A parameter control method for ultrasonic motor

Abstract

The present invention relates to a parameter control method for an ultrasonic motor, which uses sensors to collect key operating parameters including driving frequency, driving voltage, output power, operating temperature, and internal strain parameters and load change data in real time. The collected data is input into a pre-trained fault prediction neural network model to generate fault risk assessment and prediction results of potential abnormality types. Combining the prediction results with the current operating parameters, an adjustment strategy is dynamically generated, including adjusting the driving frequency to reduce internal strain accumulation, adjusting the driving voltage to reduce operating stress and controlling the strain within a safe range. The adjustment strategy is executed in real time to optimize the motor operating parameters, and the operating parameters, prediction results and control strategies are stored in a database, and the real-time status, historical records and fatigue trends of the motor are intuitively displayed through a human-computer interaction interface. The present invention provides a dynamic parameter optimization method based on predictive analysis, which can effectively extend the life of the motor and improve operational reliability and safety.

Description

Technical Field

[0001] The present invention relates to the technical field of ultrasonic motor control, and in particular to a parameter control method of an ultrasonic motor. Background Art

[0002] Ultrasonic motors are high-precision drive devices that use the principle of electromagnetic induction to generate high-frequency vibrations. They are widely used in the field of miniaturization and high-precision motion control, such as drones, electric toothbrushes, etc. In the prior art, the control method of ultrasonic motors mainly relies on manual adjustment of preset parameters or simple feedback control based on limited sensor data, usually including a fixed adjustment range of driving frequency and voltage to cope with the efficiency and stability requirements under normal operating conditions.

[0003] However, there are several problems with existing technologies. First, traditional control methods lack the ability to perceive key motor operating parameters (such as strain, load changes) and operating environment changes in real time, resulting in low operating efficiency and stability. Second, existing methods rely on empirical values ​​or fixed rules for fault prediction and prevention, and cannot dynamically respond to complex operating conditions. In addition, the adjustment strategy of operating parameters is usually relatively simple, making it difficult to effectively deal with potential anomalies under different operating conditions.

[0004] Therefore, it is necessary to develop a new ultrasonic motor parameter control method. Summary of the invention

[0005] The present application provides a parameter control method for an ultrasonic motor to improve the operating reliability and safety of the ultrasonic motor.

[0006] The present application provides a parameter control method for an ultrasonic motor, comprising:

[0007] The key operating parameters of the ultrasonic motor are collected by sensors, wherein the key operating parameters include driving frequency, driving voltage, output power and operating temperature; the internal strain parameters of the motor structure and the motor load change data are collected;

[0008] Inputting the key operating parameters, internal strain parameters and motor load change data into a pre-trained fault prediction neural network model to obtain a prediction result of the current operating state of the ultrasonic motor, wherein the prediction result includes a fault risk assessment and a potential abnormality type;

[0009] According to the prediction results of the fault prediction model, combined with the current operating parameter data, an adjustment strategy is dynamically generated, wherein the adjustment strategy includes adjusting the driving frequency to reduce the internal strain accumulation; adjusting the driving voltage to reduce the operating stress, while controlling the internal strain within a safe range;

[0010] According to the generated adjustment strategy, the operating parameters of the ultrasonic motor are adjusted in real time;

[0011] The key operating parameters, internal strain parameters, motor load change data, prediction results and control strategies are stored in a database, and the real-time operating status, historical adjustment records and fatigue trend diagram of the motor structure are displayed to the user through a human-computer interaction interface.

[0012] Furthermore, the fault prediction neural network model includes a data preprocessing module, a feature extraction module, a fault classification module and a risk assessment module; wherein the data preprocessing module is used to receive the key operating parameters, internal strain parameters and motor load change data; a piecewise linear denoising method of a multi-dimensional time series signal is used to filter out high-frequency noise in the key operating parameters; and a dynamic time warping algorithm is used to perform time alignment on the key operating parameters, internal strain parameters and load change data to ensure the consistency and synchronization of multi-source data, and generate high-quality data after denoising and time alignment; the feature extraction module is used to receive the high-quality data provided by the data preprocessing module; a convolutional neural network is used to extract local features in the high-quality data; a long short-term memory network is used to capture the long-term dependencies in the high-quality data; the local features and long-term dependencies are fused to obtain a multi-dimensional feature vector; wherein the multi-dimensional feature vector includes the statistical characteristics of the operating parameters, the strain accumulation trend characteristics, and the local features of the high-quality data. The fault classification module is used to receive the multi-dimensional feature vector provided by the feature extraction module, and uses a classifier composed of a support vector machine and a multi-layer perceptron to process the received data to obtain the potential abnormality type under the current operating state; the risk assessment module is used to receive the abnormality type and historical operating parameters provided by the fault classification module; according to the received data, the probability of the fault occurrence is calculated through the Bayesian network to obtain the probability distribution of each fault type; the fuzzy logic model is used to input the probability of the fault type and the abnormal impact characteristics extracted from the historical data to calculate the impact range and severity of the abnormality on the motor operating state; combined with the abnormality type, the probability of the fault occurrence and its impact range, the strain safety threshold, safety stress upper limit, fatigue safety upper limit and fatigue cumulative value corresponding to each fault are obtained; a comprehensive evaluation result including the probability of fault occurrence, impact range, potential impact, strain safety threshold, safety stress upper limit, fatigue safety upper limit and fatigue cumulative value is generated.

[0013] Furthermore, the piecewise linear denoising method of the data preprocessing module comprises the following steps:

[0014] The key operating parameters, internal strain parameters and motor load change data are divided into time windows of fixed length; for the signal in each time window, a linear trend curve is fitted by calculating the mean and slope of the multidimensional data; the fitted linear trend curve is subtracted from the original signal to obtain a signal with low-frequency drift removed; a bandpass filter is used to filter out high-frequency noise from the signal in each time window, and the passband of the bandpass filter is dynamically adjusted according to the operating characteristics of the ultrasonic motor; the processing results of each time window are spliced ​​through a signal reconstruction method to generate an overall denoised signal.

[0015] Furthermore, the specific method of using the convolutional neural network in the feature extraction module to extract local features includes the following steps: dividing the high-quality data into multi-channel time-series inputs, wherein each channel corresponds to an operating parameter or a sensor signal; adjusting the size and step size of the convolution kernel in real time according to the dynamic changes in the operating state of the ultrasonic motor, wherein the adjustment is based on the rate of change of the operating parameters and the amplitude of load fluctuations, thereby optimizing the ability to capture local patterns; after the convolution operation, adopting a dynamically weighted multi-core feature fusion strategy, combined with the weight distribution of key operating parameters, to proportionally integrate low-frequency information and high-frequency information into a feature vector; screening the pooled features through a specific gating mechanism, retaining only features with a high correlation with the abnormal type to reduce redundancy; bidirectionally interacting the fused feature vector with the output of the long short-term memory network, extracting deep-level patterns and long-term dependency characteristics in the signal through a multi-layer feature fusion network, and obtaining a multi-dimensional feature vector containing statistical characteristics of operating parameters, strain accumulation trend characteristics, and load change rate characteristics.

[0016] Furthermore, the fault classification module further includes a dynamic feature weight allocation mechanism, specifically including: dynamically adjusting feature weights based on the correlation between the features of each dimension in the multidimensional feature vector and the known fault types in the historical operation data; wherein the correlation is calculated by a feature importance analysis model, and the feature importance analysis model is based on random forest or gradient boosting decision tree training, which is used to quantify the contribution of each feature to the classification accuracy.

[0017] Furthermore, the risk assessment module further includes a dynamic weighting mechanism for abnormal impact, specifically including: dynamically adjusting the impact weight of the fault type according to the abnormal type output by the fault classification module and the motor operating conditions extracted from the historical operating data; wherein, the calculation of the impact weight is based on a fuzzy logic model, and the fuzzy logic model combines the real-time fault probability, the abnormal impact range, the historical abnormal frequency, and the fluctuation range of key operating parameters to generate a comprehensive weight value through dynamic adjustment rules; the comprehensive weight value is used to correct the probability distribution of each fault type node in the Bayesian network, and affects the finally generated strain safety threshold, safety stress upper limit, fatigue safety upper limit and fatigue accumulation value, so as to ensure the adaptability and accuracy of the assessment results under different operating conditions.

[0018] Furthermore, the adjustment strategy is dynamically generated based on the prediction results of the fault prediction model and in combination with the current operating parameter data, including:

[0019] According to the following formula (1), the adjusted driving frequency is calculated :

[0020] ;

[0021] in, is the current driving frequency; A strain safety threshold predicted by the fault prediction neural network model; is the internal strain value measured at the current moment; The fatigue cumulative value predicted by the fault prediction neural network model; The fatigue safety upper limit predicted by the fault prediction neural network model; is the frequency adjustment coefficient; is the fatigue accumulation adjustment coefficient;

[0022] According to the following formula (2), the adjusted driving voltage is calculated :

[0023] ;

[0024] in, is the current driving voltage; is the front running stress; The upper limit of the safety stress predicted by the fault prediction neural network model; The fatigue cumulative value predicted by the fault prediction neural network model; The fatigue safety upper limit predicted by the fault prediction neural network model; is the stress adjustment coefficient; is the fatigue adjustment factor.

[0025] Furthermore, the human-computer interaction interface includes a real-time status display module and an abnormal alarm module, wherein the real-time status display module is used to display the real-time change curves of the driving frequency, driving voltage, output power, operating temperature and internal strain in the form of a visual chart; the abnormal alarm module uses multi-level color coding and sound alarms to prompt the user when a high-risk abnormality occurs based on the fault risk assessment in the prediction results, and provides a detailed description of the specific abnormality type and its impact range.

[0026] Furthermore, the human-computer interaction interface includes a custom analysis module, through which the user can select key operating parameters and adjustment records within a specific time period for playback analysis; the custom analysis module supports interactive switching of historical trend charts and adjustment strategy execution effects of different abnormality types, including dynamic changes in fatigue accumulation values ​​and strain safety thresholds, to help users evaluate the long-term effectiveness of adjustment strategies.

