physical parameterization model. A drive execution unit is used to generate a drive signal based on the optimal start control parameters to control the motor to start. An industrial internet communication module, integrated into the edge computing control module, is used to establish a secure and reliable two-way communication link with the cloud. A cloud-based intelligent analysis server, which communicates with multiple of the aforementioned industrial internet communication modules, and includes at least: The macroscopic anomaly detection unit is used to analyze group data and generate global thermal environment early warning commands; The global model training unit is used to iteratively optimize the global hot model using aggregated edge data; The model distribution unit is used to distribute the optimized model parameters to the edge side to update the local model.

[0019] The multimodal sensor array module is directly embedded inside the motor body, and its hardware composition includes the following parts: For temperature sensing, three negative temperature coefficient thermistors are used, which are embedded in the end coil slots of the three phases of the stator winding through a potting process to detect the temperature of local hot spots in the winding with high spatial resolution. At the same time, a surface-mount platinum resistance thermometer is fastened to the outer wall of the bearing housing at the front end of the motor to monitor the thermal state of the bearing area.

[0020] In terms of current detection, a set of three-phase current sensors based on the Hall effect is configured and connected in series in the power circuit from the inverter bridge output to the three-phase windings of the motor to realize the synchronous acquisition of three-phase instantaneous current.

[0021] For voltage monitoring, a resistor voltage divider network is connected in parallel between the positive and negative poles of the DC bus, and the voltage divider point is connected to the analog input channel of the analog-to-digital converter for real-time monitoring of bus voltage fluctuations.

[0022] The input layer has 500 nodes, corresponding to the vector length after flattening the 5×100 feature matrix; the first hidden layer contains 128 long short-term memory units; the second hidden layer also contains 128 long short-term memory units; the output layer consists of 10 fully connected nodes, each corresponding to the predicted winding temperature at a series of future time points.

[0023] The model weight parameters W and bias b are uniformly mapped to 8-bit signed integers using the following linear quantization formula:

[0024] Where Q is the quantized integer value, and W is the original weight or bias to be quantized. and These are the maximum and minimum values ​​in the weight or bias matrix, respectively, and `round` represents the rounding function. The quantized parameter ranges from -128 to 127.

[0025] For the tanh and sigmoid activation functions in the long short-term memory (LSM) unit, a lookup table method is used to implement their fixed-point approximation calculation to adapt to the computational constraints of embedded platforms. After training and quantization, the model occupies approximately tens of KB of space in flash memory, and the time for a single forward inference is in the order of milliseconds. A dedicated buffer of fixed size is allocated in static random access memory, thus meeting the overall resource constraints of embedded devices.

[0026] In summary, the edge computing control module clearly defines its hardware configuration, functional architecture, and implementation details of the core prediction model, covering the entire process from data preprocessing and feature construction to quantization model deployment. This design ensures high-precision, low-latency temperature trajectory prediction in real time at the edge, providing a crucial basis for the subsequent dynamic generation of optimal startup parameters and serving as the core technical support for achieving reliable startup across a wide temperature range.

[0027] The industrial internet communication module is integrated within the edge computing control module. The hardware foundation of this module consists of an Ethernet media access controller built into the microcontroller and an external physical layer chip, thereby enabling wired network connectivity and supporting adaptive communication rates of 10 Mbps and 100 Mbps.

[0028] At the software protocol level, the module runs a lightweight message queue telemetry transport protocol stack, which is built on Transport Layer Security (TLS) version 1.2 and is used to establish an encrypted bidirectional communication link with the cloud server.

[0029] In summary, the industrial internet communication module defines the physical connection method and communication protocol standard for data interaction and command transmission between the edge and the cloud, ensuring that the edge control unit can reliably and securely receive cloud-based early warning commands and upload them for operation. Logs are a key communication foundation for achieving edge-cloud collaboration and closed-loop optimization.

[0030] The cloud-based intelligent analysis server is deployed on an industrial private cloud platform and is built using a distributed microservice architecture.

[0031] Its hardware foundation consists of a multi-node server cluster, which is managed and scheduled in a unified manner through a container orchestration platform and has elastic computing and storage resources.

[0032] At the software level, a server system comprises multiple functional units that work together.

[0033] The second unit is the data aggregation and storage unit. This unit is based on a distributed message queue system and is responsible for receiving startup data logs uploaded from thousands of edge devices and storing them persistently to provide a data foundation for subsequent analysis.

[0034] The second is the global thermal model training unit, which is built on a scalable machine learning platform framework. It uses multimodal time-series data from stored historical logs to continuously train a Transformer architecture global thermal model that incorporates an attention mechanism, in order to learn a wider range of motor thermal dynamics.

[0035] The third unit is the macroscopic anomaly detection unit, which is implemented based on a stream processing engine. It uses a sliding time window to statistically analyze the variance of temperature changes across all motor groups within a specific region and sets a threshold based on mathematical statistics principles. When the observed variance exceeds this threshold, a regional thermal disturbance is determined to exist. This threshold is set based on the principle of three standard deviations, i.e., 3σ, where σ represents the standard deviation of the temperature change sequence within the statistical window.

[0036] The fourth unit is the model distribution and update unit, which implements its communication mechanism based on a high-performance remote procedure call protocol. This unit is responsible for performing integer quantization on the trained and optimized new global model parameters and pushing them securely and accurately to the designated edge device nodes to complete the model update.

[0037] In summary, the cloud server, as the "intelligent brain" of the system, plays a crucial role in data aggregation, model iteration, group anomaly detection, and knowledge distribution. Close collaboration with the edge computing platform forms a complete closed loop from local perception to global optimization, and then from global knowledge to local applications, which is the core guarantee for the system's continuous evolution and environmental adaptability.

[0038] After the above system architecture is built, the workflow of this embodiment is carried out according to the following steps: Step S1 describes the initialization and data acquisition process of the multimodal sensor array module after the system is powered on. The specific execution steps are as follows: S101: Signal Generation and Acquisition; After the system is powered on, each sensor in the sensor array immediately starts working. Three negative temperature coefficient thermistors embedded in the stator windings continuously output analog voltage signals corresponding to the temperature at a frequency of 10 kHz. These signals are processed by a conditioning circuit including low-pass filtering and level shifting, and then sent to the first three analog input channels of the analog-to-digital converter. At the same time, the platinum resistance thermometer used to monitor the bearing temperature converts its resistance value into a voltage signal through a constant current source circuit and connects to the fourth channel of the analog-to-digital converter.

