Intelligent prediction and maintenance system and method for ultrasonic vibrator group

Abstract

The invention discloses an ultrasonic vibrator group intelligent prediction maintenance system and method, and relates to the technical field of industrial equipment fault diagnosis and predictive maintenance, the ultrasonic vibrator group intelligent prediction maintenance system is composed of a data acquisition module, a data transmission module, a data processing module, a model analysis module, an application module, a feedback optimization module and the like, and vibration, current, voltage, resistance, temperature, power and other data are acquired in real time through a multi-source sensor; the sampling frequency is dynamically adjusted under the abnormal working condition; multi-dimensional feature fusion and fault prediction are carried out by utilizing a pre-trained Transform model, mechanical, electrical, thermal and other fault types are automatically identified, maintenance suggestions are output, a user can obtain fault early warning, residual service life and maintenance guidance through a visual interface, and an actual maintenance result is transmitted back to a system to continuously optimize the model, and compared with a traditional mode, the method has the advantages that the operation is simple, and the efficiency is high. According to the invention, accidental shutdown can be effectively reduced, the maintenance cost is reduced, the service life of equipment is prolonged, the real-time monitoring and accurate prediction of the operation state of the equipment are realized, and the production efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial equipment fault diagnosis and predictive maintenance, and in particular to an ultrasonic vibrator group intelligent predictive maintenance system and method thereof. Background Art

[0002] The specification of Chinese invention patent application CN118962513A discloses a device status judgment system and method in ultrasonic equipment based on digital twins, specifically involving digital twin technology and ultrasonic power supply technology. The patent realizes online monitoring of device status and fault warning through the following means: Use voltage and current sampling units to obtain real-time electrical signals from ultrasonic power supply equipment; The device parameters are set to initial values ​​using the equivalent circuit part in the digital twin system; Intelligent algorithms are used to find the optimal solution to the objective function and compare it with normal parameters to determine the health status of each component in real time.

[0003] This patent has certain advantages in real-time monitoring of device parameters, but because it mainly relies on equivalent circuit modeling and electrical sampling, it lacks comprehensive analysis capabilities for other physical quantities (such as vibration frequency and amplitude, temperature, environmental conditions, etc.). As the application of ultrasonic vibrators in various working conditions continues to expand, single or fewer dimensional monitoring may be difficult to timely capture complex failure modes caused by mechanical, thermal or multi-fault coupling. In addition, although this patent utilizes the concept of digital twins, it still has certain limitations in multi-source data fusion, fault identification accuracy improvement and remote intelligent maintenance guidance.

[0004] In summary, most of the existing technologies perform status judgment based on a single or a few sensor data, and lack comprehensive and fine-grained monitoring of the health status of ultrasonic vibrators in complex and multi-condition application scenarios. Especially when multiple factors such as mechanical, electrical, and thermal faults are coupled, traditional models are prone to omissions or false alarms, resulting in equipment downtime or increased maintenance costs. At the same time, many existing methods can only provide fault warnings and cannot further provide maintenance strategies and decision-making recommendations. Summary of the invention

[0005] The purpose of the present invention is to overcome the deficiencies of the prior art and to provide an ultrasonic vibrator group intelligent predictive maintenance system and method thereof to solve the above-mentioned problems.

[0006] Summarize the overall technical means and corresponding technical effects of this application in one paragraph The object of the present invention is achieved through the following technical solutions: an ultrasonic vibrator group intelligent predictive maintenance system, comprising: The data acquisition module is used to collect data related to the operating status of each ultrasonic vibrator using various sensors deployed on the vibrator, including: The vibration sensor collects the vibration frequency and amplitude; The temperature sensor collects the working temperature of the oscillator; The power sensor collects the power consumption information of the oscillator; The electrical sensor collects the current, voltage and resistance during the operation of the vibrator; Environmental sensors can optionally collect environmental parameters such as humidity and pressure; A data transmission module, used to transmit the multi-source data acquired by the data acquisition module to the data center through the industrial Internet of Things network; The data processing module is used to perform data cleaning, normalization and key feature extraction on the received multi-source data; The model analysis module is used to predict and analyze key features based on the pre-trained Transformer architecture model. The model combines a multi-head attention mechanism to focus on changes in vibration characteristics, current, resistance, voltage, temperature, power, and environmental data, and outputs the failure probability, remaining service life, and possible failure type of each vibrator. Application module, used to present fault prediction results and maintenance suggestions through the user interface, including real-time status monitoring, fault warning, maintenance guidance and historical data query; The feedback optimization module is used to obtain the actual maintenance results of maintenance personnel, and update and adjust the parameters of the pre-trained model based on the accuracy of fault diagnosis, so as to continuously improve the accuracy of fault prediction and the rationality of maintenance recommendations.