[0027] The beneficial effects of the technical solution provided by this application include: (1) This method introduces a pre-trained fault prediction neural network model, based on key operating parameters, internal strain parameters and load change data, to achieve real-time assessment of ultrasonic motor fault risks and accurate prediction of potential abnormal types, thereby effectively improving the reliability and safety of motor operation. (2) According to the fault prediction results and current operating parameters, an adjustment strategy is dynamically generated, which can adjust the drive frequency and drive voltage in real time to reduce internal strain accumulation and reduce operating stress, ensuring that the motor always operates efficiently within a safe range and prolonging the service life of the equipment. This application provides a parameter control method for an ultrasonic motor to improve the operating reliability and safety of an ultrasonic motor.

[0028] The present application provides a parameter control method for an ultrasonic motor, comprising:

[0029] The key operating parameters of the ultrasonic motor are collected by sensors, wherein the key operating parameters include driving frequency, driving voltage, output power and operating temperature; the internal strain parameters of the motor structure and the motor load change data are collected;

[0030] Inputting the key operating parameters, internal strain parameters and motor load change data into a pre-trained fault prediction neural network model to obtain a prediction result of the current operating state of the ultrasonic motor, wherein the prediction result includes a fault risk assessment and a potential abnormality type;

[0031] According to the prediction results of the fault prediction model, combined with the current operating parameter data, an adjustment strategy is dynamically generated, wherein the adjustment strategy includes adjusting the driving frequency to reduce the internal strain accumulation; adjusting the driving voltage to reduce the operating stress, while controlling the internal strain within a safe range;

[0032] According to the generated adjustment strategy, the operating parameters of the ultrasonic motor are adjusted in real time;

[0033] The key operating parameters, internal strain parameters, motor load change data, prediction results and control strategies are stored in a database, and the real-time operating status, historical adjustment records and fatigue trend diagram of the motor structure are displayed to the user through a human-computer interaction interface.

[0034] Furthermore, the fault prediction neural network model includes a data preprocessing module, a feature extraction module, a fault classification module and a risk assessment module; wherein the data preprocessing module is used to receive the key operating parameters, internal strain parameters and motor load change data; a piecewise linear denoising method of a multi-dimensional time series signal is used to filter out high-frequency noise in the key operating parameters; and a dynamic time warping algorithm is used to perform time alignment on the key operating parameters, internal strain parameters and load change data to ensure the consistency and synchronization of multi-source data, and generate high-quality data after denoising and time alignment; the feature extraction module is used to receive the high-quality data provided by the data preprocessing module; a convolutional neural network is used to extract local features in the high-quality data; a long short-term memory network is used to capture the long-term dependencies in the high-quality data; the local features and long-term dependencies are fused to obtain a multi-dimensional feature vector; wherein the multi-dimensional feature vector includes the statistical characteristics of the operating parameters, the strain accumulation trend characteristics, and the local features of the high-quality data. The fault classification module is used to receive the multi-dimensional feature vector provided by the feature extraction module, and uses a classifier composed of a support vector machine and a multi-layer perceptron to process the received data to obtain the potential abnormality type under the current operating state; the risk assessment module is used to receive the abnormality type and historical operating parameters provided by the fault classification module; according to the received data, the probability of the fault occurrence is calculated through the Bayesian network to obtain the probability distribution of each fault type; the fuzzy logic model is used to input the probability of the fault type and the abnormal impact characteristics extracted from the historical data to calculate the impact range and severity of the abnormality on the motor operating state; combined with the abnormality type, the probability of the fault occurrence and its impact range, the strain safety threshold, safety stress upper limit, fatigue safety upper limit and fatigue cumulative value corresponding to each fault are obtained; a comprehensive evaluation result including the probability of fault occurrence, impact range, potential impact, strain safety threshold, safety stress upper limit, fatigue safety upper limit and fatigue cumulative value is generated.

[0035] Furthermore, the piecewise linear denoising method of the data preprocessing module comprises the following steps:

[0036] The key operating parameters, internal strain parameters and motor load change data are divided into time windows of fixed length; for the signal in each time window, a linear trend curve is fitted by calculating the mean and slope of the multidimensional data; the fitted linear trend curve is subtracted from the original signal to obtain a signal with low-frequency drift removed; a bandpass filter is used to filter out high-frequency noise from the signal in each time window, and the passband of the bandpass filter is dynamically adjusted according to the operating characteristics of the ultrasonic motor; the processing results of each time window are spliced ​​through a signal reconstruction method to generate an overall denoised signal.

[0037] Furthermore, the specific method of using the convolutional neural network in the feature extraction module to extract local features includes the following steps: dividing the high-quality data into multi-channel time-series inputs, wherein each channel corresponds to an operating parameter or a sensor signal; adjusting the size and step size of the convolution kernel in real time according to the dynamic changes in the operating state of the ultrasonic motor, wherein the adjustment is based on the rate of change of the operating parameters and the amplitude of load fluctuations, thereby optimizing the ability to capture local patterns; after the convolution operation, adopting a dynamically weighted multi-core feature fusion strategy, combined with the weight distribution of key operating parameters, to proportionally integrate low-frequency information and high-frequency information into a feature vector; screening the pooled features through a specific gating mechanism, retaining only features with a high correlation with the abnormal type to reduce redundancy; bidirectionally interacting the fused feature vector with the output of the long short-term memory network, extracting deep-level patterns and long-term dependency characteristics in the signal through a multi-layer feature fusion network, and obtaining a multi-dimensional feature vector containing statistical characteristics of operating parameters, strain accumulation trend characteristics, and load change rate characteristics.

[0038] Furthermore, the fault classification module further includes a dynamic feature weight allocation mechanism, specifically including: dynamically adjusting feature weights based on the correlation between the features of each dimension in the multidimensional feature vector and the known fault types in the historical operation data; wherein the correlation is calculated by a feature importance analysis model, and the feature importance analysis model is based on random forest or gradient boosting decision tree training, which is used to quantify the contribution of each feature to the classification accuracy.

[0039] Furthermore, the risk assessment module further includes a dynamic weighting mechanism for abnormal impact, specifically including: dynamically adjusting the impact weight of the fault type according to the abnormal type output by the fault classification module and the motor operating conditions extracted from the historical operating data; wherein, the calculation of the impact weight is based on a fuzzy logic model, and the fuzzy logic model combines the real-time fault probability, the abnormal impact range, the historical abnormal frequency, and the fluctuation range of key operating parameters to generate a comprehensive weight value through dynamic adjustment rules; the comprehensive weight value is used to correct the probability distribution of each fault type node in the Bayesian network, and affects the finally generated strain safety threshold, safety stress upper limit, fatigue safety upper limit and fatigue accumulation value, so as to ensure the adaptability and accuracy of the assessment results under different operating conditions.

[0040] Furthermore, the adjustment strategy is dynamically generated based on the prediction results of the fault prediction model and in combination with the current operating parameter data, including:

[0041] According to the following formula (1), the adjusted driving frequency is calculated :

[0042] ;

[0043] in, is the current driving frequency; A strain safety threshold predicted by the fault prediction neural network model; is the internal strain value measured at the current moment; The fatigue cumulative value predicted by the fault prediction neural network model; The fatigue safety upper limit predicted by the fault prediction neural network model; is the frequency adjustment coefficient; is the fatigue accumulation adjustment coefficient;

[0044] According to the following formula (2), the adjusted driving voltage is calculated :

[0045] ;

[0046] in, is the current driving voltage; is the front running stress; The upper limit of the safety stress predicted by the fault prediction neural network model; The fatigue cumulative value predicted by the fault prediction neural network model; The fatigue safety upper limit predicted by the fault prediction neural network model; is the stress adjustment coefficient; is the fatigue adjustment factor.

[0047] Furthermore, the human-computer interaction interface includes a real-time status display module and an abnormal alarm module, wherein the real-time status display module is used to display the real-time change curves of the driving frequency, driving voltage, output power, operating temperature and internal strain in the form of a visual chart; the abnormal alarm module uses multi-level color coding and sound alarms to prompt the user when a high-risk abnormality occurs based on the fault risk assessment in the prediction results, and provides a detailed description of the specific abnormality type and its impact range.

[0048] Furthermore, the human-computer interaction interface includes a custom analysis module, through which the user can select key operating parameters and adjustment records within a specific time period for playback analysis; the custom analysis module supports interactive switching of historical trend charts and adjustment strategy execution effects of different abnormality types, including dynamic changes in fatigue accumulation values ​​and strain safety thresholds, to help users evaluate the long-term effectiveness of adjustment strategies.

[0049] The beneficial effects of the technical solution provided by this application include: (1) This method introduces a pre-trained fault prediction neural network model, based on key operating parameters, internal strain parameters and load change data, to achieve real-time assessment of ultrasonic motor fault risks and accurate prediction of potential abnormal types, thereby effectively improving the reliability and safety of motor operation. (2) According to the fault prediction results and current operating parameters, an adjustment strategy is dynamically generated, which can adjust the drive frequency and drive voltage in real time to reduce internal strain accumulation and reduce operating stress, ensuring that the motor always operates efficiently within a safe range and prolonging the service life of the equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a flow chart of a parameter control method of an ultrasonic motor provided in the first embodiment of the present application. DETAILED DESCRIPTION

[0051] Many specific details are described in the following description to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of the present application, so the present application is not limited by the specific implementation disclosed below.

[0052] The first embodiment of the present application provides a method for controlling parameters of an ultrasonic motor. Figure 1 , which is a schematic diagram of the first embodiment of the present application. Figure 1 A parameter control method for an ultrasonic motor is described in detail in the first embodiment of the present application.

[0053] Step S101: collecting key operating parameters of the ultrasonic motor through sensors, wherein the key operating parameters include driving frequency, driving voltage, output power and operating temperature; collecting internal strain parameters of the motor structure and motor load change data.

[0054] In step S101, the key operating parameters and other relevant data of the ultrasonic motor are collected by sensors to provide basic information for subsequent fault prediction and dynamic adjustment strategies. The following is a detailed implementation method of step S101:

[0055] First, based on the operating status of the ultrasonic motor, select sensors that can accurately collect the driving frequency, driving voltage, output power and operating temperature. These sensors can include the following specific types:

[0056] Drive frequency sensor: Use the frequency meter built into the motor control system to detect changes in the motor drive frequency in real time and record its current value. The frequency meter achieves accurate frequency measurement by synchronizing with the signal of the drive circuit.

[0057] Drive voltage sensor: Select a Hall effect voltage sensor or a high-precision voltage divider circuit to collect the real-time change value of the drive voltage and digitize it into data that can be processed by the motor control module.