[0039] S102: Current and voltage signals are synchronously input; the analog signal reflecting the instantaneous current output by the three-phase Hall current sensor is processed by the isolation amplifier and then input to the fifth, sixth, and seventh channels of the analog-to-digital converter respectively. The DC bus voltage is sampled through a resistor divider network, and its divider point signal is input to the eighth channel of the analog-to-digital converter.

[0040] Steps S101 and S102 above complete the generation of analog signals and channel allocation for all temperature, current and voltage physical quantities, preparing for the next step of synchronous digitization.

[0041] S103: Synchronous Sampling and Data Storage; Driven by a 10kHz synchronous clock provided by the main control chip, the analog-to-digital converter performs strict simultaneous sampling of the eight analog input channels. After each sampling, the resulting 24-bit digital conversion result is directly written to a pre-allocated circular buffer in static random access memory via the direct memory access controller in block transfer mode. This process does not consume CPU resources.

[0042] In summary, step S1 clarifies the complete path from synchronous acquisition of multiple analog signals to efficient caching of digital results during the system initialization phase. Among these, a unified synchronous clock and direct memory access mechanism are crucial for ensuring high-precision time alignment and real-time processing of multi-source data, providing accurate and synchronized raw data for the subsequent construction of the time-series feature matrix.

[0043] Regarding step S2, it describes the processing flow of the data preprocessing unit on the raw sampled data, which is executed periodically according to the following sub-steps: S201: Periodic data extraction and zero-point calibration; The data preprocessing unit extracts the most recent 100 sampling points from the circular buffer of the static random access memory at a period of 10 milliseconds, forming a data window with a length of 1 second. For each data channel, the zero-point drift compensation of the extracted raw digital code is first performed using the calibration register value built into the analog-to-digital converter.

[0044] S202: Physical Quantity Conversion; After zero-point calibration, the calibrated digital code is converted into a value with actual physical meaning based on the unique physical conversion relationship of each sensor. Specifically, for temperature sensors, the conversion is performed according to their calibrated temperature-resistance characteristic curve; for current and voltage sensors, the conversion is performed according to their calibrated transmission coefficient. After conversion, the corresponding physical quantity for each channel is obtained, with units of Celsius, amperes, or volts, respectively.

[0045] The above steps S201 and S202 together complete the process of restoring the original digital signal to the standard physical quantity, eliminating the inherent zero drift error of the hardware, and providing accurate input for subsequent unified mathematical processing.

[0046] S203: Range Normalization; To ensure consistent weights for different physical quantities in subsequent models, all transformed physical quantities are normalized. A minimum-maximum normalization method is used to map each physical quantity to the interval between 0 and 1. The normalization formula is: in, Represents the original physical quantity value of a certain channel. This represents its normalized value. and These are the theoretical minimum and maximum values ​​of the physical quantity under the allowable operating conditions of the motor design. For example, the winding temperature... and The temperature can be set to -40℃ and 150℃ respectively.

[0047] S204: Current characteristic calculation and matrix construction; For the normalized values ​​of the three-phase currents, the Clarke transformation is used to further transform them from the three-phase stationary coordinate system to the two-phase stationary coordinate system, resulting in... and Components. Then, the characteristic value of the current RMS at that moment is calculated. The calculation formula is: Finally, the normalized winding average temperature, bearing temperature, calculated RMS current, bus voltage, and data from a reserved dimension are arranged in chronological order to construct a 5-row, 100-column multidimensional time-series feature matrix. In the matrix, the first... Line number Column elements Indicates the first The sampling time point, the first The normalized values ​​of each feature.

[0048] In summary, step S2's data preprocessing workflow systematically transforms the raw cached data into a normalized feature matrix. Through periodic extraction, error calibration, physical quantity restoration, unified normalization, and specific feature calculations, a standardized data representation that combines time series characteristics with multi-dimensionality is generated. This feature matrix... It is a key input that triggers the subsequent local temperature prediction model, and its quality directly determines the accuracy of the prediction trajectory. It is the core data processing link that connects high-speed acquisition and intelligent prediction.

[0049] For step S3, the specific process of the local temperature prediction unit using a quantization model to predict the temperature trajectory is executed according to the following sub-steps: All the aforementioned analog signal channels are connected to the same analog-to-digital converter (ADC) chip with multi-channel synchronous sampling capability. The sampling frequency of this chip is configurable; in this embodiment, it is set to a fixed frequency to meet the sampling requirements of the thermodynamic process. The digital output interface of the ADC is connected to the direct memory access controller (DRAM) of the motor controller's main control chip via a serial bus, enabling high-speed data transfer without consuming central processing unit resources, ensuring that the time synchronization error of the multi-source signals is controlled within the design range.

[0050] The multimodal sensor array module is a hardware implementation scheme for constructing a multimodal sensor array, covering high-precision synchronous acquisition of three key physical quantities: temperature, current, and voltage. It provides a reliable and synchronous data source for subsequent time-series feature extraction and temperature prediction at the edge, serving as the fundamental hardware guarantee for achieving zero-hysteresis prediction and look-ahead control.

[0051] The edge computing control module uses an industrial-grade microcontroller based on a high-performance core. It has sufficient clock speed, flash memory and static random access memory resources, and integrates hardware floating-point units and digital signal processing instruction sets to support complex calculations.

[0052] This module is logically divided into five functional units, including a data preprocessing unit, a local temperature prediction unit, a prediction result fusion unit, a startup parameter generation unit, and a drive execution unit.

[0053] The data preprocessing unit works in concert with the microcontroller’s general-purpose timer and direct memory access channel to process the raw sampled data stream uploaded by the multimodal sensor array. The specific process includes: timestamp alignment using the synchronous sampling trigger signal of the analog-to-digital converter as a unified time reference, zero-point calibration of each channel, normalization of each physical quantity to the interval [0,1] according to the design range, and sliding window truncation according to the preset length.

