Motor working condition monitoring method with high-speed data transmission capability

Abstract

The invention discloses a motor working condition monitoring method with a high-speed data transmission capability, and relates to the field of motor operation working condition detection and information transmission. The method comprises the following steps of: synchronously acquiring multi-dimensional time sequence data of a motor under multiple working conditions, and performing data segmentation, standardization, fixed-step downsampling, frequency domain conversion and feature fusion to form a multi-dimensional feature vector; then building a BiLSTM-Attention deep learning model, extracting bidirectional time sequence dependence of the sequence, focusing key time step features, and training to obtain a working condition monitoring model; the monitoring reliability is guaranteed through a dual detection mechanism, and high-speed synchronous transmission of multi-sensor data is realized based on an EtherCAT communication protocol stack. According to the method, through collaborative modeling of multi-dimensional time-frequency feature fusion and a deep learning model, a high-speed synchronous transmission protocol stack and a dual detection mechanism are combined, motor working condition feature expression can be remarkably enhanced, and high-precision data transmission and reliable monitoring can be achieved.

Description

Technical Field

[0001] The present invention relates to the field of motor operating condition detection and information transmission, and in particular to a motor operating condition monitoring method with high-speed data transmission capability. Background Art

[0002] In modern industrial production, electric motors, as core power equipment, are crucial for the smooth operation of the entire production system. However, motor operation is affected by a variety of factors, making them prone to failure. This can lead to production interruptions, equipment damage, and even safety accidents, resulting in significant economic losses. Therefore, accurately monitoring motor operating conditions and promptly identifying potential faults have become critical issues that need to be addressed in the industrial sector.

[0003] Currently, existing technologies in the field of motor operating condition monitoring and control have many shortcomings that are difficult to ignore. In terms of multi-dimensional coupling analysis, traditional monitoring methods are often limited to independent analysis of single-dimensional data. For example, they only focus on the electrical parameters of the motor, such as voltage and current, but ignore the close connection between mechanical parameters (such as speed, torque, vibration acceleration) and thermal characteristic parameters (such as stator winding and core temperature, and rotor infrared temperature). This single-dimensional analysis method cannot fully capture the complex coupling relationship between the electrical, mechanical, and thermal multi-physics fields during motor operation. In fact, during motor operation, electrical faults may cause abnormal mechanical vibration, which in turn leads to temperature increase; conversely, wear of mechanical components can also affect the electrical performance of the motor. These parameters are interrelated and influence each other. However, traditional methods cannot effectively tap into these potential connections, resulting in a lack of in-depth understanding of the motor's operating status and difficulty in accurately judging the motor's true operating status under complex operating conditions.

[0004] In terms of identifying non-steady-state conditions, monitoring methods based on shallow neural networks or traditional machine learning algorithms (such as support vector machines, hidden Markov models, etc.) have obvious shortcomings. These methods are difficult to effectively mine long-distance time series dependencies when processing time series data of motor operation. Under non-steady-state conditions such as the start-stop transition process and load mutation, the operating parameters of the motor will undergo rapid and complex changes, showing dynamic characteristic patterns. For example, at the moment of motor startup, parameters such as speed, torque, and current will fluctuate violently in a short period of time, and the change trends of these parameters are interrelated. However, due to the limitations of their own algorithms, the existing methods are unable to fully capture the characteristic patterns of these dynamic changes, and the recognition accuracy of non-steady-state conditions is low. It is easy to make misjudgments or missed judgments, which makes it difficult to meet the needs of accurate motor monitoring in actual production.

[0005] Data transmission is also a weak link in existing technologies. Currently, many industrial sites still use traditional industrial bus protocols such as Modbus and CAN for data transmission. These protocols have significant limitations in data synchronization accuracy and transmission speed, typically with data transmission rates below 10 Mbps. In motor condition monitoring scenarios, large amounts of multi-dimensional data from various sensors must be collected and transmitted in real time, including electrical, mechanical, and thermal characteristics. Due to the limited transmission rates of traditional protocols, they cannot meet the high-speed requirements of multi-sensor synchronous data collection, resulting in high data transmission latency. For example, in high-speed motor systems, if sensor data cannot be transmitted to the monitoring center for analysis and processing in a timely manner, the monitoring system will be unable to respond to motor anomalies, delaying the optimal time for troubleshooting. Furthermore, traditional protocols also have low data synchronization accuracy, unable to ensure precise temporal alignment of sensor data. This significantly reduces the accuracy of condition analysis and fault diagnosis based on this data, seriously impacting the real-time performance and reliability of the monitoring system.

[0006] In summary, the shortcomings of existing motor condition monitoring technologies in multi-dimensional coupling analysis, identification of non-steady-state conditions, and data transmission have become bottlenecks hindering the intelligent and efficient operation of industrial motors. To improve the safety, stability, and efficiency of industrial production, an innovative technical solution is urgently needed that can integrate multi-dimensional feature deep modeling, achieve high-precision timing analysis, and possess high-speed data transmission capabilities. This effectively addresses the limitations of existing technologies and meets the stringent requirements of modern industry for motor condition monitoring. Summary of the Invention

[0007] The present invention provides a motor operating condition monitoring method with high-speed data transmission capability, which aims to solve the problems of insufficient multi-dimensional coupling analysis, low accuracy in identifying non-steady-state operating conditions, and data transmission delay in the prior art.