[0007] The sampling frequency of the data acquisition module is set between 10 Hz and 100 Hz according to the equipment operating characteristics and fault response requirements, and can be dynamically adjusted according to abnormal fluctuations in the oscillator.

[0008] The data processing module includes: Data cleaning unit, used to remove noise data, missing values ​​and transient outliers; Normalization units are used to convert data of different dimensions and ranges to the same standard range; The feature extraction unit is used to extract the frequency domain features, time domain features, temperature change rate, power fluctuation range, and change rate or transient response features of the vibration signal.

[0009] The multi-head attention mechanism in the model analysis module includes at least: The first attention head is used to focus on the vibration frequency and amplitude to capture the early signs of mechanical failure; A second attention head for focusing on changes in current, voltage, and resistance to identify electrical faults or abnormal energy loss; A third attention head for focusing on temperature changes to detect overheating or poor heat dissipation faults; At least one additional attention head for comprehensive power and environmental characteristics to analyze the impact of load changes and environmental factors on the operation of the oscillator; Among them, the weights of each attention head can be dynamically adjusted automatically according to the real-time operating conditions.

[0010] During the fine-tuning process, the pre-trained model is trained specifically by combining the historical current, resistance, voltage, vibration, and temperature data of the ultrasonic oscillator, so as to improve the recognition and prediction capabilities for multiple types of fault modes.

[0011] The application module includes: A fault warning sub-module for automatically sending an alarm to maintenance personnel when the predicted fault probability is higher than a preset threshold; A maintenance suggestion sub-module for providing corresponding maintenance guidance plans according to the fault type, remaining service life, and abnormal electrical / mechanical parameter conditions, including replacing parts in advance, repairing electrical circuits, or adjusting process parameters, etc.; A historical data query sub-module for visually displaying historical fault records and operating parameter curves to facilitate maintenance personnel to analyze the reasons and trends of abnormalities.

[0012] An intelligent predictive maintenance method for an ultrasonic oscillator group includes the following steps: S1. Data collection: Real-time obtain data such as vibration, temperature, power, current, voltage, and resistance through various sensors deployed on the ultrasonic oscillator, and adjust the sampling frequency according to the working condition requirements; S2. Data transmission: Transmit the collected data to the data center through the industrial Internet of Things network; S3. Data preprocessing: Denoise, normalize, and extract various features from the received data, including vibration features, power fluctuation features, abnormal electrical parameter features, and environmental parameter features; S4. Model analysis: Input the features into the pre-trained Transformer model, and obtain the fault probability, remaining service life, and fault type of each oscillator through the multi-head attention mechanism; S5. Prediction results and maintenance suggestions: Set a warning threshold according to the prediction results. When the fault probability exceeds the threshold, push an alarm to maintenance personnel, and give maintenance suggestions for possible fault parts or types; S6. Feedback and model update: After the fault is processed, collect the actual maintenance results, compare the prediction results with the actual situation, and update the pre-trained model and system parameters regularly or in real time to improve the overall prediction accuracy and maintenance decision-making level.

[0013] When collecting current, voltage, and resistance in step S1, further monitor the transient electrical waveform changes. When the current or voltage shows a sharp fluctuation and the duration exceeds the preset threshold, automatically increase the sampling frequency of the sensor.

[0014] In step S4, the multi-head attention mechanism automatically adjusts the weight distribution of each attention head according to the real-time operating environment of the oscillator, including increasing the attention to temperature, current, and voltage characteristics in high-load or high-temperature situations, and increasing the attention to vibration characteristics when the vibration frequency is abnormal.