[0054] After the above processing, the system constructs a multi-dimensional time-series feature matrix X with 5 rows and 100 columns. Each row of the matrix represents a feature sequence: the first row is the average temperature of the three-phase windings, obtained by the arithmetic mean of the readings of three negative temperature coefficient thermistors; the second row is the bearing temperature; the third row is the effective value of the three-phase current calculated after Clarke transformation; the fourth row is the bus voltage; and the fifth row is a reserved dimension for subsequent integration with early warning instructions issued from the cloud.

[0055] The core of the local temperature prediction unit is a long short-term memory (LSTM) neural network model quantized to 8 bits, which is embedded in the microcontroller's flash memory. The network topology consists of an input layer, two hidden layers, and an output layer. The input layer has 500 nodes, corresponding to the vector length after flattening the 5×100 feature matrix; the first hidden layer contains 128 LSM units; the second hidden layer also contains 128 LSM units; the output layer consists of 10 fully connected nodes, each corresponding to a predicted winding temperature at a series of future time points.

[0056] The model weight parameters W and bias b are uniformly mapped to 8-bit signed integers using the following linear quantization formula: Q="round" ((W-W_"min" ) / (W_"max" -W_"min" )×255)-128 Where Q is the quantized integer value, W is the original weight or bias to be quantized, W_"max" and W_"min" are the maximum and minimum values ​​in the weight or bias matrix, respectively, and round represents the rounding function. The range of the quantized parameters is [-128, 127].

[0057] For the tanh and sigmoid activation functions in the long short-term memory (LSM) unit, a lookup table method is used to implement their fixed-point approximation calculation to adapt to the computational constraints of embedded platforms. After training and quantization, the model occupies approximately tens of KB of space in flash memory, and the time for a single forward inference is in the order of milliseconds. A dedicated buffer of fixed size is allocated in static random access memory, thus meeting the overall resource constraints of embedded devices.

[0058] In summary, the edge computing control module clearly defines its hardware configuration, functional architecture, and implementation details of the core prediction model, covering the entire process from data preprocessing and feature construction to quantization model deployment. This design ensures high-precision, low-latency temperature trajectory prediction in real time at the edge, providing a crucial basis for the subsequent dynamic generation of optimal startup parameters and serving as the core technical support for achieving reliable startup across a wide temperature range.

[0059] The industrial internet communication module is integrated within the edge computing control module. The hardware foundation of this module consists of an Ethernet media access controller built into the microcontroller and an external physical layer chip, thereby enabling wired network connectivity and supporting adaptive communication rates of 10 Mbps and 100 Mbps.

[0060] At the software protocol level, the module runs a lightweight message queue telemetry transport protocol stack, which is built on Transport Layer Security (TLS) version 1.2 and is used to establish an encrypted bidirectional communication link with the cloud server.

[0061] In summary, the industrial internet communication module determines the physical connection method and communication protocol standard for data interaction and command transmission between the edge and the cloud, ensuring that the edge control unit can reliably and securely receive cloud warning commands and upload operation logs. It is a key communication foundation for realizing edge-cloud collaboration and closed-loop optimization.

[0062] The cloud-based intelligent analysis server is deployed on an industrial private cloud platform and is built using a distributed microservice architecture.

[0063] Its hardware foundation consists of a multi-node server cluster, which is managed and scheduled in a unified manner through a container orchestration platform and has elastic computing and storage resources.

[0064] At the software level, a server system comprises multiple functional units that work together.

[0065] The second unit is the data aggregation and storage unit. This unit is based on a distributed message queue system and is responsible for receiving startup data logs uploaded from thousands of edge devices and storing them persistently to provide a data foundation for subsequent analysis.

[0066] The second is the global thermal model training unit, which is built on a scalable machine learning platform framework. It uses multimodal time-series data from stored historical logs to continuously train a Transformer architecture global thermal model that incorporates an attention mechanism, in order to learn a wider range of motor thermal dynamics.

[0067] The third unit is the macroscopic anomaly detection unit, which is implemented based on a stream processing engine. It uses a sliding time window to statistically analyze the variance of temperature changes across all motor groups within a specific region and sets a threshold based on mathematical statistics principles. When the observed variance exceeds this threshold, a regional thermal disturbance is determined to exist. This threshold is set based on the principle of three standard deviations, i.e., 3σ, where σ represents the standard deviation of the temperature change sequence within the statistical window.

[0068] The fourth unit is the model distribution and update unit, which implements its communication mechanism based on a high-performance remote procedure call protocol. This unit is responsible for performing integer quantization on the trained and optimized new global model parameters and pushing them securely and accurately to the designated edge device nodes to complete the model update.

[0069] In summary, the cloud server, as the "intelligent brain" of the system, plays a crucial role in data aggregation, model iteration, group anomaly detection, and knowledge distribution. Close collaboration with the edge computing platform forms a complete closed loop from local perception to global optimization, and then from global knowledge to local applications, which is the core guarantee for the system's continuous evolution and environmental adaptability.

[0070] After the above system architecture is built, the workflow of this embodiment is carried out according to the following steps: Step S1 describes the initialization and data acquisition process of the multimodal sensor array module after the system is powered on. The specific execution steps are as follows: S101: Signal Generation and Acquisition; After the system is powered on, each sensor in the sensor array immediately starts working. Three negative temperature coefficient thermistors embedded in the stator windings continuously output analog voltage signals corresponding to the temperature at a frequency of 10 kHz. These signals are processed by a conditioning circuit including low-pass filtering and level shifting, and then sent to the first three analog input channels of the analog-to-digital converter. At the same time, the platinum resistance thermometer used to monitor the bearing temperature converts its resistance value into a voltage signal through a constant current source circuit and connects to the fourth channel of the analog-to-digital converter.

[0071] S102: Current and voltage signals are synchronously input; the analog signal reflecting the instantaneous current output by the three-phase Hall current sensor is processed by the isolation amplifier and then input to the fifth, sixth, and seventh channels of the analog-to-digital converter respectively. The DC bus voltage is sampled through a resistor divider network, and its divider point signal is input to the eighth channel of the analog-to-digital converter.