[0058] Output power sensor: The instantaneous value of the motor output power is calculated through a current transformer combined with a voltage sensor, and the power data is converted into a high-precision value using a signal amplification and analog-to-digital conversion module.

[0059] Operating temperature sensor: Use a thermistor or infrared thermometer installed on the motor housing or inside to measure the operating temperature and obtain information about the motor's heat dissipation performance and working environment changes.

[0060] At the same time, the internal strain parameters of the motor structure and the motor load change data are collected. The internal strain parameters are collected by strain gauge sensors attached to key structural parts of the motor (such as drive components or fixed frames). These strain gauge sensors convert the deformation of the structure after force into electrical signals based on the principle of resistance change, and obtain stable strain data through signal conditioning circuits (such as amplification and filtering modules). For load change data, real-time monitoring can be carried out through vibration sensors (such as piezoelectric vibration sensors) and torque sensors. Vibration sensors are used to detect the vibration characteristics of the motor housing or mounting platform, while torque sensors are used to collect load changes on the motor output shaft. All data collected by sensors are centrally processed through the data acquisition module.

[0061] After data collection is completed, all types of data need to be preprocessed to improve accuracy and consistency. Specifically, for signals with significant high-frequency noise (such as vibration and strain data), a piecewise linear denoising method can be used to remove background noise through filters. At the same time, the dynamic time warping algorithm is used to time align multi-source data to ensure the synchronization of drive parameters, strain parameters, and load data in the time dimension. In addition, all signals are normalized to unify the data scale for subsequent model analysis and processing.

[0062] Finally, the preprocessed key operating parameters, internal strain parameters and motor load change data are stored in a temporary buffer in a unified data format for use by subsequent fault prediction models and adjustment strategy modules. This detailed data collection and processing process can ensure the accuracy, timeliness and integrity of the collected data, thus providing a reliable foundation for the entire parameter control method.

[0063] Step S102: inputting the key operating parameters, internal strain parameters and motor load change data into a pre-trained fault prediction neural network model to obtain a prediction result of the current operating state of the ultrasonic motor, wherein the prediction result includes a fault risk assessment and a potential abnormality type.

[0064] In step S102, the key operating parameters, internal strain parameters and motor load change data collected in step S101 are input into the pre-trained fault prediction neural network model to obtain the prediction result of the current operating state of the ultrasonic motor. In order to achieve this function, it is necessary to ensure the formatting of data input and the efficient operation of the model.

[0065] First, the key operating parameters (such as drive frequency, drive voltage, output power and operating temperature), internal strain parameters and motor load change data received from the data acquisition module need to be preprocessed to ensure the uniformity of data input and match the input requirements of the model. The preprocessing step includes data normalization, mapping all parameter values ​​to the interval of 0 to 1 to eliminate the dimensional differences between parameter values. At the same time, the time series filling method is used to interpolate the missing values ​​in the data to fill the short interruptions that may exist during the acquisition process.

[0066] Secondly, the processed data is organized according to the input format of the neural network model. The input layer of the fault prediction neural network model is designed as a multi-dimensional feature input, and each dimension of the feature corresponds to an acquisition parameter (such as frequency, strain, etc.). The input data is divided into time series segments of fixed length to form a tensor form that can be directly processed by the model. This time series segment is generated under the sliding window mechanism to provide continuous time series information and ensure that the model can capture the dynamic changes in the operating status.

[0067] The sorted data is input into the pre-trained neural network model. The structure of the model includes an input layer, several hidden layers, and an output layer. After the input layer receives the multi-dimensional feature sequence, it performs preliminary feature fusion through the first layer of fully connected network. Subsequently, the recursive neural network module in the hidden layer (such as the long short-term memory network LSTM or the gated recurrent unit GRU) captures the long-term dependencies and complex dynamic patterns in the time series. The feature representation after passing through the recursive network will enter the output layer, which consists of several neurons, each of which corresponds to a prediction result of a fault risk or potential abnormality type.

[0068] The prediction results include two parts: one is the risk assessment of each fault, which represents the probability of different faults in the form of probability distribution; the other is the classification result of potential abnormal types, indicating the specific problems that may exist in the current operating state (such as overload, resonance, overtemperature, etc.). These prediction results are output in the form of standardized numerical values ​​and stored in the buffer area for use in subsequent steps.

[0069] The operation of the neural network model relies on pre-trained weights and parameters, which are obtained by offline training using a large amount of historical operation data. The supervised learning method is used in the model training process, with the labeled operation data and the corresponding fault results as training samples, combined with optimization algorithms (such as Adam or RMSprop) to gradually adjust the weights until the model has high prediction accuracy and generalization ability.

[0070] In the above manner, step S102 ensures the efficient input of multi-source data and the accuracy of the model prediction results, providing a reliable basis for the generation of subsequent dynamic adjustment strategies.

[0071] Furthermore, the fault prediction neural network model includes a data preprocessing module, a feature extraction module, a fault classification module and a risk assessment module;

[0072] The data preprocessing module is used to receive the key operating parameters, internal strain parameters and motor load change data; the piecewise linear denoising method of multi-dimensional time series signals is used to filter out the high-frequency noise in the key operating parameters; and the dynamic time warping algorithm is used to perform time series alignment on the key operating parameters, internal strain parameters and load change data to ensure the consistency and synchronization of multi-source data and generate high-quality data after denoising and time series alignment;

[0073] The feature extraction module is used to receive the high-quality data provided by the data preprocessing module; use a convolutional neural network to extract local features in the high-quality data; use a long short-term memory network to capture the long-term dependencies in the high-quality data; fuse the local features and the long-term dependencies to obtain a multi-dimensional feature vector; wherein the multi-dimensional feature vector includes the statistical characteristics of the operating parameters, the strain accumulation trend characteristics and the load change rate characteristics;

[0074] The fault classification module is used to receive the multidimensional feature vector provided by the feature extraction module, and use a classifier composed of a support vector machine and a multi-layer perceptron to process the received data to obtain the potential abnormality type in the current operating state;

[0075] The risk assessment module is used to receive the abnormal type and historical operating parameters provided by the fault classification module; based on the received data, the probability of fault occurrence is calculated through a Bayesian network to obtain the probability distribution of each fault type; using a fuzzy logic model to input the probability of occurrence of the fault type and the abnormal impact characteristics extracted from the historical data, the impact range and severity of the abnormality on the motor operation state are calculated; combining the abnormal type, the probability of fault occurrence and its impact range, the strain safety threshold, safety stress upper limit, fatigue safety upper limit and fatigue cumulative value corresponding to each fault are obtained; and a comprehensive assessment result including the probability of fault occurrence, impact range, potential impact, and strain safety threshold, safety stress upper limit, fatigue safety upper limit and fatigue cumulative value is generated.

[0076] The fault prediction neural network model described in this embodiment includes multiple modules, namely a data preprocessing module, a feature extraction module, a fault classification module and a risk assessment module. Each module works together to achieve accurate prediction and comprehensive assessment of the operating status of the ultrasonic motor.

[0077] The data preprocessing module is responsible for receiving the key operating parameters, internal strain parameters, and motor load change data collected from the sensor. These data usually exist in the form of multi-dimensional time series signals and may be affected by the noise of the acquisition environment. In order to ensure data quality, this module uses a piecewise linear denoising method to filter out high-frequency noise in key operating parameters and eliminate random interference in the signal. At the same time, the key operating parameters, internal strain parameters, and load change data are time-aligned through a dynamic time warping algorithm to solve the problem of asynchrony during multi-source data acquisition. This process generates high-quality data after denoising and time series alignment, providing an accurate and consistent basis for subsequent feature extraction.

[0078] The feature extraction module receives the high-quality data generated by the data preprocessing module and uses a combination of convolutional neural networks (CNNs) and long short-term memory networks (LSTMs) to perform deep feature extraction on the data. First, the convolutional neural network extracts local features from the high-quality data to capture short-term patterns and changing trends in the timing signals, such as instantaneous changes in the drive frequency or voltage. Then, the long short-term memory network captures the long-term dependencies in the data and extracts deeper time series information, such as the periodicity of load changes and the trend of strain accumulation. Finally, these local features and long-term dependencies are combined through a feature fusion mechanism to form a high-dimensional feature vector. The content of the feature vector includes the statistical characteristics of the operating parameters (such as mean, variance), strain accumulation trend characteristics (such as cumulative strain rate), and load change rate characteristics, which fully describe the operating status of the motor.

[0079] The fault classification module receives the multi-dimensional feature vector generated by the feature extraction module and processes the data using a classifier that combines a support vector machine (SVM) with a multi-layer perceptron (MLP). The SVM module uses its ability to classify high-dimensional features to quickly identify major abnormal types, such as potential faults such as overload, resonance, or overheating; the MLP module further refines the preliminary classification results to improve the accuracy and granularity of the classification. The classification results include potential abnormal types and their associated characteristics under the current operating state, providing accurate input for subsequent risk assessment.

[0080] The risk assessment module receives the abnormal type and historical operating parameters provided by the fault classification module, and calculates the probability of occurrence of each fault type through the Bayesian network model. The Bayesian network uses the conditional probability relationship between historical data and real-time input to dynamically update the fault prediction results and generate the probability distribution of different fault types. At the same time, the fuzzy logic model is used to input the fault occurrence probability and the abnormal impact characteristics extracted from the historical data to further calculate the scope and severity of the impact of the abnormality on the motor operation state. Based on these evaluation results, the risk assessment module combines the abnormal type, the probability of the fault occurrence and its impact range to derive the strain safety threshold, safety stress upper limit, fatigue safety upper limit and fatigue accumulation value corresponding to each fault. These parameters are used to define the allowable range of internal strain, the maximum safe operating stress and the allowable fatigue accumulation, respectively, to ensure the safety and reliability of the motor in long-term operation.

[0081] Finally, the comprehensive assessment results generated by the risk assessment module include the probability of failure, the impact range of abnormalities, potential impacts, strain safety thresholds, safety stress upper limits, fatigue safety upper limits, and fatigue accumulation values. These results can be used to dynamically adjust the motor operating parameters and provide a scientific basis for subsequent maintenance and optimization, thereby comprehensively improving the operating efficiency and reliability of ultrasonic motors.