[0008] To achieve the above object, the present invention provides a technical solution: a motor operating condition monitoring method with high-speed data transmission capability, characterized in that it includes the following steps:

[0009] Step A1: Testing the motor under different operating conditions, including rated load, variable load, start-stop transition process, and fault simulation conditions. Using voltage sensors, current sensors, speed-torque sensors, temperature sensors, vibration sensors, and infrared thermal imaging devices in combination with high-speed data transmission methods, synchronously collect time-aligned multi-dimensional time series data to establish a dataset for model training and validation; the dataset includes:

[0010] Electrical parameters: d-axis and q-axis voltage, d-axis and q-axis current;

[0011] Mechanical parameters: speed, torque, vibration acceleration;

[0012] Thermal characteristic parameters: stator winding and core temperature, rotor infrared temperature;

[0013] Step A2: preprocessing the collected data, wherein the data preprocessing specifically includes data segmentation, Z-score normalization, fixed-step downsampling, FFT, and feature fusion, wherein the FFT refers to fast Fourier transform, to obtain a preprocessed multi-dimensional feature vector;

[0014] Step A3: Build the BiLSTM-Attention deep learning model architecture, including the input layer, bidirectional LSTM layer, attention mechanism layer, and fully connected layer;

[0015] Step A4: Input the preprocessed multi-dimensional feature vector into the model for training, use the cross entropy loss function as the target, optimize the parameters through back propagation, and evaluate the accuracy of the working condition classification based on the validation set until convergence to obtain a trained BiLSTM-Attention deep learning model;

[0016] Step A5: Collect motor operation data through a high-speed data transmission method, input the collected motor operation data into the trained model, and obtain a judgment of the motor operation condition, wherein the motor operation condition includes normal operation, start-stop transition, load mutation, and fault state.

[0017] Furthermore, the data in step A2 is divided into 70% training set, 20% validation set, and 10% test set using stratified random sampling, using random states to ensure repeatability; the Z-score normalization is achieved by a Standard Scaler;

[0018] The downsampling adopts a fixed step size strategy, and the step size is determined according to the Nyquist sampling theorem of the highest frequency signal;

[0019] The fast Fourier transform technology converts time domain signals such as vibration acceleration and current into frequency domain to extract fundamental frequency, harmonic components and sideband frequency characteristics;

[0020] The feature fusion refers to performing FFT conversion on the vibration acceleration and current time domain signals into frequency domain signals, extracting the fundamental frequency, harmonic components, and sideband frequencies as frequency domain features, and splicing them with the mean, variance, and peak value of the original time domain signals according to the time step to form a multi-dimensional feature vector containing time domain-frequency domain information.

[0021] Furthermore, the deep learning training model BiLSTM-Attention in step A3, that is, a bidirectional long short-term memory-attention neural network model, includes an input layer, a bidirectional LSTM layer, an attention mechanism layer and a fully connected layer; the input layer is used to receive preprocessed data, and the bidirectional LSTM layer includes a forward LSTM and a backward LSTM layer; the number of hidden units in each layer is the same, the forward LSTM layer processes data from the starting point to the end point of the sequence, and outputs a forward hidden state sequence, and the backward LSTM layer processes data in reverse from the end point to the starting point of the sequence, outputs a backward hidden state sequence, and splices the forward and backward hidden states according to the time step and serves as the input of the attention mechanism layer; the attention mechanism layer adopts Bahdanau addition attention to calculate the attention weight of each time step and generate a context vector.

[0022] Furthermore, the forward LSTM layer can be expressed by the following formula:

[0023]

[0024] The calculation formula of the backward LSTM layer is as follows:

[0025]

[0026] Where, the superscript (f) represents forward propagation, and the superscript (b) represents backward propagation; t is the input feature vector at time step t, W f is the input weight matrix of the forget gate, U f is the hidden state weight matrix of the forget gate, W i 、W c 、W o is the input weight matrix of the input gate, cell state, and output gate, U i 、U c 、U o is the hidden state weight matrix of the input gate, cell state, and output gate; b f is the forget gate bias vector, b i is the input gate bias vector, b c is the candidate cell state bias vector, b o is the output gate bias vector; h t is the hidden state at time step t, h t-1 is the hidden state at the previous moment; i t is the output of the input gate; is the cell state, c t is the candidate cell state; o tis the output of the output gate; σ is the sigmoid function, which maps the value to the interval (0,1) and is used for the gating mechanism; ⊙ is the element-wise multiplication; the forward and backward hidden states are concatenated according to the time step and serve as the input of the force mechanism layer; tanh is the hyperbolic tangent activation function, which is used to map the value to the interval (-1,1);

[0027] The Bahdanau additive attention can be expressed as follows:

[0028]

[0029] e t is the energy value of time step i; v a is the attention weight vector; h i is the concatenated hidden state, which represents the concatenation result of the forward and backward hidden states at time step t; W a is the hidden state weight matrix; α t is the attention weight at time step t; the output of the LSTM layer is used as the input of the attention layer at each time step; c is the context vector, which is the weighted sum of the output of the attention mechanism and is passed to the fully connected layer; T is the total length of the time series;

[0030] The fully connected layer can be expressed by the following formula:

[0031] y=softmax(W d ·h t +b d )

[0032] Where y is the class probability vector, W d is the weight matrix of the fully connected layer, b d is the fully connected layer bias vector, h t is the hidden state at time step t, and the softmax function is a normalized exponential function.

[0033] Furthermore, for the future values ​​predicted by the LSTM model, dual testing is applied to the prediction results: statistical anomaly detection based on isolation forests and physical constraint verification based on the motor thermal balance equation; after the abnormal data is marked, it is reversely input into the model to trigger online fine-tuning and dynamically update the model parameters.

[0034] Furthermore, the high-speed data transmission method is implemented based on the EtherCAT communication protocol stack, which includes a physical layer, a data link layer, and an application layer. The method includes the following steps:

[0035] Step B1: At the physical layer, the EtherCAT master device generates an Ethernet frame at the beginning of each communication cycle. The frame contains process data of all slave stations. The EtherCAT master device, i.e., the device master station, is a programmable logic controller. The slave stations include motors and voltage sensors, current sensors, speed-torque sensors, temperature sensors, and vibration sensors. The process data includes sensor configuration instructions and model parameter update instructions. The Ethernet frame is sent to the EtherCAT network, and each slave station in the network receives the frame in turn.