[0072] Steps S101 and S102 above complete the generation of analog signals and channel allocation for all temperature, current and voltage physical quantities, preparing for the next step of synchronous digitization.

[0073] S103: Synchronous Sampling and Data Storage; Driven by a 10kHz synchronous clock provided by the main control chip, the analog-to-digital converter performs strict simultaneous sampling of the eight analog input channels. After each sampling, the resulting 24-bit digital conversion result is directly written to a pre-allocated circular buffer in static random access memory via the direct memory access controller in block transfer mode. This process does not consume CPU resources.

[0074] In summary, step S1 clarifies the complete path from synchronous acquisition of multiple analog signals to efficient caching of digital results during the system initialization phase. Among these, a unified synchronous clock and direct memory access mechanism are crucial for ensuring high-precision time alignment and real-time processing of multi-source data, providing accurate and synchronized raw data for the subsequent construction of the time-series feature matrix.

[0075] Regarding step S2, it describes the processing flow of the data preprocessing unit on the raw sampled data, which is executed periodically according to the following sub-steps: S201: Periodic data extraction and zero-point calibration; The data preprocessing unit extracts the most recent 100 sampling points from the circular buffer of the static random access memory at a period of 10 milliseconds, forming a data window with a length of 1 second. For each data channel, the zero-point drift compensation of the extracted raw digital code is first performed using the calibration register value built into the analog-to-digital converter.

[0076] S202: Physical Quantity Conversion; After zero-point calibration, the calibrated digital code is converted into a value with actual physical meaning based on the unique physical conversion relationship of each sensor. Specifically, for temperature sensors, the conversion is performed according to their calibrated temperature-resistance characteristic curve; for current and voltage sensors, the conversion is performed according to their calibrated transmission coefficient. After conversion, the corresponding physical quantity for each channel is obtained, with units of Celsius, amperes, or volts, respectively.

[0077] The above steps S201 and S202 together complete the process of restoring the original digital signal to the standard physical quantity, eliminating the inherent zero drift error of the hardware, and providing accurate input for subsequent unified mathematical processing.

[0078] S203: Range Normalization; To ensure consistent weights for different physical quantities in subsequent models, all transformed physical quantities are normalized. A minimum-maximum normalization method is used to map each physical quantity to the interval between 0 and 1. The normalization formula is: in, Represents the original physical quantity value of a certain channel. This represents its normalized value. and These are the theoretical minimum and maximum values ​​of the physical quantity under the allowable operating conditions of the motor design. For example, the winding temperature... and The temperature can be set to -40℃ and 150℃ respectively.

[0079] S204: Current characteristic calculation and matrix construction; For the normalized values ​​of the three-phase currents, the Clarke transformation is used to further transform them from the three-phase stationary coordinate system to the two-phase stationary coordinate system, resulting in... and Components. Then, the characteristic value of the current RMS at that moment is calculated. The calculation formula is: Finally, the normalized winding average temperature, bearing temperature, calculated RMS current, bus voltage, and data from a reserved dimension are arranged in chronological order to construct a 5-row, 100-column multidimensional time-series feature matrix. In the matrix, the first... Line number Column elements Indicates the first The sampling time point, the first The normalized values ​​of each feature.

[0080] In summary, step S2's data preprocessing workflow systematically transforms the raw cached data into a normalized feature matrix. Through periodic extraction, error calibration, physical quantity restoration, unified normalization, and specific feature calculations, a standardized data representation that combines time series characteristics with multi-dimensionality is generated. This feature matrix... It is a key input that triggers the subsequent local temperature prediction model, and its quality directly determines the accuracy of the prediction trajectory. It is the core data processing link that connects high-speed acquisition and intelligent prediction.

[0081] For step S3, the specific process of the local temperature prediction unit using a quantization model to predict the temperature trajectory is executed according to the following sub-steps: S301: Model Triggering and Data Preparation; When Multidimensional Temporal Feature Matrix Once the construction is complete, the local temperature prediction unit is triggered. This unit first generates a matrix with dimensions of 5 rows and 100 columns. The vector is flattened in memory to form an input vector containing 500 elements.

[0082] S302: Subsequently, this 500-dimensional vector is fed into an 8-bit integer-quantized long short-term memory neural network model stored in flash memory. The model performs forward inference computation according to its fixed topology.

[0083] The above steps S301 and S302 complete the triggering of the prediction task and the format adaptation of the input data, ensuring that the data can correctly flow into the deployed neural network model for calculation.

[0084] S303: Temporal feature extraction computation; both hidden layers of the model perform computation using Long Short-Term Memory (LSTM) units. For each time step, the internal computation of each unit specifically includes: First, the activation values ​​of the forget gate, input gate, and output gate are calculated. The calculation of these gates follows the standard Long Short-Term Memory (LSTM) formula, but all operations are implemented using fixed-point integer arithmetic.

[0085] Specifically, for the input vector, the hidden state of the previous time step, and the corresponding 8-bit integer weights and biases, the original weighted sum of the gate is obtained through fixed-point multiplication and addition operations.

[0086] Subsequently, these weighted sums, which are fixed-point approximations of the sigmoid or tanh activation functions implemented by a lookup table, are mapped to 8-bit unsigned integers between 0 and 255 to simulate the activation value range of 0 to 1 or -1 to 1.

[0087] During the calculation process, all intermediate accumulation results are stored in a 32-bit accumulator to prevent numerical overflow during continuous fixed-point multiplication and addition operations, thus ensuring calculation accuracy.

[0088] After these gating mechanisms perform calculations and information filtering, the unit updates its internal cell state and outputs the hidden state at the current moment.

[0089] S304: After timing processing through two layers of networks, the final hidden state is passed to the output layer. The output layer consists of 10 fully connected nodes, each performing an 8-bit fixed-point linear weighted summation on the input.

[0090] The raw output value produced by each node in the output layer is an 8-bit signed integer. To obtain a temperature prediction value with physical meaning in degrees Celsius, these integer outputs need to be dequantized.