[0082] The strain safety threshold is the maximum internal strain range allowed in the motor structure during operation, which is used to avoid structural failure due to excessive deformation. When deriving this value, it is first necessary to identify the strain-related abnormal types through the fault prediction model. For example, in the case of overload or resonance, the structure may be subjected to large dynamic strain. By analyzing the historical operation data, the strain characteristics corresponding to these abnormal types are extracted, including the average strain value and the fluctuation amplitude of the strain. These data can be obtained by statistical methods. For example, for a motor that has been run many times, the maximum strain value under each overload condition is recorded and used as the upper limit of the typical strain. Combined with the probability of abnormal occurrence in the fault prediction model, the strain safety threshold needs to be set lower for high-risk abnormal types, so as to provide a greater safety margin in actual operation. For example, if historical data shows that the typical value of strain in the resonant state is in the range of 400 to 450 microstrains, and the probability of resonance is high, the system may set the safety threshold to 420 microstrains to ensure long-term operation safety.

[0083] The upper limit of the safety stress is the maximum stress that the motor material can withstand under the current operating conditions. First, according to the real-time measured internal strain value, it is converted into the actual stress in the current operation through the Young's modulus of the material. For example, for a certain material, its Young's modulus is 200 GPa, when the measured strain is 400 microstrain, the calculated actual stress is 80MPa. Combined with the abnormal type and impact range of the fault prediction model, it can be judged whether these stresses are close to the yield point or fatigue limit of the material. If the yield stress of the material is 120 MPa and the current operating stress is close to 100 MPa, it is necessary to reduce the upper limit of the safety stress under the operating conditions to avoid premature failure of the material due to long-term high-stress operation. In addition, the upper limit of the safety stress is adjusted according to historical data and risk assessment results. If the risk corresponding to a certain fault type is high, for example, the probability of occurrence is 30%, the system may reduce the upper limit of the safety stress to 70% of the material yield point (i.e., 84 MPa) to enhance safety.

[0084] The fatigue safety upper limit defines the maximum amount of fatigue damage that the motor structure can withstand over multiple operating cycles. When deriving, the fatigue life at different stress levels needs to be analyzed from historical data to help users evaluate the long-term effectiveness of the adjustment strategy. For example, at a certain stress level, the material may be able to withstand 10^6 cycles without failure. When combined with the fault prediction model to analyze the anomaly type, the system can identify the impact of different fault types on fatigue life. If the stress level caused by a certain fault type is high, the fatigue life will be significantly reduced. Under actual operating conditions, the current stress level of the motor and its long-term cumulative impact on fatigue are evaluated. For example, for low-stress operation, the allowable fatigue safety upper limit may be close to 90% of the theoretical fatigue life of the material; while under higher stress, the upper limit may be reduced to 50% of the theoretical life to reduce the risk of long-term material failure. Based on the current fault risk assessment results, the system can further dynamically adjust the value of the fatigue safety upper limit. For example, when the anomaly risk is high, the allowable fatigue safety upper limit may be reduced by 10%.

[0085] The fatigue accumulation value is used to measure the degree of fatigue damage that the motor has suffered under the current operating state. This value is obtained by real-time analysis of the stress changes and number of cycles experienced by the motor. For example, the system records the stress level of each period of time when the motor is running in real time, and counts the number of cycles at each stress level. Combined with the fatigue life curve in the historical data, such as the maximum number of cycles that the material can withstand under a specific stress level, the system can calculate the cumulative contribution of the motor's current operation to fatigue. If the motor's stress level is 80 MPa under the current operating conditions and runs 500 cycles per hour, and the historical data shows that the maximum number of cycles allowed for the material at this stress level is 10^6, then the current fatigue accumulation contribution per hour is approximately 0.05% of the total allowable value. The system dynamically updates this value and monitors its changing trend in real time. If the fault prediction model indicates that a certain type of abnormality has a high risk, the system will pay special attention to the growth rate of the fatigue accumulation value and adopt a more stringent operation parameter adjustment strategy to ensure that the motor's cumulative fatigue does not approach the fatigue safety upper limit.

[0086] Furthermore, the piecewise linear denoising method of the data preprocessing module comprises the following steps:

[0087] Dividing the key operating parameters, internal strain parameters and motor load change data into time windows of fixed length;

[0088] For the signal in each time window, a linear trend curve is fitted by calculating the mean and slope of the multidimensional data;

[0089] Subtract the fitted linear trend curve from the original signal to obtain a signal with low-frequency drift removed;

[0090] A bandpass filter is used to filter out high-frequency noise from the signal in each time window, wherein the passband of the bandpass filter is dynamically adjusted according to the operating characteristics of the ultrasonic motor;

[0091] The processing results of each time window are spliced ​​together through the signal reconstruction method to generate an overall denoised signal.

[0092] In this embodiment, the piecewise linear denoising method of the data preprocessing module is used to improve the signal quality of key operating parameters, internal strain parameters and motor load change data, and ensure the consistency and accuracy of input data of subsequent models. The entire denoising process is implemented through multiple steps, as follows.

[0093] First, the collected multi-dimensional time series signal data is divided into time windows of fixed length. The length of the time window needs to be selected according to the operating characteristics of the ultrasonic motor. For example, for a motor with an operating frequency of 28 kHz, a window containing 100 sampling points can be selected to cover a sufficient time range. The time window is divided in a sliding manner, with a fixed step size each time to ensure continuity and avoid the loss of key signal characteristics due to the boundary of the window.

[0094] In each time window, by calculating the mean and slope of the signal in the window, an optimal linear trend curve is fitted. This trend curve represents the low-frequency component of the signal and may contain slowly changing or drifting characteristics generated during the operation of the device, such as baseline offset caused by temperature changes or gradual drift of the sensor. The fitting process uses the least squares method to calculate, so that the linear trend curve can most accurately reflect the overall change trend of the signal in the time window.

[0095] Next, the fitted linear trend curve is subtracted from the original signal to remove low-frequency drift components. The subtraction operation is performed point by point to ensure that the processed signal can retain its dynamic changes, such as high-frequency vibrations or sudden changes in operating conditions. After removing low-frequency drift, the signal in each time window is more focused on describing the actual operating characteristics of the equipment, rather than being disturbed by slow environmental changes.

[0096] Subsequently, a bandpass filter is applied to the de-drifted signal of each time window to filter out high-frequency noise. The passband of the bandpass filter needs to be dynamically adjusted according to the characteristics of the ultrasonic motor. For example, for a motor with an operating frequency between 20-40kHz, the passband of the bandpass filter can be set to 18-42 kHz to cover the operating frequency range while retaining harmonics and other critical high-frequency information. The filter design adopts the finite impulse response (FIR) or infinite impulse response (IIR) method, and the specific choice depends on the requirements of real-time performance and processing accuracy.

[0097] Finally, the signal fragments processed in all time windows are spliced ​​into a complete time series signal through signal reconstruction method. During the splicing process, it is necessary to ensure that the boundary signals of adjacent windows transition smoothly to avoid the boundary effect introduced by window cutting. The overlapping area can be processed by weighted averaging to ensure the continuity and integrity of the signal on the time axis.

[0098] Through this series of processing steps, the data preprocessing module can effectively remove low-frequency drift and high-frequency noise in the signal while retaining core information related to the motor operating status.

[0099] Furthermore, the specific method of using the convolutional neural network in the feature extraction module to extract local features includes the following steps:

[0100] Dividing the high-quality data into multi-channel time-series inputs, wherein each channel corresponds to an operating parameter or a sensor signal;

[0101] According to the dynamic changes of the operating state of the ultrasonic motor, the size and step size of the convolution kernel are adjusted in real time, and the adjustment is based on the rate of change of the operating parameters and the load fluctuation amplitude, so as to optimize the capture ability of the local mode;

[0102] After the convolution operation, a dynamic weighted multi-core feature fusion strategy is adopted to integrate low-frequency information and high-frequency information into feature vectors in proportion, combined with the weight distribution of key operating parameters;

[0103] The pooled features are screened through a specific gating mechanism, and only features with high correlation with the abnormal type are retained to reduce redundancy;

[0104] The fused feature vector is bidirectionally interacted with the output of the long short-term memory network. The deep patterns and long-term dependency characteristics in the signal are extracted through a multi-layer feature fusion network to obtain a multidimensional feature vector containing the statistical characteristics of operating parameters, strain accumulation trend characteristics and load change rate characteristics.

[0105] The feature extraction module uses convolutional neural network (CNN) to extract local features. Its specific method combines the complexity and dynamic change characteristics of ultrasonic motor operation to optimize the efficiency and accuracy of signal processing and feature extraction.

[0106] First, the high-quality data generated by the data preprocessing module is divided into multi-channel time series inputs, where each channel represents an independent operating parameter or sensor signal, such as drive frequency, drive voltage, output power, internal strain or motor load change data. The data of each channel is independently input into the convolutional network for processing to ensure that the characteristics of multi-source signals can be fully captured.

[0107] In the convolution operation, the size and step size of the convolution kernel are not fixed, but are adjusted in real time according to the dynamic operating state of the ultrasonic motor. Specifically, the system determines the parameters of the convolution kernel based on the rate of change of the operating parameters and the amplitude of the load fluctuation. For example, when the operating parameters of the motor fluctuate greatly, the size and step size of the convolution kernel are increased to capture patterns in a larger local range; when the operating parameters are relatively stable, the size of the convolution kernel is reduced to focus on capturing fine change patterns. Such a dynamic adjustment mechanism improves the adaptability of feature extraction and can capture effective information in high-frequency vibration signals and low-frequency slowly varying signals at the same time.

[0108] After the convolution operation is completed, the output features of different convolution kernels are integrated through a dynamically weighted multi-core feature fusion strategy. The system first assigns weights to each operating parameter, which are determined by the importance analysis of the parameters in the historical data. For example, in overload conditions, internal strain parameters may have higher weights, while the importance of operating temperature parameters increases when the temperature changes drastically. During the fusion process, low-frequency information and high-frequency information are proportionally integrated into a feature vector to ensure the complete expression of the full-band signal characteristics.

[0109] After feature fusion is completed, the system applies a specific gating mechanism to the features after the pooling operation for screening. This mechanism dynamically screens out features with high correlation with the fault type by analyzing the correlation between the features and the known abnormality types, while eliminating redundant or irrelevant features. For example, when a resonance-type abnormality is identified, features related to frequency characteristics are retained first, while features unrelated to the load change rate are suppressed. This step can significantly reduce the feature dimension and improve the efficiency of subsequent model processing.