[0036] Step B2: The EtherCAT slave controller (ESC chip) of each slave station processes the received Ethernet frames online at the data link layer and extracts the process data related to the slave station in real time, including sensor configuration instructions and model parameter update instructions;

[0037] Step B3: Each sensor in the slave station triggers the corresponding motor data acquisition function according to the sensor configuration instructions in the process data. The motor in the slave station adjusts the operating state parameters according to the model parameter update instructions in the process data, and sends the process data to the microcontroller through the process data interface (PDI), starting sensor data acquisition and adjustment of operating state parameters.

[0038] Step B4: At the application layer, the microcontroller preprocesses the data collected by the sensor, including the time-aligned multi-dimensional time series data and the motor operation data collected by the sensor, to generate a multi-dimensional feature vector, converts the multi-dimensional feature vector and the motor operation data into a digital signal as return data, encapsulates the return data, and inserts it into a return frame through the ESC chip. The return data is the data sent from the station back to the device master station, including motor status information, electrical parameters, mechanical parameters, and thermal characteristic parameters;

[0039] Step B5: At the application layer, the master device receives frames containing data returned by all slaves and parses the data for model training and updates the actual state of the motor.

[0040] Step B6: Synchronize the local clock of each slave station through the distributed clock mechanism to ensure that all slave stations trigger events simultaneously in each communication cycle.

[0041] Furthermore, the physical layer is based on Fast Ethernet cables and RJ45 connectors, with a data transmission rate of 100Mbps and support for full-duplex communication; the data link layer is implemented by the ESC chip, which is responsible for real-time processing of Ethernet frames at the hardware level, including data extraction and insertion, and the single slave processing delay is ≤100ns; the application layer is based on the CANopen over EtherCAT protocol, namely the Coe protocol, and supports the CiA402 device profile for defining the data structure and communication behavior of motors and sensors, while expanding support for model parameter object dictionaries.

[0042] Furthermore, the structure of the Ethernet frame in step B1 includes a frame header, a frame body, and a frame trailer.

[0043] The frame header contains a global timestamp with an accuracy of ≤100ns and a frame priority tag;

[0044] The frame body contains multiple EtherCAT datagrams, each datagram corresponds to the process data of a slave station, and the datagram format is extended to support model parameter transmission;

[0045] The frame tail contains a CRC32 checksum.

[0046] Furthermore, in step B3, the specific steps of processing at the application layer include:

[0047] Step B3-1: The slave's microcontroller reads data from the ESC chip via an SPI interface with a clock frequency ≥ 20 MHz, using a double buffer mechanism to ensure that data reading and processing are performed in parallel;

[0048] Step B3-2: The microcontroller maps the read data to the corresponding object in the CANopen object dictionary, where:

[0049] Sensor configuration commands are mapped to standard CiA402 objects;

[0050] Model parameter update instructions are mapped to the extended object dictionary;

[0051] Step B3-3: The microcontroller parses the data according to the CiA402 device profile:

[0052] For sensor data, convert PDO data into physical measurement values;

[0053] For model parameters, verify the checksum and update the local model weights, and trigger the model hot loading mechanism.

[0054] Furthermore, the distributed clock mechanism in step B6 includes:

[0055] Step B6-1: The master device sends multiple test frames at startup, measures the frame arrival time of each slave station on the forward path and return path, and calculates the propagation delay of each slave station;

[0056] Step B6-2: The master device sends the propagation delay information of each slave station to the slave station, and the slave station adjusts its local clock according to the information;

[0057] Step B6-3: The master device periodically reads the global time from the reference slave and sends it to other slaves to compensate for clock drift.

[0058] Step B6-4: The local clock of each slave station generates a SYNC0 signal according to the global time, triggering a control loop.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] (1) The present invention proposes a motor operating condition monitoring method with high-speed data transmission capability. By using multiple sensors to synchronously collect multi-dimensional time series data such as the electrical, mechanical, and thermal characteristics of the motor under different operating conditions, a complete data set is constructed, avoiding the one-sidedness of single-dimensional data and laying a solid foundation for accurately judging the motor operating status.

[0061] (2) The BiLSTM-Attention deep learning training model is adopted, which combines the bidirectional temporal modeling capability of the bidirectional long short-term memory network with the dynamic weight allocation advantage of the attention mechanism. It is suitable for processing high-dimensional, multi-noise and multi-sensor synchronous acquisition signals of motors with complex coupling relationships, effectively improving the working condition classification accuracy and fault detection recall rate.

[0062] (3) The introduction of fast Fourier transform to perform frequency domain conversion on time domain signals such as vibration acceleration and current can convert periodic anomalies that are difficult to identify in the time domain into characteristic frequencies with clear physical meaning in the frequency domain, thus solving the problem that traditional methods are unable to capture the hidden characteristics of early faults.

[0063] (4) For the future values ​​predicted by the deep learning model, dual detection is performed using statistical anomaly detection based on isolation forests and physical constraint verification based on the motor thermal balance equation, so that the big data model can continuously optimize itself, continuously improve the accuracy and reliability of the prediction, and better adapt to various changes in the motor operation process.

[0064] (5) The present invention proposes a motor operating condition monitoring method with high-speed data transmission capability. The high-speed data transmission method implemented based on the EtherCAT communication protocol stack has efficient data transmission capability and can accurately parse and process data.

[0065] (6) The present invention proposes a motor condition monitoring method with high-speed data transmission capability, which synchronizes the local clock of each slave station through a distributed clock mechanism to ensure that all slave stations trigger events simultaneously in each communication cycle, further improving the accuracy and real-time performance of data transmission, and providing a strong guarantee for the stable operation of the motor condition monitoring system.