[0091] The inverse quantization formula is: in, It is the first Temperature values ​​at each predicted time point (unit: °C). It is the first output of the model 8-bit signed integer values It is the preset quantization scaling factor. This is the preset zero-point offset. Scaling factor. With zero point The model is determined and fixed in the program after training.

[0092] Finally, a 10-dimensional prediction output vector is generated. Each element in the vector Corresponding to a future time interval Predicted winding temperature at milliseconds.

[0093] Steps S303 and S304 are the core of model calculation. Through resource-optimized fixed-point network operations, the input historical time series data is mapped into temperature prediction sequences for multiple future time points.

[0094] In summary, this prediction process is completed independently on the edge microcontroller, achieving fully local millisecond-level real-time inference. By inputting a normalized feature matrix into a specially quantized and optimized long short-term memory network, the system can obtain the temperature change trajectory of the motor windings in a forward-looking manner over the next few hundred milliseconds. This prediction result serves as the direct basis for subsequent control parameter optimization and fusion with cloud commands, forming the technological cornerstone for achieving "zero-hysteresis" temperature sensing and forward-looking control strategies.

[0095] Step S4 discloses the process of receiving early warning instructions from the cloud at the edge, which is specifically executed according to the following sub-steps: S401: Continuously monitors downlink messages from the cloud; the industrial internet communication module runs continuously in the background, monitoring downlink messages from the cloud-based intelligent analysis server in real time through the established encrypted communication link. This monitoring process is asynchronous and does not block local control tasks.

[0096] S402: Generates and issues global early warning commands; the macroscopic anomaly detection unit in the cloud server is responsible for statistical analysis of the temperature data of a large number of motors in a specific area. When this unit detects that hundreds of motors deployed in the same industrial park have experienced a synchronous temperature surge within the last 5-minute time window, it determines that there is a regional thermal disturbance.

[0097] For example, if a cooling system malfunction causes the ambient temperature in the entire area to rise by 20 degrees Celsius within 10 minutes, an alert will be triggered.

[0098] At this point, a global thermal environment early warning command will be generated in the cloud, denoted as G.

[0099] S403: Instruction content and transmission mechanism; The generated warning instruction G contains two core data fields: a boolean flag to indicate whether a valid warning exists; and a floating-point temperature change rate δT / δt, in degrees Celsius per second, to quantify the rate of change in the regional environment.

[0100] The instruction is published to a pre-agreed specific topic via a message queue telemetry transport protocol with a service quality level of 1 and guaranteed reliability.

[0101] S404: Edge reception and command buffering; the industrial internet communication module at the edge has pre-subscribed to the above-mentioned specific topics. When a command is issued for this topic, the edge module immediately receives and parses the command message.

[0102] After successful parsing, the two key data points, flag and temperature change rate δT / δt, from instruction G are stored in a pre-defined dedicated global variable in static random access memory for subsequent reading and use by the prediction result fusion unit.

[0103] In summary, step S4 establishes a low-latency, highly reliable information channel from cloud-based anomaly detection to edge command reception. By real-time monitoring and caching of macroscopic thermal environment early warnings issued from the cloud, the edge control system gains environmental perception capabilities beyond its own local sensors, providing crucial information input for subsequent correction of local predictions and enhancing the system's robustness under sudden environmental disturbances.

[0104] Step S5 is the process by which the prediction result fusion unit fuses local predictions and cloud-based early warnings. This process is executed according to the following sub-steps: S501: Data Reading and Status Judgment; The prediction result fusion unit first reads two key data from memory: one is the 10-dimensional prediction vector generated by the local temperature prediction unit. ,in Corresponding to the future The first is the predicted temperature at a given time; the second is the global early warning command read from dedicated variables. This instruction includes a Boolean flag. and floating-point temperature change rate The unit is degrees Celsius per second.

[0105] The unit first checks the flag bits in the instruction. The state.

[0106] S502: Direct output when there is no warning; if the flag is checked. A value of false indicates that no effective regional thermal disturbance was detected in the cloud. In this case, the prediction result fusion unit does not perform any correction calculations and directly converts the local prediction vector. As the final temperature trajectory output, i.e. .

[0107] S503: Weighted fusion calculation when there is an early warning; if the judgment flag is present. A value of true indicates a sudden change in the macroscopic environment. The unit will execute a weighted fusion algorithm to correct for each prediction time point.

[0108] For the Predicted time points Its corresponding future time value millisecond.

[0109] First, calculate the temperature shift caused by environmental disturbance at that time point. The calculation formula is: in, It is the rate of regional temperature change carried in the early warning instruction. It is a time-increasing weighting function used to simulate the cumulative amplification effect of environmental disturbances over time. This weighting function is defined as: In the formula, It is an empirical decay coefficient used to control the rate of weight growth; in this embodiment, its value is 0.02. It is a natural constant.

[0110] The core of step S503 is to calculate the expected temperature increment caused by environmental disturbances at each future time point based on the macroscopic environmental change rate provided by the cloud and a weight that grows non-linearly over time.

[0111] S504: Generate the corrected final predicted trajectory; obtain the temperature offset for each predicted point. Then, it is compared with the corresponding local predicted value. Add them together to get the corrected predicted temperature value for that point. : right Perform the above calculations on all points to generate a corrected 10-dimensional final predicted temperature trajectory vector. .

[0112] In summary, step S5 intelligently identifies the warning status and, upon confirming the existence of macroscopic thermal disturbances, dynamically overlays the global environmental change trend onto the local high-precision prediction in a time-weighted manner. This process combines point-like local sensing with area-like cloud-based early warning to generate a comprehensive temperature evolution trajectory that reflects both the motor's own thermal dynamics and responds to sudden environmental changes, providing a more reliable and robust temperature input benchmark for the accurate calculation of subsequent control parameters.

[0113] Regarding step S6, it describes the process by which the startup parameter generation unit calculates the optimal startup parameters based on the corrected temperature trajectory. This process is executed according to the following sub-steps: S601: Data input and preparation; the parameter generation unit is started to receive the corrected temperature trajectory vector from the prediction result fusion unit. This vector represents the predicted winding temperature at a series of future time points starting from the current moment.