[0110] Finally, the fused feature vector is bidirectionally interacted with the output of the long short-term memory network (LSTM). This process further extracts deep patterns and long-term dependency characteristics in the signal through a multi-layer feature fusion network. For example, LSTM captures the cumulative strain trend in the time series, while CNN captures the characteristics under the instantaneous operating state. The bidirectional interaction enables the complementary fusion of the characteristics of the two to generate the final multi-dimensional feature vector, which contains the statistical characteristics of the operating parameters, the characteristics of the cumulative strain trend, and the characteristics of the load change rate.

[0111] Through the above method, the feature extraction module can dynamically adapt to the data characteristics under the complex operating conditions of the ultrasonic motor, ensure that the extracted features accurately describe the motor operating status, and provide high-quality input for the fault classification module.

[0112] Furthermore, the fault classification module further includes a dynamic feature weight allocation mechanism, specifically including:

[0113] Based on the correlation between the features of each dimension in the multidimensional feature vector and the known fault types in the historical operation data, the feature weights are dynamically adjusted; wherein the correlation is calculated by a feature importance analysis model, and the feature importance analysis model is based on random forest or gradient boosting decision tree training to quantify the contribution of each feature to the classification accuracy.

[0114] The fault classification module further optimizes the accuracy and robustness of classification by introducing a dynamic feature weight allocation mechanism. This mechanism can dynamically adjust the weight of each feature based on the correlation between the multi-dimensional feature vector and the known fault types in the historical operation data, thereby enhancing the classifier's responsiveness to key features.

[0115] First, a multi-dimensional feature vector is generated by the feature extraction module, where each dimension of the feature represents a specific statistical characteristic of the ultrasonic motor operating parameters, strain accumulation trend, or load change rate. In the fault classification process, in order to improve the accuracy of the classifier, it is necessary to evaluate the correlation between each feature and different fault types. The correlation analysis is achieved through a feature importance analysis model, which is trained based on random forests or gradient boosted decision trees. These algorithms can effectively quantify the contribution of each feature to the target variable (i.e., fault type) by iteratively partitioning the data space.

[0116] During the training process, feature importance is calculated using specific metrics. For example, a random forest-based model can use average information gain or Gini index reduction to represent feature importance. Gradient boosted decision trees quantify feature importance by accumulating the split contribution of each tree. By analyzing the classification contribution of different features in historical running data, an initial weight can be assigned to each feature.

[0117] In actual operation, these initial weights are dynamically adjusted according to the real-time operating status. For example, when the values ​​of certain features in the multidimensional feature vector change significantly (such as fluctuations in the drive frequency or an increase in the strain accumulation rate), the system will re-evaluate the weights of these features to make them more influential in the classification process. The feature weight adjustment also takes into account specific feature patterns of known fault types. For example, if the current classification task involves resonance-type faults, features related to the drive frequency characteristics will be given higher weights, while the weights of features that are not related to the load change rate will be reduced.

[0118] Dynamically adjusted feature weights directly affect the behavior of the classifier. For support vector machines, feature weights are used to adjust kernel function parameters so that the classifier has higher sensitivity on high-weight features. For multi-layer perceptrons, feature weights enhance the contribution of highly correlated features to the classification results by changing the weight strength of hidden layer connections.

[0119] Through this dynamic feature weight allocation mechanism, the fault classification module can respond to changes in the operating status of the ultrasonic motor in real time, while adaptively optimizing the classification process for different fault types. This mechanism ensures high accuracy and robustness of the classification results, providing reliable basic data for the subsequent risk assessment module.

[0120] Furthermore, the risk assessment module further includes a dynamic weighting mechanism for abnormal impact, specifically including:

[0121] According to the abnormal type output by the fault classification module and the motor operating conditions extracted from the historical operating data, the influence weight of the fault type is dynamically adjusted; wherein, the calculation of the influence weight is based on a fuzzy logic model, which combines the real-time fault probability, the abnormal impact range, the historical abnormal frequency and the fluctuation range of key operating parameters to generate a comprehensive weight value through dynamic adjustment rules; the comprehensive weight value is used to correct the probability distribution of each fault type node in the Bayesian network, and affects the finally generated strain safety threshold, safety stress upper limit, fatigue safety upper limit and fatigue accumulation value, so as to ensure the adaptability and accuracy of the evaluation results under different operating conditions.

[0122] The risk assessment module in this embodiment significantly improves the adaptability and accuracy of the assessment results under complex operating conditions by introducing a dynamic weighting mechanism for abnormal impact. This mechanism uses a fuzzy logic model to dynamically adjust the impact weight of each fault type based on the abnormal type output by the fault classification module and the motor operating conditions extracted from the historical operating data.

[0123] During operation, the risk assessment module first receives the abnormal type information provided by the fault classification module, such as overload, overtemperature or resonance, etc. These abnormal types correspond to different operating characteristics and potential impacts, and are combined with the motor operating conditions recorded in the historical operating data, such as the long-term strain accumulation trend, load fluctuations, and the rate of change of operating parameters, to provide basic data for weight calculation.

[0124] The calculation of impact weights is based on a fuzzy logic model. The model uses a series of fuzzy rules to comprehensively consider the real-time probability of fault occurrence, the impact range of abnormalities, the historical abnormal frequency, and the fluctuation range of key operating parameters. For example, if the probability of a certain fault type is high and its impact range is wide, the weight will be dynamically adjusted to a higher value to increase the priority of this fault in the risk assessment. In addition, for abnormal types with a high historical frequency, such as frequent minor overloads, the weight will also be appropriately increased, but lower than the fault types with greater potential harm. The fluctuation range of key operating parameters is another important factor. For example, when the rate of change of drive frequency or internal strain exceeds the normal range, the weight of the fault type related to these parameters will be significantly increased.

[0125] The fuzzy logic model defines a fuzzy set of input variables (such as probability of occurrence, impact range, etc.) and maps each input variable to a fuzzy value, such as "high", "medium", and "low". Then, through fuzzy rules, such as "if the probability of a fault is high and the impact range is large, the weight is high", multiple input variables are comprehensively reasoned. Finally, a defuzzification method is used to generate an accurate comprehensive weight value, which is used to dynamically adjust the node probability distribution of each fault type in the Bayesian network.

[0126] The modified Bayesian network updates the probability of occurrence of each fault type using the comprehensive weight value. These probabilities directly affect the key parameters that are ultimately generated, including strain safety threshold, safety stress upper limit, fatigue safety upper limit, and fatigue accumulation value. For example, when the weight of a certain fault type increases, the strain safety threshold associated with it may be lowered to provide a greater safety margin; the corresponding fatigue safety upper limit may also be lowered to reduce fatigue risks in long-term operation.

[0127] Through this dynamic weighting mechanism, the risk assessment module can adapt to changing operating conditions under different operating conditions, and make intelligent adjustments to real-time and historical data to ensure the scientificity and rationality of the assessment results.

[0128] The following is the reference implementation code of the fault prediction neural network model:

[0129] import numpy as np

[0130] from scipy.signal import butter, filtfilt

[0131] from sklearn.ensemble import RandomForestClassifier

[0132] from sklearn.metrics import accuracy_score

[0133] from sklearn.tree import DecisionTreeClassifier

[0134] import tensorflow as tf

[0135] from tensorflow.keras.models import Model

[0136] from tensorflow.keras.layers import Input, Conv1D, MaxPooling1D,Flatten, Dense, LSTM, concatenate

[0137] # Data preprocessing module

[0138] class DataPreprocessing:

[0139] def __init__(self, sample_rate, lowcut, highcut):

[0140] # Initialize bandpass filter parameters

[0141] self.sample_rate = sample_rate

[0142] self.lowcut = lowcut

[0143] self.highcut = highcut

[0144] def butter_bandpass_filter(self, data, lowcut, highcut, sample_rate):

[0145] # Implementation of a bandpass filter to filter out high frequency noise

[0146] nyquist = 0.5 * sample_rate

[0147] low = lowcut / nyquist

[0148] high = highcut / nyquist

[0149] b, a = butter(4, [low, high], btype='band')

[0150] return filtfilt(b, a, data)

[0151] def linear_detrend(self, data):

[0152] # Calculate the linear trend of the signal and remove low-frequency drift

[0153] length = len(data)

[0154] x = np.arange(length)

[0155] slope, intercept = np.polyfit(x, data, 1)

[0156] trend = slope * x + intercept

[0157] return data - trend

[0158] def preprocess(self, data, window_size):

[0159] # Segment the data by time window

[0160] processed_data = []

[0161] for i in range(0, len(data) - window_size, window_size):

[0162] segment = data[i:i + window_size]

[0163] detrended_segment = self.linear_detrend(segment)

[0164] filtered_segment = self.butter_bandpass_filter(

[0165] detrended_segment, self.lowcut, self.highcut, self.sample_rate)

[0166] processed_data.append(filtered_segment)

[0167] return np.concatenate(processed_data)

[0168] # Feature extraction module

[0169] class FeatureExtraction:

[0170] def __init__(self, input_shape, num_channels):

[0171] # Initialize convolution and LSTM parameters

[0172] self.input_shape = input_shape

[0173] self.num_channels = num_channels

[0174] self.model = self.build_model()

[0175] def build_model(self):

[0176] # Define a neural network combining convolution and LSTM

[0177] inputs = Input(shape=(self.input_shape, self.num_channels))

[0178] x = Conv1D(64, kernel_size=3, activation='relu', padding='same')(inputs)

[0179] x = MaxPooling1D(pool_size=2)(x)

[0180] x = Conv1D(128, kernel_size=5, activation='relu', padding='same')(x)

[0181] x = MaxPooling1D(pool_size=2)(x)

[0182] cnn_out = Flatten()(x)

[0183] lstm_out = LSTM(64, return_sequences=False)(inputs)

[0184] combined = concatenate([cnn_out, lstm_out])

[0185] output = Dense(128, activation='relu')(combined)

[0186] output = Dense(64, activation='relu')(output)

[0187] output = Dense(10, activation='softmax')(output) # Assume there are 10 types of failures

[0188] return Model(inputs, output)

[0189] def extract_features(self, data):

[0190] # Extract features

[0191] return self.model.predict(data)

[0192] # Fault classification module

[0193] class FaultClassification:

[0194] def __init__(self):

[0195] # Initialize the classifier

[0196] self.rf_classifier = RandomForestClassifier()

[0197] self.feature_importances = None

[0198] def train(self, X_train, y_train):