[0066] (7) The present invention proposes a motor operating condition monitoring method with high-speed data transmission capability. The master device periodically reads the global time from the reference slave station and sends it to other slave stations to compensate for clock drift.

[0067] (8) The present invention proposes a motor operating condition monitoring method with high-speed data transmission capability. The dynamic adjustment mechanism adopted can continuously maintain the synchronization state of the clocks of each slave station. Even during long-term operation, it can effectively avoid the accumulation of synchronization errors caused by clock drift, greatly improving the synchronization accuracy between devices and ensuring that the transmission and processing of motor signals and sensor data are highly coordinated in time. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0069] Figure 1 This is a flow chart of a motor operating condition monitoring method with high-speed data transmission capability proposed by the present invention;

[0070] Figure 2 This is the internal structure diagram of the deep learning model LSTM;

[0071] Figure 3 This is the internal structure diagram of the deep learning model BiLSTM-Attention;

[0072] Figure 4 It is a flow chart of a high-speed data transmission method;

[0073] Figure 5 This is the principle diagram of the distributed clock synchronization mechanism. DETAILED DESCRIPTION

[0074] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and examples. The present invention discloses a motor operating condition monitoring method with high-speed data transmission capability, which relates to the field of motor operating condition detection and information transmission, and aims to solve the problems of insufficient multi-dimensional coupling analysis, low accuracy in identifying non-steady-state conditions, and data transmission delay in the prior art. The flow chart is as follows: Figure 1 As shown, the following steps are included:

[0075] Step S1: Testing the motor under different operating conditions, including rated load, variable load, start-stop transition process, and fault simulation conditions. Time-aligned multi-dimensional time series data are collected synchronously using voltage sensors, current sensors, temperature sensors, speed-torque sensors, temperature sensors, vibration sensors, and infrared thermal imaging devices to establish a data set for model training and verification. The data set includes:

[0076] Electrical parameters: d-axis and q-axis voltage, d-axis and q-axis current;

[0077] Mechanical parameters: speed, torque, vibration acceleration;

[0078] Thermal characteristic parameters: stator winding and core temperature, rotor infrared temperature;

[0079] The established multi-dimensional time series is not limited to electrical data, but integrates electrical parameters, mechanical parameters, and thermal characteristic parameters to construct a complete data set that includes electrical-mechanical-thermal coupling characteristics. It can not only capture the explicit state of steady-state operation (such as current stability under rated load), but also identify the implicit characteristics of dynamic processes or early faults (such as timing delays in speed and torque during start-stop, and abnormal vibration sideband frequency and temperature rise associated with bearing wear), providing the model with cross-domain complementary time series information and avoiding the one-sidedness of single-dimensional data.

[0080] This data acquisition strategy enables the model to learn complex relationships between different parameters, such as the synchronization anomalies of vibration peaks and current harmonics, and enhance its adaptability to non-steady-state operating conditions such as start-stop transitions and sudden load changes. At the same time, it resists sensor noise through redundant features, improves the sensitivity and robustness of fault detection, and lays a multi-dimensional feature foundation for deep learning models to accurately judge the operating status of the motor.

[0081] Step S2: preprocessing the collected data, wherein the data preprocessing specifically includes data segmentation, Z-score normalization, fixed-step downsampling, FFT and feature fusion, wherein FFT refers to fast Fourier transform;

[0082] FFT can convert periodic anomalies that are difficult to identify in the time domain, such as the periodic vibration shock caused by bearing wear and the harmonic current distortion caused by broken rotor bars, into specific frequency components in the frequency domain, such as higher harmonics of the fundamental frequency and frequency-shifting sidebands. For example, a motor bearing failure will produce sideband frequencies in the vibration signal related to the bearing roller frequency, while abnormal harmonic components in the current signal directly indicate inverter failure or winding short circuits. These frequency domain characteristics are key indicators for early fault warning, enabling the early identification of subtle anomalies that are difficult to detect in the time domain waveform, thus resolving the problem of single-time domain analysis being insensitive to periodic faults.

[0083] Frequency domain features such as fundamental frequency, harmonics, and sideband frequency extracted by FFT are combined with statistical features such as mean, variance, and peak value in the time domain to form a multi-dimensional feature vector containing dual "time-frequency" information. Time domain features describe the amplitude changes of the signal, while frequency domain features reveal the frequency composition of the signal. The two complement each other to fully characterize the operating status of the motor. For example, during the start-stop transition process, the time domain step change of the speed combined with the frequency domain fundamental frequency offset of the torque signal can more accurately describe the dynamics of electromechanical coupling; and the increase in current time domain fluctuations and frequency domain harmonic components during sudden load changes can assist the model in distinguishing between "normal load changes" and "abnormal fluctuations caused by faults" to avoid misjudgment in a single dimension. FFT can effectively remove high-frequency noise through frequency domain filtering, retain characteristic frequencies related to the physical characteristics of the motor, reduce data redundancy, and improve the feature signal-to-noise ratio.

[0084] Step S3: Build a BiLSTM-Attention deep learning training model, including the input layer, bidirectional LSTM layer, attention mechanism layer, and fully connected layer;

[0085] The emergence of long short-term memory (LSTM) recurrent neural networks aims to overcome the challenges faced by traditional recurrent neural networks (RNNs), especially the gradient disappearance or gradient explosion problem. Figure 2 This figure shows a common LSTM cell architecture. A standard LSTM cell consists of a memory cell and three gates: an input gate, a forget gate, and an output gate. These gates control the proportion of information discarded or passed to the next time step. The unique feature of LSTM networks lies in their memory cells, which can preserve long-term dependencies between events in the input sequence, distinguishing them from traditional RNNs. These three gates play a crucial role in managing the flow of information into and out of the memory cell.