[0015] The maintenance suggestions in step S5 include the detection of electrical circuits, the fastening of mechanical structures, the overhaul of the cooling system, the replacement or upgrade of key components, and specific guidance on adjusting the production process or operating environment in case of abnormal power consumption.

[0016] The beneficial effects of the present invention are: 1. Through the collaborative monitoring of multiple sensors (mechanical, electrical, thermal, environmental, etc.) and the intelligent analysis based on the multi-head attention mechanism, the present invention can capture potential fault signs of the ultrasonic oscillator at an early stage and give early warnings in a timely manner, thus significantly reducing the downtime risk caused by sudden failures and ensuring the continuous and stable operation of the production line.

[0017] 2. Compared with the traditional regular maintenance and after-failure repair methods, the present invention adopts a predictive maintenance mode, which can carry out targeted repairs or preventive measures before the equipment actually has serious failures, greatly reducing unnecessary component replacements and over-maintenance, delaying the aging and wear of the equipment, and extending the overall service life of the ultrasonic oscillator.

[0018] 3. By performing multi-modal data fusion (vibration, current, voltage, resistance, temperature, power, etc.) on the pre-trained Transformer model and using the dynamically adjusted multi-head attention mechanism, the present invention can adaptively capture key features under different working conditions, improve the recognition accuracy of multiple types of faults (mechanical, electrical, thermal, etc.); at the same time, give differentiated maintenance suggestions, providing more accurate decision-making support for on-site operation and maintenance personnel.

[0019] 4. Establish a closed loop between the model prediction results and the actual maintenance feedback. By continuously collecting and updating the fault and maintenance result data and fine-tuning the model regularly or in real time, the present invention can continuously iterate and optimize the fault detection and analysis strategies over time to adapt to the changing production environment and equipment status.

[0020] 5. The present invention has set up intuitive warning, maintenance advice, historical data query and other functional interfaces in the application layer, enabling operation and maintenance personnel to conveniently view the current device status, understand the historical fault development trend and take prompt actions. This human-computer interaction design not only improves the usability of the system, but also facilitates the accumulation of maintenance knowledge and experience for front-line personnel. Brief Description of the Drawings

[0021] Figure 1 is the system architecture diagram of the present invention; Figure 2 is the working flow chart of the present invention. Detailed Embodiments

[0022] Next, the technical solutions of the present invention will be clearly and completely described in conjunction with the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative efforts shall fall within the protection scope of the present invention.

[0023] It should be noted that in the following solutions, the orientation concepts of "left", "right", "up", "down", "front", "back", "inside", and "outside" are all relative directions, and will not be listed one by one here.

[0024] Embodiment 1: As Figure 1 and Figure 2 shown, in this embodiment, in order to achieve intelligent predictive maintenance of the ultrasonic oscillator group, the overall system is hierarchically designed according to the process of "data acquisition - data transmission - data processing - model analysis - application and feedback", that is: data acquisition module, data transmission module, data processing module, model analysis module, application module, feedback optimization module The basic working process of the entire system can be briefly described as: The data acquisition module is deployed on the on-site ultrasonic oscillators to obtain multi-source data such as vibration, electricity, temperature, power, and environment in real time; The data transmission module sends the collected various data to the data center or cloud through the industrial Internet of Things network; The data processing module cleans, normalizes, and performs preliminary feature extraction on the received raw data in the data center; The model analysis module uses a pre-trained Transformer model to perform fault prediction on these features, and obtains the fault probability, remaining service life, and possible fault types of each oscillator; The application module presents these prediction results and maintenance advice to the operation and maintenance personnel through the user interface, and triggers a fault warning when necessary; The feedback optimization module updates and iterates the pre-trained model according to the actual fault situation and maintenance feedback in the later stage to continuously improve the prediction accuracy and the rationality of maintenance suggestions.

[0025] In practical applications, these functional modules can be either centrally deployed on the same hardware platform or dispersed according to requirements, and they cooperate through network interconnection to achieve an efficient and flexible system structure.