[0114] Meanwhile, the unit obtains the predicted temperature value of the bearing area from the data preprocessing stage. .

[0115] S602: Calculation of equivalent winding resistance; the cell uses the internally built-in winding resistance temperature model. This model describes the linear relationship between winding resistance and temperature, as shown in the following formula: in, The temperature is The equivalent resistance of the winding at that time. The winding at the reference temperature The reference resistance value (usually 20 degrees Celsius) can be obtained from the motor nameplate parameters. It is the temperature coefficient of resistance of the winding material. For copper windings, its typical value is 0.00393 per degree Celsius. Take the temperature trajectory vector The first element in That is, the predicted temperature 50 milliseconds after startup.

[0116] Substituting the above parameters into the formula, the equivalent winding resistance at the moment of startup can be calculated. .

[0117] The above steps S601 and S602 complete the preparation of key input data and calculate the actual resistance value of the winding at startup based on the forward temperature prediction, which is the basis for subsequent current calculation.

[0118] S603: Calculation of bearing viscous friction torque; the element uses the internally fixed bearing viscous friction torque model. This model adopts an Arrhenius exponential relationship, as shown in the following formula: in, The temperature is The bearing viscous friction torque at that time. and These are the characteristic constants of the grease used in bearings, namely the pre-exponential factor and the activation energy. It is the universal gas constant, with a value of 8.314 joules per mole per Kelvin. Take the predicted temperature of the bearing area The unit is Celsius.

[0119] Substituting the parameters into the formula, the bearing viscous friction torque at the current temperature can be calculated. .

[0120] S604: Establish and solve the startup dynamic equations; based on the calculated winding resistance and bearing friction torque The unit establishes the second-order dynamic system equations describing the motor starting process: In this equation, It is the moment of inertia of the rotor. It is the damping coefficient of the system. These are the torque constants of the motor, and they are all inherent parameters of the motor. It is a function of the starting current as a function of time. Since the motor is stationary and unloaded before starting, the load torque... Set it to 0.

[0121] Set the starting current as a ramp function: ,in Let be the current ramp-up slope to be determined. For time.

[0122] To ensure in time to Within the preset start-up phase time, the integral torque generated by the current just overcomes the frictional torque. At the same time, the instantaneous current should not exceed the preset maximum allowable current. (For example, 15 amperes), the unit uses numerical integration and optimization algorithms to solve the above equations, and finally obtains the optimal current ramp-up slope. .

[0123] The above steps S603 and S604 calculate the bearing resistance torque based on the thermodynamic model, and use this as the core constraint to determine the optimal current variation law required to overcome the resistance by solving the motor motion equation.

[0124] S605: Calculate other key control parameters; the unit continues to calculate other parameters required for the startup sequence.

[0125] 1. Pre-positioning current amplitude Considering the remanence of permanent magnets at low temperatures To enhance the electromagnetic attraction and maintain a constant electromagnetic force during the pre-positioning phase, the current amplitude needs to be reduced proportionally. The calculation formula is as follows: in, This is the rated prepositioning current at room temperature. It is the magnetic temperature coefficient of the permanent magnet (typically 0.0012 per degree Celsius). Take the predicted low temperature value (e.g.) (Minimum value in the range).

[0126] 2. Initial commutation frequency This frequency is determined by the back electromotive force constant at the predicted temperature. By working backward, it can be concluded that commutation and back EMF must be synchronized during the open-loop startup phase.

[0127] 3. Closed-loop switching speed threshold The threshold is dynamically adjusted based on the predicted temperature rise rate. If the predicted temperature rise is too fast, the switching speed is increased to avoid loss of control synchronization due to rapid parameter changes causing the system to enter the closed loop in the low-speed region.

[0128] In summary, step S6 is the crucial link between temperature prediction and final drive execution. Through a series of physically parameterized models, the abstract temperature trajectory is transformed into a specific, executable set of optimal control parameters, including current slope, amplitude, frequency, and switching threshold. This mechanism ensures a precise match between the startup energy injection and the motor's current thermodynamic state, fundamentally avoiding stalling, overcurrent, or torque oscillation problems caused by parameter mismatch, and providing direct parameter basis for achieving high-reliability startup over a wide temperature range.

[0129] Step S7 describes the process by which the drive execution unit performs startup control based on the optimal startup parameters. This process is executed in the following sub-steps: S701: Parameter reception and hardware configuration; The drive execution unit receives the optimal startup control parameter packet from the startup parameter generation unit, which includes the current ramp-up slope. Pre-positioning current amplitude Initial commutation frequency and closed-loop switching speed threshold .

[0130] Based on these parameters, the unit configures the relevant registers of the pulse width modulation generator. Specifically, this includes setting the value of the comparator register to define the pulse width, thereby controlling the output voltage amplitude; and configuring the dead time to prevent shoot-through short circuits of the power transistors on the same bridge arm. After configuration, the pulse width modulation generator generates six complementary pulse width modulation signals with dead time.

[0131] S702: Execute the first stage of the startup sequence – rotor pre-positioning; The drive unit outputs six pulse width modulation signals to the three-phase full-bridge inverter circuit to drive its power switching transistors.

[0132] The startup sequence first enters the rotor pre-positioning stage, which lasts for 200 milliseconds. During this stage, the inverter circuit applies an amplitude of [missing value] to the motor windings. The DC current locks the motor rotor at a known initial angular position, preparing it for subsequent commutation acceleration.

[0133] The above steps S701 and S702 complete the hardware parameterization configuration before execution and implement the pre-positioning operation based on predicted temperature adjustment to ensure that the rotor is in a controllable initial state.

[0134] S703: Execute the second stage of the startup sequence – open-loop acceleration; After the pre-positioning phase, the system seamlessly enters the open-loop acceleration phase. During this phase, the phase current amplitude applied to the motor follows the optimal slope. The electromagnetic torque increases linearly over time, resulting in a gradual increase in electromagnetic torque. Simultaneously, the commutation logic of the inverter circuit is based on the initial frequency. It starts working and gradually increases the commutation frequency according to the preset pattern, driving the motor speed to rise from zero.