[0086] Bidirectional long short-term memory (BiLSTM) networks extend the capabilities of traditional LSTM networks by processing the input sequence both forward and backward. This bidirectional processing enables BiLSTM to capture dependencies between past and future contexts in the sequence, providing a more comprehensive understanding of temporal relationships in the data. By combining information from both directions, BiLSTM performs better in tasks that require understanding the full context of the sequence, such as time series analysis. The BiLSTM architecture is essentially a combination of two LSTM layers: one processes the input sequence forward, and the other processes it backward, with their outputs typically concatenated or merged before being passed to subsequent layers.

[0087] The attention mechanism mimics the human ability to focus on key information. By calculating weights for each time step in a sequence, it dynamically allocates the model's attention to features at different locations. In time series data processing, this mechanism enables the model to automatically focus on the time step most relevant to the task when generating current state judgments. For example, this can be done by focusing on vibration peaks or temperature spikes at the moment of a fault, thus avoiding interference from redundant information. The common Bahdanau additive attention method uses an energy function to calculate feature relevance, generating a context vector containing key information, thereby improving the model's accuracy in capturing local key features.

[0088] like Figure 3 As shown, BiLSTM-Attention combines the bidirectional time series modeling capabilities of BiLSTM with the dynamic weight allocation advantages of the attention mechanism: First, BiLSTM extracts bidirectional contextual features from the sequence. Then, the attention mechanism uses the bidirectional hidden state to filter the most critical time step features for current operating condition classification or fault detection, ultimately outputting a context vector focused on key information. This architecture preserves the global dependencies of time series data while highlighting key local features. It is suitable for processing high-dimensional, noisy time series data with complex coupling relationships, such as the motor-mechanical-thermal signals collected simultaneously by multiple motor sensors.

[0089] Step S4: The preprocessed multi-dimensional feature vector is input into the model training, with the cross entropy loss function as the target, and the parameters are optimized through back propagation. The working condition classification accuracy, fault detection recall rate and other indicators are evaluated in combination with the validation set to train a large model of motor parameters; the optimizer is selected as Adam, the learning rate is 0.001, the model is trained for 50 epochs, and Early Stopping is adopted, that is, if the validation set loss does not decrease for 10 consecutive rounds, it will stop.

[0090] The cross entropy loss function is used for multi-classification problems. It calculates the difference between the category probability distribution predicted by the model and the true label distribution. For training data with a batch size of N, the encoding vector [y i1 ,y i2 ,....,y iC] Where C is the number of working condition categories, and the category probability vector predicted by the model is [p i1 ,p i2 ,....,p iC ], then the cross entropy loss of a single sample is:

[0091]

[0092] The batch average cross entropy loss is:

[0093]

[0094] Among them, y ic Indicates the indicator value of the i-th sample in the true label belonging to category C, 1 means it belongs to the category, and 0 means it does not belong to the category; p ic represents the probability that the model predicts that the i-th sample belongs to category C and

[0095] Step S5: input the collected motor operation data into the trained model to obtain the motor operation condition judgment, wherein the motor operation condition specifically includes normal operation, start-stop transition, load mutation, and fault state.

[0096] Furthermore, the data segmentation in step S2 is divided into 70% training set, 20% validation set, and 10% test set by stratified random sampling, and random state is used to ensure repeatability; the Z-score normalization is achieved by StandardScaler, and the specific formula is: Where μ is the feature mean, σ is the feature standard deviation, and the standardized data satisfies the mean of 0 and the standard deviation of 1. The Standard Scaler is a standard scaler;

[0097] The downsampling adopts a fixed step size strategy, and the step size is determined according to the Nyquist sampling theorem of the highest frequency signal;

[0098] The fast Fourier transform technology converts time-domain signals such as vibration acceleration and current into frequency domain to extract features such as fundamental frequency, harmonic components, and sideband frequency;

[0099] The feature fusion refers to performing FFT conversion on time domain signals such as vibration acceleration and current into frequency domain signals, extracting the fundamental frequency, harmonic components, and sideband frequencies as frequency domain features, and splicing them with the mean, variance, and peak value of the original time domain signal according to the time step to form a multi-dimensional feature vector containing time domain-frequency domain information.

[0100] Furthermore, a method for constructing a large model of motor parameters according to claim 1 is characterized in that the deep learning training model BiLSTM-Attention in step S3, that is, a bidirectional long short-term memory-attention neural network model, includes an input layer, a bidirectional LSTM layer, an attention mechanism layer and a fully connected layer; the input layer is used to receive preprocessed data, and the bidirectional LSTM layer includes a forward LSTM and a backward LSTM layer; the number of hidden units in each layer is the same, the forward LSTM layer processes data from the starting point to the end point of the sequence, and outputs a forward hidden state sequence, the backward LSTM layer processes data in reverse from the end point of the sequence to the starting point, outputs a backward hidden state sequence, and splices the forward and backward hidden states according to the time step and uses them as the input of the attention mechanism layer; the attention mechanism layer adopts Bahdanau addition attention to calculate the attention weight of each time step and generate a context vector.

[0101] The forward LSTM layer can be expressed as follows:

[0102]

[0103] The calculation formula of the backward LSTM layer is as follows:

[0104]

[0105] Where, the superscript (f) represents forward propagation, and the superscript (b) represents backward propagation; t is the input feature vector at time step t, W f is the input weight matrix of the forget gate, U f is the hidden state weight matrix of the forget gate, W i 、W c 、W o is the input weight matrix of the input gate, cell state, and output gate, U i 、U c 、U o is the hidden state weight matrix of the input gate, cell state, and output gate; b f is the forget gate bias vector, b i is the input gate bias vector, b c is the candidate cell state bias vector, b o is the output gate bias vector; h t is the hidden state at time step t, h t-1 is the hidden state at the previous moment; i t is the input gate output; c t is the cell state, c t is the candidate cell state; o tis the output of the output gate; σ is the sigmoid function, which maps the value to the (0,1) interval for the gating mechanism; ⊙ is the element-wise multiplication; the forward and backward hidden states are concatenated into As the input of the force mechanism layer; tanh is the hyperbolic tangent activation function, which is used to map the value to the (-1,1) interval;