[0026] The data acquisition module makes full use of a variety of sensors to comprehensively monitor the operating status of the ultrasonic oscillator (corresponding to the sensor arrangements in claim 1). In this embodiment, it specifically includes: A vibration sensor, mainly used to monitor the vibration frequency and amplitude of the ultrasonic oscillator. Deviations in vibration frequency and abnormal changes in amplitude usually can reflect problems such as loosening, wear, or failure of mechanical components. A temperature sensor, used to monitor the temperature status of the oscillator and its surrounding environment in real time. When the temperature exceeds the limit or rises rapidly, it often implies potential hazards inside the oscillator or in the surrounding environment, such as overheating, cooling system failure, or harsh environment, etc. A power sensor, which collects the power consumption information of the oscillator during operation and conducts corresponding diagnoses based on the fluctuations of the electric power. If the power fluctuates significantly under the same working conditions, there may be abnormal energy loss or unstable load, etc. An electrical sensor, including the collection of current, resistance, and voltage, used to comprehensively monitor the electrical characteristics of the oscillator under different working loads. For example, when the voltage deviates from the normal range, the current is too large, or the resistance changes drastically, it may indicate risks such as electrical system failure, short circuit, poor contact, or insulation deterioration. An environmental sensor (optional). In some special working environments, sensors such as humidity, air pressure, and dust concentration can be additionally deployed to better understand the impact of the equipment's environment on the operation of the oscillator.

[0027] Sampling frequency and dynamic adjustment. The sampling frequency of this system is basically controlled within the range of 10Hz to 100Hz. For the following considerations, this range can capture the characteristics of most potential faults without overly consuming system resources: Under normal circumstances, the sampling frequency can be default set to a medium frequency (such as 30Hz or 50Hz) to balance the data volume and monitoring accuracy. When the system detects obvious abnormal fluctuations, such as instantaneous mutations in current or voltage, sudden increases in vibration peaks, etc., the sampling frequency can be temporarily increased to 100Hz through software instructions to obtain more detailed fault information.

[0028] Meanwhile, the sampling strategy can also be configured accordingly according to different types of oscillators or different working modes (such as no-load, full-load, high-temperature environment, etc.) to implement a flexible and hierarchical data acquisition mechanism.

[0029] Basic process of the data processing module This embodiment further introduces a series of steps executed by the data processing module after receiving the raw data to ensure the data quality and feature completeness required for subsequent model analysis, mainly including: Data cleaning. By setting reasonable anomaly detection rules, noise data, missing values, and instantaneous outliers are removed or corrected. For example, when the sensor malfunctions temporarily or is affected by external electromagnetic interference, the extreme values or invalid values collected will be determined as abnormal data and removed; For some local small-range fluctuations, a moving average filter or other signal smoothing algorithms can be used for processing to enhance the stability of subsequent analysis.

[0030] Normalization processing. Since there are various physical quantities such as vibration, temperature, current, voltage, and resistance in the system, and their original numerical ranges vary greatly, if normalization processing is not performed, it may cause subsequent algorithms to be overly sensitive or ignore certain dimensions; Therefore, in this embodiment, data of different dimensions are usually mapped to the same standard interval (such as [0,1]) or the Zscore standardization method is used to make the subsequent model inputs maintain a consistent order of magnitude.

[0031] Feature extraction. Time-domain features: including the average value, variance, peak factor, etc. of the vibration signal; frequency-domain features: performing Fourier transform (FFT) on the vibration signal to extract spectral energy distribution, main frequency, harmonic components, etc.; electrical features: such as the transient change rate, mean shift of current, voltage, and resistance; temperature change rate: observing the heating curve and cooling curve features of the oscillator temperature in combination with time series information; power fluctuation range: collecting the average power per unit time and conducting statistical analysis in combination with the instantaneous peak value. The processed data above will be marked and stored in a database or memory buffer for subsequent model analysis modules to call. By classifying and labeling the above features, the system can quickly retrieve and transfer relevant data to the model analysis module to achieve more efficient and accurate fault prediction.

[0032] Diversity of sensor arrangement: It lays a foundation for the system to obtain multi-dimensional information including mechanical, thermal, power, electrical, and environmental aspects; Necessity of data processing: The data cleaning, normalization, and feature extraction links can effectively improve the accuracy and stability of subsequent model analysis; Flexible sampling strategy: The basic sampling frequency range from 10Hz to 100Hz, as well as the dynamic boost during abnormal moments, ensures the ability to capture short-term or sudden faults.