[0199] # Train random forest and extract feature importance

[0200] self.rf_classifier.fit(X_train, y_train)

[0201] self.feature_importances = self.rf_classifier.feature_importances_

[0202] def predict(self, X_test):

[0203] # Fault classification prediction

[0204] return self.rf_classifier.predict(X_test)

[0205] # Risk Assessment Module

[0206] class RiskAssessment:

[0207] def __init__(self):

[0208] # Initialize the Bayesian network model to calculate the probability of failure

[0209] self.bayesian_network = BayesianNetwork([

[0210] ('FaultType', 'FaultProbability'),

[0211] ('FaultType', 'ImpactSeverity'), ])

[0213] self.inference_engine = None

[0214] def train_bayesian_network(self, data):

[0215] """

[0216] Initialize and train the Bayesian network using training data;

[0217] Assume that `data` is a record containing historical fault types, occurrence probabilities, and impact scopes;

[0218] """

[0219] # Initialize the inference engine

[0220] self.inference_engine = VariableElimination(self.bayesian_network)

[0221] def calculate_fault_probabilities(self, fault_types):

[0222] """

[0223] Receive the fault type and infer the probability distribution of the fault through the Bayesian network;

[0224] """

[0225] fault_probabilities = {}

[0226] for fault in fault_types:

[0227] # Assuming the model has been trained, the probability of each fault can be calculated

[0228] prob_dist = self.inference_engine.query(variables=['FaultProbability'], evidence={'FaultType': fault})

[0229] fault_probabilities[fault] = prob_dist.values[0] # Assume that the first probability value is taken

[0230] return fault_probabilities

[0231] def fuzzy_logic_evaluation(self, fault_probabilities, historical_data):

[0232] """

[0233] Use fuzzy logic models to calculate the extent and severity of abnormalities on the motor's operating status;

[0234] """

[0235] impact_results = {}

[0236] for fault, prob in fault_probabilities.items():

[0237] # Extract the impact features related to the fault in the historical data

[0238] historical_impact = historical_data.get(fault, {'impact_mean':0.5, 'impact_variance': 0.1})

[0239] # Fuzzy logic rules: Relationship between impact range, probability and historical data

[0240] if prob>0.7:

[0241] impact_severity = "High"

[0242] elif prob>0.3:

[0243] impact_severity = "Medium"

[0244] else:

[0245] impact_severity = "Low"

[0246] # Calculate the impact range

[0247] impact_range = prob * historical_impact['impact_mean'] + (1 -prob) * historical_impact['impact_variance']

[0248] impact_results[fault] = {

[0249] "severity": impact_severity,

[0250] "impact_range": impact_range

[0251] }

[0252] return impact_results

[0253] def calculate_safety_metrics(self, impact_results):

[0254] """

[0255] Calculate the safety parameters corresponding to each fault according to the fault impact range;

[0256] """

[0257] safety_metrics = {}

[0258] for fault, impact in impact_results.items():

[0259] # Strain safety threshold

[0260] strain_safety_threshold = max(0.8 - impact['impact_range'],0.5)

[0261] # Safety stress upper limit

[0262] stress_safety_limit = max(1.0 - impact['impact_range'], 0.6)

[0263] # Fatigue safety limit

[0264] fatigue_safety_limit = max(0.9 - impact['impact_range'], 0.7)

[0265] # Fatigue accumulation value (example logic)

[0266] fatigue_accumulated = 0.5 + impact['impact_range'] * 0.2

[0267] safety_metrics[fault] = {

[0268] "strain_threshold": strain_safety_threshold,

[0269] "stress_limit": stress_safety_limit,

[0270] "fatigue_limit": fatigue_safety_limit,

[0271] "fatigue_accumulated": fatigue_accumulated

[0272] }

[0273] return safety_metrics

[0274] def assess_risks(self, fault_types, historical_data):

[0275] """

[0276] Comprehensively generate risk assessment results, including failure probability, impact scope and safety parameters;

[0277] """

[0278] fault_probabilities = self.calculate_fault_probabilities(fault_types)

[0279] impact_results = self.fuzzy_logic_evaluation(fault_probabilities,historical_data)

[0280] safety_metrics = self.calculate_safety_metrics(impact_results)

[0281] risk_assessment = {}

[0282] for fault in fault_types:

[0283] risk_assessment[fault] = {

[0284] "probability": fault_probabilities[fault],

[0285] "impact": impact_results[fault],

[0286] "safety_metrics": safety_metrics[fault]

[0287] }

[0288] return risk_assessment

[0289] Training a fault prediction neural network model requires the following main steps: First, prepare historical data, including key operating parameters, internal strain parameters, motor load change data, and fault type labels, and complete denoising and time series alignment through the data preprocessing module to generate high-quality input.

[0290] The training of the feature extraction module combines convolutional neural network (CNN) and long short-term memory network (LSTM) to extract local features and time-dependent characteristics respectively. The preprocessed time series data is input into CNN and LSTM, and the weights are optimized using supervised learning methods to minimize the error through the loss function.

[0291] The fault classification module uses random forests and gradient boosting decision trees to train the mapping relationship between multi-dimensional feature vectors and fault types, and optimizes classification performance through feature importance analysis. The risk assessment module estimates the probability of fault occurrence by training Bayesian networks, and uses fuzzy logic models combined with historical data to adjust rule parameters, thereby generating accurate risk assessment results.

[0292] After the overall training is completed, the performance of the model is verified to ensure that it can reliably predict faults and provide risk assessment results under actual operating conditions.

[0293] Step S103: According to the prediction results of the fault prediction model and in combination with the current operating parameter data, an adjustment strategy is dynamically generated, wherein the adjustment strategy includes adjusting the driving frequency to reduce internal strain accumulation; adjusting the driving voltage to reduce operating stress, while controlling the internal strain within a safe range.

[0294] In step S103, an adjustment strategy is dynamically generated based on the prediction results of the fault prediction model combined with the current operating parameter data to ensure that the operating parameters of the ultrasonic motor can always be maintained within a safe range and optimize performance.

[0295] First, the prediction results output by step S102 are obtained, including fault risk assessment and potential abnormality types. Fault risk assessment usually expresses the possibility of each fault in the form of a probability value, while potential abnormality types provide specific operating status information, such as the degree of accumulation of internal strain, the risk of stress load exceeding the limit, or early signs of fatigue characteristics. At the same time, the key operating parameters collected in step S101 are called in real time, including the current drive frequency, drive voltage, output power and operating temperature, as well as internal strain parameters and load change data related to the motor structure.

[0296] In order to dynamically generate adjustment strategies, it is necessary to first conduct a comprehensive analysis of these data. By calculating the safety margin of the current operating parameters, the relative relationship between them and the safety threshold provided by the prediction model is evaluated. For example, the residual safety margin of the internal strain is calculated by the formula, and the current strain value is compared with the predicted strain safety threshold to determine whether the accumulation of internal strain is close to the dangerous range. Similarly, the real-time operating stress value is compared with the predicted safety stress upper limit to determine whether the stress exceeds the allowable range.

[0297] Based on these analysis results, specific adjustment strategies are generated. For the adjustment of the drive frequency, the strain accumulation degree and fatigue risk are used as weight factors to dynamically calculate the frequency adjustment amount to avoid further accumulation of internal strain. For example, when the internal strain is close to the safety threshold, the adjustment strategy of reducing the frequency is preferred to reduce the resonance effect of the structure. For the adjustment of the drive voltage, the voltage increase or decrease is determined by analyzing the current stress state and fatigue risk to balance the relationship between stress reduction and motor output performance. At the same time, ensure that the voltage adjustment does not cause the output power to exceed the load demand.

[0298] All adjustment strategies are generated through mathematical models and logical rules. For example, when adjusting the drive frequency, the dynamic adjustment formula of exponential or logarithmic function is introduced in combination with the nonlinear influence of fatigue accumulation value; and for voltage regulation, linear weight or weighted mean is used in combination with multi-dimensional input data to ensure the rationality of the adjustment range and the smoothness of control. The generated adjustment strategy is verified by the data verification module to ensure that the adjustment parameters will not have a negative impact on other operating conditions of the motor.

[0299] Finally, the dynamic adjustment strategy is output in the form of structured instructions, including specific values ​​such as the adjusted drive frequency, drive voltage, and target power. These instructions will be directly passed to the motor control module to provide an accurate basis for the next step of real-time parameter adjustment. In this way, step S103 can efficiently generate optimized adjustment strategies for different operating states to ensure the safety and performance stability of the motor under complex working conditions.

[0300] Furthermore, the adjustment strategy is dynamically generated based on the prediction results of the fault prediction model and the current operating parameter data, including:

[0301] According to the following formula (1), the adjusted driving frequency is calculated :

[0302] ;

[0303] in, is the current driving frequency, which is collected in real time by the driving circuit.

[0304] A strain safety threshold predicted by the fault prediction neural network model;

[0305] It is the internal strain value measured at the current moment, which is monitored in real time by the sensor.

[0306] The fatigue cumulative value predicted by the fault prediction neural network model;

[0307] The fatigue safety upper limit predicted by the fault prediction neural network model;

[0308] is the frequency adjustment coefficient, which determines the response sensitivity of the driving frequency to the strain. Its value is calibrated by experiments, and generally ranges from 0.1 to 0.5. For example, the appropriate value obtained in the experiment is 0.3.

[0309] is the fatigue accumulation adjustment coefficient, which reflects the weight of fatigue accumulation on frequency adjustment. Its value is determined by optimization calculation. For example, the commonly used value is 1.5.

[0310] According to the following formula (2), the adjusted driving voltage is calculated :

[0311] ;

[0312] in, is the current driving voltage, measured in real time by the driving circuit.

[0313] is the previous operating stress, which is calculated from the strain and material properties collected by the sensor.

[0314] The upper limit of the safety stress predicted by the fault prediction neural network model;

[0315] The fatigue cumulative value predicted by the fault prediction neural network model;

[0316] The fatigue safety upper limit predicted by the fault prediction neural network model;

[0317] It is the stress adjustment coefficient, which determines the adjustment strength of voltage to stress. Its value is usually between 0.1 and 0.5. For example, the calibrated value is 0.25.

[0318] Fatigue adjustment coefficient determines the compensating effect of fatigue on voltage. Its value is calibrated through experiments and the common value is 5.