[0106] The Bahdanau additive attention can be expressed as follows:

[0107]

[0108] e t is the energy value of time step i; v a is the attention weight vector; h i is the concatenated hidden state, which represents the concatenation result of the forward and backward hidden states at time step t; W a is the hidden state weight matrix; α t is the attention weight at time step t; the output of the LSTM layer is used as the input of the attention layer at each time step; c is the context vector, which is the weighted sum of the output of the attention mechanism and is passed to the fully connected layer; T is the total length of the time series;

[0109] The fully connected layer can be expressed by the following formula:

[0110] y=softmax(W d ·h t +b d )

[0111] Where y is the class probability vector, W d is the weight matrix of the fully connected layer, b d is the fully connected layer bias vector, h t is the hidden state at time step t, and the softmax function is a normalized exponential function.

[0112] Furthermore, for the future values ​​predicted by the BiLSTM-Attention model, dual checks are applied to the prediction results: statistical anomaly detection based on isolation forests and physical constraint verification based on the motor's thermal balance equation. Abnormal data is marked and fed back into the model, triggering online fine-tuning and dynamically updating model parameters. The isolation algorithm calculates the residual, and an anomaly is flagged if the residual exceeds three times the standard deviation for five consecutive time steps. The physical constraint verification is based on the thermal balance equation:

[0113]

[0114] T rotor_pred ≤T stator +80℃

[0115] Among them, Q loss Indicates the total motor loss, Q conduction is the conduction heat dissipation power, Q radiation is the radiation heat dissipation power, C p is the equivalent heat capacity of the motor, Indicates the rate of change of temperature over time, T rotor_pred is the predicted rotor temperature, T stator is the stator temperature. 80°C is the safety temperature threshold set based on Class B insulation.

[0116] Furthermore, the high-speed transmission method is implemented based on the EtherCAT communication protocol stack, which includes a physical layer, a data link layer, and an application layer. Figure 4 As shown, the method includes the following steps:

[0117] Step B1: At the physical layer, the EtherCAT master device generates an Ethernet frame at the beginning of each communication cycle. The frame contains process data of all slave stations. The EtherCAT master device, i.e., the device master station, is a programmable logic controller. The slave stations include motors and voltage sensors, current sensors, speed-torque sensors, temperature sensors, and vibration sensors. The process data includes sensor configuration instructions and model parameter update instructions. The Ethernet frame is sent to the EtherCAT network, and each slave station in the network receives the frame in turn.

[0118] Step B2: The EtherCAT slave controller (ESC chip) of each slave station processes the received Ethernet frames online at the data link layer and extracts the process data related to the slave station in real time, including sensor configuration instructions and model parameter update instructions;

[0119] Step B3: Each sensor in the slave station triggers the corresponding motor data acquisition function according to the sensor configuration instructions in the process data. The motor in the slave station adjusts the operating state parameters according to the model parameter update instructions in the process data, and sends the process data to the microcontroller through the process data interface (PDI), starting sensor data acquisition and adjustment of operating state parameters.

[0120] Step B4: At the application layer, the microcontroller preprocesses the data collected by the sensor, including the time-aligned multi-dimensional time series data and the motor operation data collected by the sensor, to generate a multi-dimensional feature vector, converts the multi-dimensional feature vector and the motor operation data into a digital signal as return data, encapsulates the return data, and inserts it into a return frame through the ESC chip. The return data is the data sent from the station back to the device master station, including motor status information, electrical parameters, mechanical parameters, and thermal characteristic parameters;

[0121] Step B5: At the application layer, the master device receives frames containing data returned by all slaves and parses the data for model training and updates the actual state of the motor.

[0122] Step B6: Synchronize the local clock of each slave station through the distributed clock mechanism to ensure that all slave stations trigger events simultaneously in each communication cycle.

[0123] Furthermore, the physical layer is based on Fast Ethernet cables and RJ45 connectors, with a data transmission rate of 100Mbps and support for full-duplex communication; the data link layer is implemented by the ESC chip, which is responsible for real-time processing of Ethernet frames at the hardware level, including data extraction and insertion, and the single slave processing delay is ≤100ns; the application layer is based on the CANopen over EtherCAT protocol, namely the Coe protocol, and supports the CiA402 device profile for defining the data structure and communication behavior of motors and sensors, while expanding support for model parameter object dictionaries.

[0124] Furthermore, the structure of the Ethernet frame in step B1 includes a frame header, a frame body, and a frame trailer.

[0125] The frame header contains a global timestamp with an accuracy of ≤100ns and a frame priority tag;

[0126] The frame body contains multiple EtherCAT datagrams, each datagram corresponds to the process data of a slave station, and the datagram format is extended to support model parameter transmission;

[0127] The frame tail contains a CRC32 checksum.

[0128] Furthermore, in step B3, the specific steps of processing at the application layer include:

[0129] Step B3-1: The slave's microcontroller reads data from the ESC chip via an SPI interface with a clock frequency ≥ 20 MHz, using a double buffer mechanism to ensure that data reading and processing are performed in parallel;

[0130] Step B3-2: The microcontroller maps the read data to the corresponding object in the CANopen object dictionary, where:

[0131] Sensor configuration commands are mapped to standard CiA402 objects;

[0132] Model parameter update instructions are mapped to the extended object dictionary;

[0133] Step B3-3: The microcontroller parses the data according to the CiA402 device profile:

[0134] For sensor data, convert PDO data into physical measurement values;

[0135] For model parameters, verify the checksum and update the local model weights, and trigger the model hot loading mechanism.