[0135] S704: Execute the third stage of the startup sequence – switch to closed-loop control; During open-loop acceleration, the drive execution unit reads the actual motor speed fed back by the encoder in real time. When the actual rotational speed is detected Exceeding the preset closed-loop switching threshold At this point, the system immediately and seamlessly switches from open-loop commutation control mode to field-oriented closed-loop control mode. Afterward, motor operation is taken over by the closed-loop controller, achieving high-precision speed regulation and torque control.

[0136] For the above steps S703 and S704, the transition from accelerated start-up to stable operation is dynamically executed, wherein the current ramp-up and commutation frequency strictly follow the optimal parameters generated based on temperature prediction, and the switching timing is also determined by dynamic threshold.

[0137] In summary, step S7 transforms the theoretically optimal parameters calculated in the preceding steps into a crucial step in the actual physical operation. By precisely configuring pulse width modulation and strictly executing the three-stage startup sequence, the system achieves "just right" energy injection guided by the predicted temperature trajectory. This process ensures that the startup current can be precisely controlled regardless of whether the resistance increases due to extreme low temperatures or the winding resistance decreases due to high temperatures, thereby completely eliminating the startup failure or surge current problems common in traditional fixed-parameter strategies, completing the final closed loop from intelligent prediction to reliable execution.

[0138] Step S8 describes the closed-loop process of data upload and cloud model iteration after startup. This process is executed in the following sub-steps: S801: Event Context Data Packaging; After the startup process is completed, regardless of success or failure, the edge computing control module immediately starts the data packaging program. This program integrates the complete context information of this startup event and organizes it into a JSON-formatted data log according to a predefined key-value pair structure.

[0139] The log contains the following specific contents: 1. Original multimodal state data: A numerical matrix containing 100 consecutive time steps, each time step covering 5 feature dimensions.

[0140] 2. Predicted data: The original predicted temperature trajectory vector generated by the local temperature prediction unit. And the final predicted temperature trajectory vector after fusion correction. .

[0141] 3. Control Parameters: The optimal set of startup control parameters calculated by the startup parameter generation unit, including... , , , wait.

[0142] 4. Result Identifier: A digital status identifier used to indicate the startup result, such as 0 for success, 1 for stall, and 2 for overcurrent.

[0143] S802: Encrypted upload to the cloud; after the data logs are packaged, they are transmitted via the industrial internet communication module. Before transmission, the communication module encrypts the data payload using a transport layer security protocol. Subsequently, the encrypted log data is uploaded to the cloud-based intelligent analysis server via a message queue telemetry transmission protocol at a specified quality of service level. The cloud-based data aggregation and storage unit is responsible for receiving and persistently storing these logs from a large number of edge devices.

[0144] S803: Cloud-based periodic model training; The global thermal model training unit on the cloud server automatically executes training tasks at fixed time intervals (e.g., every 24 hours). This unit extracts a certain number (e.g., 100,000) of the latest accumulated startup data logs from the data storage unit as the training dataset. During training, a transfer learning strategy is used to fine-tune the existing global thermal model. Specifically, while keeping the main architecture of the model unchanged, new log data, especially its multimodal time-series features and corresponding actual temperature evolution results, is used to adjust some hierarchical parameters of the model through the backpropagation algorithm, enabling the model to learn newer and broader motor thermal dynamics.

[0145] S804: Quantization Model Distribution and Edge Update; After the global thermal model completes one round of training and optimization, the model distribution unit is responsible for deploying it to the edge. First, the updated model weight parameters are quantized into eight-bit integers, converting them into a format suitable for edge microcontroller operation. Then, through a high-performance remote procedure call protocol, the quantized new model parameters are securely and reliably pushed to all online edge device nodes. After receiving the new parameters, the local temperature prediction unit at the edge completes the update and replacement of model parameters before the next prediction startup.

[0146] Steps S803 and S804 implement knowledge extraction and distribution in the cloud. The cloud continuously optimizes the general model using group data and synchronizes the evolved "knowledge" to every edge node through an efficient distribution mechanism.

[0147] In summary, step S8 constructs a complete data loop of "perception-decision-execution-learning." Each startup attempt, regardless of success or failure, sends data back to the cloud, fueling the continuous evolution of the entire system. Through periodic model training and distribution, the edge prediction model deployed on millions of devices continuously absorbs collective experience, thereby constantly improving its prediction accuracy and generalization ability in complex and ever-changing environments. This self-evolving mechanism of edge-cloud collaboration is the core reason why the system can maintain high reliability and environmental adaptability over the long term.

[0148] Through the aforementioned system construction and process execution, this embodiment significantly improves the startup success rate and reduces the peak startup current in cold start tests. At a high temperature of 100℃, the startup time is shortened, and the winding temperature rise rate is effectively suppressed.

[0149] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0150] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for wide-temperature start-up of a brushless motor in the industrial internet, characterized in that, Includes the following steps: Real-time synchronous acquisition of multi-modal operation data of brushless motors, and preprocessing of the acquired data to construct a multi-dimensional feature matrix containing time-series features of temperature, current and voltage; The multidimensional feature matrix is ​​input into the integer quantization recurrent neural network prediction model deployed in the edge computing control module, and forward inference is performed to output the local predicted temperature trajectory of the brushless motor within a future preset time window. Receive a global thermal environment early warning command issued by the cloud server, and perform fusion correction on the local predicted temperature trajectory based on the early warning command to generate the final temperature trajectory; Based on the final temperature trajectory, a preset motor physical parameterization model is invoked to perform calculations and dynamically generate an optimal start-up control parameter sequence that matches the current thermodynamic state. Based on the optimal start-up control parameter sequence, the brushless motor is controlled to execute a three-stage start-up process.