[0319] Step S104: According to the generated adjustment strategy, the operating parameters of the ultrasonic motor are adjusted in real time.

[0320] In step S104, according to the adjustment strategy generated in step S103, the operating parameters of the ultrasonic motor are adjusted in real time to ensure that the operating state of the motor is within a safe range and optimize its performance.

[0321] First, specific adjustment instructions are received from the adjustment strategy module. These instructions include parameter values ​​such as target drive frequency, target drive voltage, and target output power. These parameter values ​​are dynamically generated based on the current operating status and prediction results, aiming to optimize the motor operating conditions in real time. The control system compares these target parameter values ​​with the current actual operating parameters and calculates the required adjustments, such as the increase or decrease in the drive frequency, the voltage adjustment range, and the output power adjustment.

[0322] In terms of frequency regulation, the control module achieves precise regulation of the driving frequency by adjusting the generator of the driving signal, such as a phase-locked loop circuit or a digital signal processor (DSP). The frequency adjustment signal will be sent to the motor drive circuit, and the dynamic control of the ultrasonic motor drive frequency will be achieved by changing the input frequency of the signal. During the frequency adjustment process, the system will monitor the internal strain parameters in real time to ensure that the frequency adjustment will not cause new resonance or structural strain accumulation.

[0323] For voltage regulation, the control module dynamically adjusts the input voltage of the motor through the power amplifier or voltage regulator module in the motor drive circuit. The specific implementation includes adjusting the voltage amplitude using a pulse width modulation (PWM) signal, or achieving the target voltage by changing the gain of the drive circuit. During the voltage regulation process, the system will simultaneously monitor the output power and operating stress value to ensure that the voltage adjustment can both reduce the operating stress and meet the current load requirements.

[0324] The adjustment of output power is based on the coordinated control of frequency and voltage. By comprehensively controlling the drive frequency and drive voltage, the system can accurately adjust the output power, so that the motor's power output matches the load demand, while avoiding excessive strain or fatigue accumulation caused by power overload. The power detection module embedded in the system will measure the adjusted power output in real time and compare the measured value with the target power to ensure the accuracy of the power output.

[0325] While adjusting in real time, the control system monitors the execution process of the adjustment. If an abnormality occurs during the adjustment process, such as adjustment delay or execution error caused by external interference, the system will recalculate and send the corrected adjustment instructions through the feedback mechanism to ensure that the operating parameters are always maintained within the target range. After the adjustment is completed, the system will update the internal status data, including the adjusted frequency, voltage and power values, as well as the current strain parameters and stress levels, to provide basic data for the next adjustment.

[0326] Through the above process, step S104 realizes real-time parameter control based on dynamic adjustment strategy to ensure the safety, stability and operating efficiency of the ultrasonic motor under different working conditions. This process enables the motor to quickly respond to changes in the environment and operating conditions, optimize overall performance and extend service life through precise signal control and closed-loop feedback system.

[0327] Step S105: The key operating parameters, internal strain parameters, motor load change data, prediction results and control strategies are stored in a database, and the real-time operating status, historical adjustment records and fatigue trend diagram of the motor structure are displayed to the user through a human-computer interaction interface.

[0328] In step S105, the key operating parameters, internal strain parameters, motor load change data collected by sensors, and the prediction results and adjustment strategies generated by the fault prediction model are completely recorded and stored in the database. At the same time, the real-time status, historical adjustment records and fatigue trend diagrams of the motor structure are intuitively displayed to the user through the human-computer interaction interface.

[0329] First, the system will store the key operating parameters collected in real time, including drive frequency, drive voltage, output power and operating temperature, in a structured format. These parameters will be processed by the data acquisition module through pre-processing steps (such as normalization and denoising) to generate high-quality, time-stamped operating data records. Internal strain parameters and load change data will also be stored in time series and associated with key operating parameters under the same timestamp to ensure the time synchronization of multi-source data.

[0330] The prediction results generated by the fault prediction model, including fault risk assessment and potential anomaly types, are also stored as independent data items. These results are recorded in the database in the form of distributed probabilities or classification labels and are associated with the operating parameters of the corresponding timestamps. In addition, the adjustment strategy generated by the dynamic adjustment module, including the specific values ​​of the target drive frequency, target drive voltage, and target power, is also stored in the same database as part of the data record. This storage method ensures that all relevant data can be fully tracked and provides a basis for subsequent analysis and trend prediction.

[0331] The database is designed with a relational or time series database structure, depending on the storage requirements of the system. Time series databases can efficiently store and retrieve data in continuous time periods, while relational databases are suitable for complex queries and data association analysis. To improve storage reliability, the system will perform regular data backups and use distributed storage technology to improve data security and fault tolerance.

[0332] At the same time, the system presents these stored data to the user through the human-computer interaction interface. The real-time operating status will be dynamically displayed in the form of charts and numbers, such as the current drive frequency, drive voltage, and internal strain value; the historical adjustment records are displayed in the form of tables or line graphs to help users review the adjustment strategies at different time points and their corresponding operating effects. In addition, the fatigue trend diagram of the motor structure will be generated based on historical data, and the fatigue accumulation will be intuitively displayed in the form of charts. The generation of trend diagrams usually uses curve fitting or machine learning algorithms to predict possible fatigue development trends in the future.

[0333] To enhance the user experience, the human-computer interaction interface supports a variety of data filtering and display options. Users can view data within a specific time period through the time range selector, or categorize queries based on fault type or adjustment parameters. The system interface also provides a real-time alarm function. When certain operating parameters approach the safety threshold, the interface will remind you through changes in color or icons, and display the relevant fault risk assessment of the prediction model.

[0334] Through the above process, step S105 realizes complete storage and intuitive display of data, which not only meets the needs of operation monitoring and historical analysis, but also provides users with clear operation status and trend information. This function can effectively help users discover potential problems in time and optimize the operation management of the motor.

[0335] Furthermore, the human-computer interaction interface includes a real-time status display module and an abnormal alarm module, wherein the real-time status display module is used to display the real-time change curves of the driving frequency, driving voltage, output power, operating temperature and internal strain in the form of a visual chart; the abnormal alarm module uses multi-level color coding and sound alarms to prompt the user when a high-risk abnormality occurs based on the fault risk assessment in the prediction results, and provides a detailed description of the specific abnormality type and its impact range.

[0336] The human-computer interaction interface in this embodiment provides users with intuitive monitoring of the operating status of the ultrasonic motor and an efficient fault warning mechanism through a real-time status display module and an abnormal alarm module, which significantly improves the user-friendliness and operating efficiency of the system.

[0337] The real-time status display module is responsible for presenting the key operating parameters collected, including drive frequency, drive voltage, output power, operating temperature and internal strain, to the user in a visual form. This module uses dynamic chart technology to display the real-time change curves of each parameter on the user interface at the same time, ensuring that users can understand the current status of the motor at any time. For example, for the drive frequency, the chart can use time as the horizontal axis and frequency value as the vertical axis, and refresh the display of fluctuation trends in real time. For the drive voltage and output power, an overlay chart is used to show the relationship between the two, which helps users judge the load status of the motor. The operating temperature and internal strain are displayed in the form of a line chart and a bar chart respectively, which intuitively reflects the heat dissipation and structural strain characteristics of the motor.

[0338] The abnormal alarm module provides users with fault warnings by performing real-time analysis on the risk assessment results generated by the fault prediction model. When the system detects a high-risk abnormality, such as an abnormality type of overload or resonance, the alarm module immediately activates a multi-level color coding mechanism. The color coding is divided into green, yellow, and red, representing normal status, medium risk, and high-risk abnormalities, respectively. Users can judge the current operating status of the motor at a glance. At the same time as the color alarm, the module will also issue a sound alarm. For example, when a high-risk abnormality occurs, the system emits a continuous high-frequency warning sound to ensure that users can get timely fault information even if they do not view the interface.

[0339] In addition, the abnormality alarm module provides detailed descriptions of the abnormality type and impact range. For example, when the system detects a resonance abnormality, the interface will display "Resonance abnormality, which may cause excessive strain and affect structural stability", and indicate that the impact range is "critical load components" or "drive units". This information helps users quickly understand the nature of the abnormality and possible consequences, so that they can make corresponding decisions.

[0340] To enhance the user experience, the real-time status display module and the abnormal alarm module are tightly integrated. When an abnormal alarm is triggered, the real-time status display module automatically highlights the parameter curve related to the abnormality. For example, when an overload is detected, the system will highlight the change trend of the output power and drive voltage, and mark the warning threshold with a dotted line to help users quickly locate the source of the abnormality.

[0341] Through the above design, the human-machine interaction interface realizes complex data visualization and provides efficient abnormal warning function. Users can quickly understand the motor operation status and respond without technical background.

[0342] Furthermore, the human-computer interaction interface includes a custom analysis module, through which the user can select key operating parameters and adjustment records within a specific time period for playback analysis; the custom analysis module supports interactive switching to display historical trend charts and adjustment strategy execution effects of different abnormality types, including dynamic changes in fatigue accumulation values ​​and strain safety thresholds, to help users evaluate the long-term effectiveness of adjustment strategies.

[0343] The human-computer interaction interface in this embodiment provides intuitive and flexible operation data playback and analysis functions by introducing a custom analysis module, which significantly enhances the traceability of the ultrasonic motor's operating status and the evaluation ability of the adjustment strategy. This module supports users to select key operating parameters and adjustment records for a specific time period, and interactively displays the historical trends of abnormal types and the execution effects of related strategies.

[0344] The custom analysis module allows users to freely select a time period of interest, such as the past hours, days, or longer operating time. This selection can be done through a graphical timeline or date selector, and the interface is intuitive and easy to use. After selecting a time period, the module will extract the corresponding key operating parameters from the stored historical data, including drive frequency, drive voltage, output power, operating temperature, internal strain, etc. For these parameters, the module displays them in a variety of visual forms, such as line graphs, scatter graphs, or bar graphs. Users can switch between different views to observe the changing trends of the data more comprehensively.

[0345] During the data playback analysis process, the custom analysis module places special emphasis on the dynamic trend display of the anomaly type. For example, when the user chooses to view the resonance anomaly, the system will automatically filter the historical data related to the anomaly and generate a corresponding trend chart. The trend chart not only shows the time point when the anomaly occurred, but also marks the changes in the intensity of the anomaly and the scope of influence. In addition, the module provides interactive functions, and users can hover or click on specific anomaly points to view their detailed information, such as the key operating parameter values ​​at the time, the specific description of the anomaly, and the relevant adjustment strategies.