[0136] Furthermore, the distributed clock mechanism in step B6 includes:

[0137] Step B6-1: The master device sends multiple test frames at startup, measures the frame arrival time of each slave station on the forward path and return path, and calculates the propagation delay of each slave station;

[0138] Step B6-2: The master device sends the propagation delay information of each slave station to the slave station, and the slave station adjusts its local clock according to the information;

[0139] Step B6-3: The master device periodically reads the global time from the reference slave and sends it to other slaves to compensate for clock drift.

[0140] Step B6-4: The local clock of each slave station generates a SYNC0 signal according to the global time, triggering a control loop.

[0141] CANopen further defines two data exchange structures. A Process Data Object (PDO) is a structure containing object dictionary entries that are cyclically transmitted as process variables. Before cyclic communication is initiated during the configuration phase, specific object dictionary objects are mapped to this structure. Each PDO entry has a defined offset within the data set encapsulated in the Ethernet frame. This is important so that during the cyclic phase, the slave hardware knows exactly where the relevant data is located within the frame. After cyclic communication is initiated, PDO entries are exchanged between the master and slave in each cycle and cannot be changed without reconfiguring the network's communication configuration. A Service Data Object (SDO) contains object dictionary entries that are sent and received acyclically. The SDO acts as a mailbox, buffering both received and transmitted data. This type of communication is acyclic and depends on the available bandwidth during the communication cycle. This type of communication is non-deterministic and is best suited for transmitting configuration data.

[0142] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A motor operating condition monitoring method with high-speed data transmission capability, characterized in that: The steps include: Step A1: Testing the motor under different operating conditions, including rated load, variable load, start-stop transition process, and fault simulation conditions. Using voltage sensors, current sensors, speed-torque sensors, temperature sensors, vibration sensors, and infrared thermal imaging devices in combination with high-speed data transmission methods, synchronously collect time-aligned multi-dimensional time series data to establish a dataset for model training and validation; the dataset includes: Electrical parameters: d-axis and q-axis voltage, d-axis and q-axis current; Mechanical parameters: speed, torque, vibration acceleration; Thermal characteristic parameters: stator winding and core temperature, rotor infrared temperature; Step A2: preprocessing the collected data, wherein the data preprocessing specifically includes data segmentation, Z-score normalization, fixed-step downsampling, FFT, and feature fusion, wherein the FFT refers to fast Fourier transform, to obtain a preprocessed multi-dimensional feature vector; Step A3: Build the BiLSTM-Attention deep learning model architecture, including the input layer, bidirectional LSTM layer, attention mechanism layer, and fully connected layer; Step A4: Input the preprocessed multi-dimensional feature vector into the model for training, use the cross entropy loss function as the target, optimize the parameters through back propagation, and evaluate the accuracy of the working condition classification based on the validation set until convergence to obtain a trained BiLSTM-Attention deep learning model; Step A5: Collect motor operation data through a high-speed data transmission method, input the collected motor operation data into the trained model, and obtain a judgment of the motor operation condition, wherein the motor operation condition includes normal operation, start-stop transition, load mutation, and fault state.

2. A motor operating condition monitoring method with high-speed data transmission capability according to claim 1, characterized in that: The data in step A2 is divided into 70% training set, 20% validation set, and 10% test set by stratified random sampling, using random states to ensure repeatability; the Z-score normalization is achieved by Standard Scaler; The downsampling adopts a fixed step size strategy, and the step size is determined according to the Nyquist sampling theorem of the highest frequency signal; The fast Fourier transform technology converts time domain signals such as vibration acceleration and current into frequency domain to extract fundamental frequency, harmonic components and sideband frequency characteristics; The feature fusion refers to performing FFT conversion on the vibration acceleration and current time domain signals into frequency domain signals, extracting the fundamental frequency, harmonic components, and sideband frequencies as frequency domain features, and splicing them with the mean, variance, and peak value of the original time domain signals according to the time step to form a multi-dimensional feature vector containing time domain-frequency domain information.

3. The motor operating condition monitoring method with high-speed data transmission capability according to claim 1, characterized in that: The deep learning training model BiLSTM-Attention in step A3, i.e., a bidirectional long short-term memory-attention neural network model, includes an input layer, a bidirectional LSTM layer, an attention mechanism layer and a fully connected layer; the input layer is used to receive preprocessed data, and the bidirectional LSTM layer includes a forward LSTM layer and a backward LSTM layer; each layer has the same number of hidden units, the forward LSTM layer processes data from the start point to the end point of the sequence, and outputs a forward hidden state sequence, the backward LSTM layer processes data in reverse from the end point to the start point of the sequence, and outputs a backward hidden state sequence, and the forward and backward hidden states are spliced ​​according to the time step and used as the input of the attention mechanism layer; the attention mechanism layer adopts Bahdanau addition attention to calculate the attention weight of each time step and generate a context vector.

4. The motor operating condition monitoring method with high-speed data transmission capability according to claim 3, characterized in that: The forward LSTM layer can be expressed as follows: The calculation formula of the backward LSTM layer is as follows: Where, the superscript (f) represents forward propagation, and the superscript (b) represents backward propagation; t is the input feature vector at time step t, W f is the input weight matrix of the forget gate, U f is the hidden state weight matrix of the forget gate, W i 、W c 、W o is the input weight matrix of the input gate, cell state, and output gate, U i 、U c 、U o is the hidden state weight matrix of the input gate, cell state, and output gate; b f is the forget gate bias vector, b i is the input gate bias vector, b c is the candidate cell state bias vector, b o is the output gate bias vector; h t is the hidden state at time step t, h t-1 is the hidden state at the previous moment; i t is the output of the input gate; is the cell state, c t is the candidate cell state; o t is the output of the output gate; σ is the sigmoid function, which maps the value to the interval (0,1) and is used for the gating mechanism; ⊙ is the element-wise multiplication; the forward and backward hidden states are concatenated according to the time step and serve as the input of the force mechanism layer; tanh is the hyperbolic tangent activation function, which is used to map the value to the interval (-1,1); The Bahdanau additive attention can be expressed as follows: e t is the energy value of time step i; v a is the attention weight vector; h i is the concatenated hidden state, which represents the concatenation result of the forward and backward hidden states at time step t; W a is the hidden state weight matrix; α t is the attention weight at time step t; the output of the LSTM layer is used as the input of the attention layer at each time step; c is the context vector, which is the weighted sum of the output of the attention mechanism and is passed to the fully connected layer; T is the total length of the time series; The fully connected layer can be expressed by the following formula: y =softmax(W d ·h t +b d ) Where y is the class probability vector, W d is the weight matrix of the fully connected layer, b d is the fully connected layer bias vector, h t is the hidden state at time step t, and the softmax function is a normalized exponential function.