2. The method for wide-temperature start-up of a brushless motor in the industrial internet according to claim 1, characterized in that, Constructing a multidimensional feature matrix that includes time-series features of temperature, current, and voltage specifically includes: Simultaneously acquire analog signals of stator winding temperature, bearing temperature, three-phase current and DC bus voltage by using a multi-modal sensor array embedded in the motor body. Using an analog-to-digital converter driven by a unified sampling clock, synchronous sampling is performed on all the analog signal channels, and the converted digital data is written to the buffer through a direct memory access controller. A data window is extracted from the buffer at a fixed period, the data within the window is zero-point calibrated, and converted into a quantity value with physical units. The converted physical quantities are normalized, and the three-phase currents are subjected to coordinate transformation to calculate the characteristics of the effective current value. The normalized winding average temperature, bearing temperature, effective current value, bus voltage, and a reserved dimension of data are arranged in chronological order to construct the multidimensional feature matrix.

3. The method for wide-temperature start of a brushless motor in the industrial internet according to claim 1 or 2, characterized in that, The integer-quantized recurrent neural network prediction model is an eight-bit integer-quantized long short-term memory network model, and the forward inference specifically includes: The multidimensional feature matrix is ​​flattened into a one-dimensional vector and then input into the model. The model internally uses fixed-point integer arithmetic and an activation function based on a lookup table to extract and calculate time-series features. The model's output layer produces predictions, represented by eight-bit integer values, corresponding to multiple equally spaced future time points. The eight-bit integer value is dequantized to obtain the local predicted temperature trajectory expressed in degrees Celsius.

4. The method for wide-temperature start-up of a brushless motor in the industrial internet according to claim 1, characterized in that, Generating the final temperature trajectory specifically includes: Through the industrial internet communication link, continuously monitor and receive global thermal environment early warning instructions issued by the cloud server. The instructions include at least an early warning status flag and a regional temperature change rate. When the warning status flag indicates that there is a valid warning, for each future prediction time point in the local predicted temperature trajectory, the temperature compensation caused by macroscopic thermal environment disturbance is calculated based on the regional temperature change rate and a weighting function that grows non-linearly with time. The temperature compensation amount corresponding to each predicted time point is superimposed on the corresponding predicted value of the local predicted temperature trajectory to generate the final temperature trajectory.

5. The method for wide-temperature start of a brushless motor in the industrial internet according to claim 1, characterized in that, The dynamic generation of the optimal startup control parameter sequence specifically includes: Based on the initial predicted temperature value in the final temperature trajectory, the equivalent winding resistance at the initial startup moment is calculated using the winding resistance-temperature linear model. Based on the predicted temperature value of the bearing region, the bearing friction torque at the current temperature is calculated using the bearing viscous friction torque model based on the Arrhenius exponent relationship. Using the equivalent resistance of the winding and the frictional torque of the bearing as key constraints, a dynamic equation for the motor starting process is established, and the optimal starting current ramp-up slope is obtained through numerical solution. Based on the temperature information in the final temperature trajectory, combined with the temperature coefficient of the permanent magnet of the motor, the current amplitude in the prepositioning stage is calculated, and combined with the temperature characteristics of the back electromotive force, the initial commutation frequency and the speed threshold for switching from open-loop start-up to closed-loop control are determined.

6. The method for wide-temperature start-up of a brushless motor in the industrial internet according to claim 1, characterized in that, The three-stage startup process specifically includes: In the first stage, the parameters of the pulse width modulation generator are configured, and a DC current corresponding to the amplitude of the prepositioning current is applied to the motor winding to perform rotor prepositioning. In the second stage, the inverter bridge output is controlled so that the amplitude of the motor phase current increases linearly according to the optimal starting current ramp-up slope, and open-loop commutation is performed according to the initial commutation frequency to drive the motor to accelerate. In the third stage, the actual speed of the motor is monitored in real time. When the actual speed exceeds the speed threshold for switching from open-loop start-up to closed-loop control, the control mode is seamlessly switched from open-loop commutation to field-oriented control closed-loop mode.

7. The method for wide-temperature start of a brushless motor in the industrial internet according to claim 1, characterized in that, It also includes closed-loop optimization steps: On the edge side, the multimodal raw data, predicted trajectory data, generated control parameters and startup results of the entire startup process will be packaged into a structured log. The structured logs are uploaded to the cloud server via an encrypted communication link; In the cloud, the structured logs uploaded from multiple edge devices are periodically aggregated and used to train and optimize the global thermodynamic model; The optimized global thermodynamic model is then quantized by integers and distributed to each edge device to update its local integer-quantized recurrent neural network prediction model.

8. The method for wide-temperature start of an industrial internet brushless motor according to any one of claims 1 to 7, characterized in that, The method is applicable to industrial brushless DC motors that require highly reliable starting over an extreme temperature range of -40°C to 100°C or wider.

9. The method for wide-temperature start of a brushless motor in the industrial internet according to claim 7, characterized in that, The cloud server uses a stream processing engine to perform macroscopic anomaly detection on the temperature data of a group of motors in a specific area. When a coordinated temperature change that conforms to statistical laws is detected, it triggers the generation and issuance of the global thermal environment early warning command.

10. A system for wide-temperature start of an industrial internet brushless motor for implementing the method of any one of claims 1 to 9, characterized in that, include: A multimodal sensor array module is used to embed inside a brushless motor to simultaneously acquire multiple analog signals of its thermal and electrical states. An edge computing control module, connected to the multimodal sensing array module, and comprising at least: The data preprocessing unit is used to process the multiple analog signals to construct a multi-dimensional time-series feature matrix; The local temperature prediction unit has an embedded integer quantization recurrent neural network model for predicting the local temperature trajectory based on the feature matrix. The prediction result fusion unit is used to fuse the local temperature trajectory with the early warning instructions issued by the cloud to generate the final temperature trajectory; The startup parameter generation unit is used to calculate the optimal startup control parameters based on the final temperature trajectory and the built-in motor physical parameterization model. A drive execution unit is used to generate a drive signal based on the optimal start control parameters to control the motor to start. An industrial internet communication module, integrated into the edge computing control module, is used to establish a secure and reliable two-way communication link with the cloud. A cloud-based intelligent analysis server, which communicates with multiple of the aforementioned industrial internet communication modules, and includes at least: The macroscopic anomaly detection unit is used to analyze group data and generate global thermal environment early warning commands; The global model training unit is used to iteratively optimize the global hot model using aggregated edge data; The model distribution unit is used to distribute the optimized model parameters to the edge side to update the local model.

Images