[0346] The custom analysis module also supports displaying the execution effect of the adjustment strategy, including the dynamic changes of fatigue accumulation value and strain safety threshold. The display of fatigue accumulation value is mainly based on the cumulative curve, showing the gradual change of motor fatigue state over time. For the strain safety threshold, the system superimposes it with the real-time strain value to help users intuitively evaluate whether the adjustment strategy is effective in controlling the strain within the safe range. For example, if the adjustment strategy is effective, users can see that the fluctuation of the strain value gradually stabilizes and is always below the threshold.

[0347] To further enhance the analysis capability, the module allows users to view the adjustment strategies and their execution effects in different time periods in a comparative mode. For example, users can select two time periods for analysis at the same time, and the system will generate a comparative chart showing the growth rate of fatigue accumulation values, the fluctuation range of key operating parameters, and the distribution differences of abnormal types in the two time periods. In this way, users can gain a deeper understanding of the applicability and optimization effects of the adjustment strategies under different working conditions.

[0348] Interaction is an important feature of the custom analysis module. Users can customize the display range and details of the chart by dragging, zooming, clicking, and other operations. For example, users can zoom in on the time period when an anomaly occurred to observe the subtle change trend of related parameters; or use the filtering function to view only data related to a certain type of anomaly. This highly customizable analysis method provides users with more flexible tools to help them evaluate the motor operating status and the effectiveness of adjustment strategies from multiple angles.

[0349] Through the above design, the custom analysis module can significantly improve the user's ability to review the motor's operating history and the efficiency of evaluating the adjustment strategy. Its interactive and flexible features allow users to efficiently analyze data and optimize motor operation management without professional knowledge.

[0350] A second embodiment of the present application provides an electronic device, the electronic device comprising:

[0351] processor;

[0352] The memory is used to store a program, and when the program is read and executed by the processor, the program executes a parameter control method of an ultrasonic motor provided in the first embodiment of the present application.

[0353] The third embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, a parameter control method of an ultrasonic motor provided in the first embodiment of the present application is executed.

[0354] Although the present application is disclosed as above in the form of a preferred embodiment, it is not intended to limit the present application. Any technical personnel in this field may make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application shall be based on the scope defined by the claims of the present application.

Claims

1. A method for controlling parameters of an ultrasonic motor, characterized in that: include: The key operating parameters of the ultrasonic motor are collected by sensors, wherein the key operating parameters include driving frequency, driving voltage, output power and operating temperature; the internal strain parameters of the motor structure and the motor load change data are collected; Inputting the key operating parameters, internal strain parameters and motor load change data into a pre-trained fault prediction neural network model to obtain a prediction result of the current operating state of the ultrasonic motor, wherein the prediction result includes a fault risk assessment and a potential abnormality type; According to the prediction results of the fault prediction model, combined with the current operating parameter data, an adjustment strategy is dynamically generated, wherein the adjustment strategy includes adjusting the driving frequency to reduce the internal strain accumulation; adjusting the driving voltage to reduce the operating stress, while controlling the internal strain within a safe range; According to the generated adjustment strategy, the operating parameters of the ultrasonic motor are adjusted in real time; The key operating parameters, internal strain parameters, motor load change data, prediction results and control strategies are stored in a database, and the real-time operating status, historical adjustment records and fatigue trend diagram of the motor structure are displayed to the user through a human-computer interaction interface; The fault prediction neural network model includes a data preprocessing module, a feature extraction module, a fault classification module and a risk assessment module; The data preprocessing module is used to receive the key operating parameters, internal strain parameters and motor load change data; the piecewise linear denoising method of multi-dimensional time series signals is used to filter out the high-frequency noise in the key operating parameters; and the dynamic time warping algorithm is used to perform time series alignment on the key operating parameters, internal strain parameters and load change data to ensure the consistency and synchronization of multi-source data and generate high-quality data after denoising and time series alignment; The feature extraction module is used to receive the high-quality data provided by the data preprocessing module; use a convolutional neural network to extract local features in the high-quality data; use a long short-term memory network to capture the long-term dependencies in the high-quality data; fuse the local features and the long-term dependencies to obtain a multi-dimensional feature vector; wherein the multi-dimensional feature vector includes the statistical characteristics of the operating parameters, the strain accumulation trend characteristics and the load change rate characteristics; The fault classification module is used to receive the multidimensional feature vector provided by the feature extraction module, and use a classifier composed of a support vector machine and a multi-layer perceptron to process the received data to obtain the potential abnormality type in the current operating state; The risk assessment module is used to receive the abnormal type and historical operating parameters provided by the fault classification module; based on the received data, the probability of fault occurrence is calculated through a Bayesian network to obtain the probability distribution of each fault type; using a fuzzy logic model to input the probability of occurrence of the fault type and the abnormal impact characteristics extracted from the historical data, the impact range and severity of the abnormality on the motor operation state are calculated; combining the abnormal type, the probability of fault occurrence and its impact range, the strain safety threshold, safety stress upper limit, fatigue safety upper limit and fatigue cumulative value corresponding to each fault are obtained; and a comprehensive assessment result including the probability of fault occurrence, impact range, potential impact, and strain safety threshold, safety stress upper limit, fatigue safety upper limit and fatigue cumulative value is generated.

2. The method for controlling parameters of an ultrasonic motor according to claim 1, characterized in that: The piecewise linear denoising method of the data preprocessing module comprises the following steps: Dividing the key operating parameters, internal strain parameters and motor load change data into time windows of fixed length; For the signal in each time window, a linear trend curve is fitted by calculating the mean and slope of the multidimensional data; Subtract the fitted linear trend curve from the original signal to obtain a signal with low-frequency drift removed; A bandpass filter is used to filter out high-frequency noise from the signal in each time window, wherein the passband of the bandpass filter is dynamically adjusted according to the operating characteristics of the ultrasonic motor; The processing results of each time window are spliced ​​together through the signal reconstruction method to generate an overall denoised signal.

3. The method for controlling parameters of an ultrasonic motor according to claim 1, characterized in that: The specific method of using the convolutional neural network in the feature extraction module to extract local features includes the following steps: Dividing the high-quality data into multi-channel time-series inputs, wherein each channel corresponds to an operating parameter or a sensor signal; According to the dynamic changes of the operating state of the ultrasonic motor, the size and step size of the convolution kernel are adjusted in real time, and the adjustment is based on the rate of change of the operating parameters and the load fluctuation amplitude, so as to optimize the capture ability of the local mode; After the convolution operation, a dynamic weighted multi-core feature fusion strategy is adopted to integrate low-frequency information and high-frequency information into feature vectors in proportion, combined with the weight distribution of key operating parameters; The pooled features are screened through a specific gating mechanism, and only features with high correlation with the abnormal type are retained to reduce redundancy; The fused feature vector is bidirectionally interacted with the output of the long short-term memory network. The deep patterns and long-term dependency characteristics in the signal are extracted through a multi-layer feature fusion network to obtain a multidimensional feature vector containing the statistical characteristics of operating parameters, strain accumulation trend characteristics and load change rate characteristics.

4. The method for controlling parameters of an ultrasonic motor according to claim 1, characterized in that: The fault classification module further includes a dynamic feature weight allocation mechanism, specifically including: Based on the correlation between the features of each dimension in the multidimensional feature vector and the known fault types in the historical operation data, the feature weights are dynamically adjusted; wherein the correlation is calculated by a feature importance analysis model, and the feature importance analysis model is based on random forest or gradient boosting decision tree training to quantify the contribution of each feature to the classification accuracy.

5. The method for controlling parameters of an ultrasonic motor according to claim 1, characterized in that: The risk assessment module further includes a dynamic weighting mechanism for abnormal impact, specifically including: According to the abnormal type output by the fault classification module and the motor operating conditions extracted from the historical operating data, the influence weight of the fault type is dynamically adjusted; wherein, the calculation of the influence weight is based on a fuzzy logic model, which combines the real-time fault probability, the abnormal impact range, the historical abnormal frequency and the fluctuation range of key operating parameters to generate a comprehensive weight value through dynamic adjustment rules; the comprehensive weight value is used to correct the probability distribution of each fault type node in the Bayesian network, and affects the finally generated strain safety threshold, safety stress upper limit, fatigue safety upper limit and fatigue accumulation value, so as to ensure the adaptability and accuracy of the evaluation results under different operating conditions.

6. The method for controlling parameters of an ultrasonic motor according to claim 1, characterized in that: The adjustment strategy is dynamically generated based on the prediction results of the fault prediction model and the current operating parameter data, including: According to the following formula (1), the adjusted driving frequency is calculated : ; in, is the current driving frequency; A strain safety threshold predicted by the fault prediction neural network model; is the internal strain value measured at the current moment; The fatigue cumulative value predicted by the fault prediction neural network model; The fatigue safety upper limit predicted by the fault prediction neural network model; is the frequency adjustment coefficient; is the fatigue accumulation adjustment coefficient; According to the following formula (2), the adjusted driving voltage is calculated : ; in, is the current driving voltage; is the front running stress; The upper limit of the safety stress predicted by the fault prediction neural network model; The fatigue cumulative value predicted by the fault prediction neural network model; The fatigue safety upper limit predicted by the fault prediction neural network model; is the stress adjustment coefficient; is the fatigue adjustment factor.

7. The method for controlling parameters of an ultrasonic motor according to claim 1, characterized in that: The human-computer interaction interface includes a real-time status display module and an abnormal alarm module, wherein the real-time status display module is used to display the real-time change curves of the driving frequency, driving voltage, output power, operating temperature and internal strain in the form of a visual chart; the abnormal alarm module, based on the fault risk assessment in the prediction results, prompts the user with multi-level color coding and sound alarms when a high-risk abnormality occurs, and provides a detailed description of the specific abnormality type and its impact range.

8. The method for controlling parameters of an ultrasonic motor according to claim 1, characterized in that: The human-computer interaction interface includes a custom analysis module, through which users can select key operating parameters and adjustment records within a specific time period for playback analysis; the custom analysis module supports interactive switching to display historical trend charts of different abnormal types and adjustment strategy execution effects, including dynamic changes in fatigue accumulation values ​​and strain safety thresholds, to help users evaluate the long-term effectiveness of adjustment strategies.

Images