5. The method for monitoring motor operating conditions with high-speed data transmission capability according to claim 1, characterized in that: For the future values ​​predicted by the LSTM model, dual testing is applied to the prediction results: statistical anomaly detection based on isolation forests and physical constraint verification based on the motor thermal balance equation. After abnormal data is marked, it is input into the model in reverse, triggering online fine-tuning and dynamically updating the model parameters.

6. The method for monitoring motor operating conditions with high-speed data transmission capability according to claim 1, characterized in that: The high-speed data transmission method is implemented based on the EtherCAT communication protocol stack, which includes a physical layer, a data link layer, and an application layer. The method includes the following steps: Step B1: At the physical layer, the EtherCAT master device generates an Ethernet frame at the beginning of each communication cycle. The frame contains process data of all slave stations. The EtherCAT master device, i.e., the device master station, is a programmable logic controller. The slave stations include motors and voltage sensors, current sensors, speed-torque sensors, temperature sensors, and vibration sensors. The process data includes sensor configuration instructions and model parameter update instructions. The Ethernet frame is sent to the EtherCAT network, and each slave station in the network receives the frame in turn. Step B2: The EtherCAT slave controller (ESC chip) of each slave station processes the received Ethernet frames online at the data link layer and extracts the process data related to the slave station in real time, including sensor configuration instructions and model parameter update instructions; Step B3: Each sensor in the slave station triggers the corresponding motor data acquisition function according to the sensor configuration instructions in the process data. The motor in the slave station adjusts the operating state parameters according to the model parameter update instructions in the process data, and sends the process data to the microcontroller through the process data interface (PDI), starting sensor data acquisition and adjustment of operating state parameters. Step B4: At the application layer, the microcontroller preprocesses the data collected by the sensor, including the time-aligned multi-dimensional time series data and the motor operation data collected by the sensor, to generate a multi-dimensional feature vector, converts the multi-dimensional feature vector and the motor operation data into a digital signal as return data, encapsulates the return data, and inserts it into a return frame through the ESC chip. The return data is the data sent from the station back to the device master station, including motor status information, electrical parameters, mechanical parameters, and thermal characteristic parameters; Step B5: At the application layer, the master device receives frames containing data returned by all slaves and parses the data for model training and updates the actual state of the motor. Step B6: Synchronize the local clock of each slave station through the distributed clock mechanism to ensure that all slave stations trigger events simultaneously in each communication cycle.

7. The motor operating condition monitoring method with high-speed data transmission capability according to claim 6, characterized in that: The physical layer is based on Fast Ethernet cables and RJ45 connectors, with a data transmission rate of 100Mbps and support for full-duplex communication. The data link layer is implemented by an ESC chip, which is responsible for real-time processing of Ethernet frames at the hardware level, including data extraction and insertion, and the single-slave processing delay is ≤100ns. The application layer is based on the CANopen over EtherCAT protocol, namely the Coe protocol, and supports the CiA402 device profile for defining the data structure and communication behavior of motors and sensors, while also extending support for model parameter object dictionaries.

8. The method for monitoring motor operating conditions with high-speed data transmission capability according to claim 6, characterized in that: The structure of the Ethernet frame in step B1 includes a frame header, a frame body, and a frame trailer, wherein: The frame header contains a global timestamp with an accuracy of ≤100ns and a frame priority tag; The frame body contains multiple EtherCAT datagrams, each datagram corresponds to the process data of a slave station, and the datagram format is extended to support model parameter transmission; The frame tail contains a CRC32 checksum.

9. The method for monitoring motor operating conditions with high-speed data transmission capability according to claim 6, characterized in that: In step B3, the specific steps for processing at the application layer include: Step B3-1: The slave's microcontroller reads data from the ESC chip via an SPI interface with a clock frequency ≥ 20 MHz, using a double buffer mechanism to ensure that data reading and processing are performed in parallel; Step B3-2: The microcontroller maps the read data to the corresponding object in the CANopen object dictionary, where: Sensor configuration commands are mapped to standard CiA402 objects; Model parameter update instructions are mapped to the extended object dictionary; Step B3-3: The microcontroller parses the data according to the CiA402 device profile: For sensor data, convert PDO data into physical measurement values; For model parameters, verify the checksum and update the local model weights, and trigger the model hot loading mechanism.

10. The motor operating condition monitoring method with high-speed data transmission capability according to claim 6, characterized in that: The distributed clock mechanism in step B6 includes: Step B6-1: The master device sends multiple test frames at startup, measures the frame arrival time of each slave station on the forward path and return path, and calculates the propagation delay of each slave station; Step B6-2: The master device sends the propagation delay information of each slave station to the slave station, and the slave station adjusts its local clock according to the information; Step B6-3: The master device periodically reads the global time from the reference slave and sends it to other slaves to compensate for clock drift. Step B6-4: The local clock of each slave station generates a SYNC0 signal according to the global time, triggering a control loop.

Images