[0033] Thus, it provides a solid data and technical foundation for the subsequent model analysis module (fault prediction), application module (early warning and maintenance suggestions), and feedback optimization module (continuous learning and improvement).

[0034] This embodiment is carried out at the hardware and basic data processing levels, providing indispensable support for the intelligent prediction and maintenance function of the entire system. In other words, this embodiment details the main technical points, including system module settings and cooperation methods, acquisition strategies, and data cleaning and feature extraction methods. In subsequent embodiments, deeper technical details such as model analysis, application of the multi-head attention mechanism, generation of maintenance suggestions, and system feedback optimization will continue to be elaborated on this basis.

[0035] It should be emphasized that the sensor types, installation positions, sampling frequencies, etc. described in this embodiment are not limited to specific brands or specific process conditions. In actual engineering, they can be flexibly adjusted according to the oscillator type, installation environment, cost investment, and maintenance cycle requirements. As long as they conform to the overall design concept of the present invention, they fall within the protection scope of the present invention.

[0036] Embodiment 2: As Figure 1 and Figure 2 shown, the implementation of the multi-head attention mechanism In the previous embodiment, the data processing module has already cleaned, normalized, and extracted key features from the original sensor data. Next, the model analysis module inputs these features into a pre-trained model based on the Transformer architecture and relies on the multi-head attention mechanism to focus on and fuse different sensor data.

[0037] Division of multi-head attention heads The first attention head: Focus on vibration features Specifically monitor changes in frequency and amplitude to help capture early signals of mechanical faults. Typical mechanical problems such as abnormal increase in vibration and vibration frequency drift can be prominently displayed in this attention head.

[0038] The second attention head: Focus on electrical features Actively pay attention to changes in current, voltage, and resistance. If phenomena such as continuous excessive current, unstable voltage, or sudden change in resistance occur, this attention head will provide a higher weight to help the model identify electrical faults or abnormal energy losses.

[0039] The third attention head: Focus on temperature features Specifically track the temperature of the oscillator itself and the surrounding temperature conditions, which is used to detect overheating or poor heat dissipation faults. Whether the temperature rises slowly or rapidly can be promptly noticed under this attention head.

[0040] Additional attention head: Comprehensive analysis of power and environmental characteristics Mainly for power consumption levels and environmental sensor data (such as humidity, pressure, etc.), it is used to determine whether the load change is normal and whether environmental factors have an obvious impact on the operation of the oscillator.

[0041] Dynamic adjustment of attention weights In actual operation, the fault types of the oscillator often have a certain degree of uncertainty and dynamics. For example, in a high-temperature environment, abnormal temperature is more likely to be an early fault symptom; while in high-load or high-frequency operation, fluctuations in power and electrical characteristics may be more worthy of attention.

[0042] Therefore, the multi-head attention mechanism described in this embodiment can real-time sense the fluctuations of various characteristic indicators, and through the internal attention weight reallocation algorithm, automatically increase or decrease the weights of different attention heads, which can ensure that the model gives priority attention to the characteristic changes with the most fault indication significance, thereby improving the sensitivity and accuracy of fault prediction.

[0043] Case example: When the system detects that the current value of a certain oscillator continues to rise, but the temperature has not changed significantly, the model analysis module will increase the weights of the current, voltage, and resistance characteristics corresponding to the "second attention head" to more quickly detect potential electrical system hazards. And if the system simultaneously senses that the temperature is rising rapidly, the weight of the "third attention head" will also increase accordingly, so as to ensure that thermal faults can also be promptly warned.

[0044] Transformer model fine-tuning training Next, in order to make the Transformer model more adaptable to the scenario of ultrasonic oscillator fault prediction, we will conduct targeted fine-tuning training. In this embodiment, the fine-tuning process mainly starts from the following three aspects: Collection of historical data and fault labels In order to enable the pre-trained model to master more accurate industrial fault patterns, we need a large amount of historical operation data and fault labels, including normal operation data and various known fault cases.

[0045] In particular, it is necessary to focus on collecting data on mechanical faults, electrical faults, and overheating / poor heat dissipation faults that occur in different operation stages and under different working conditions, and accurately mark the fault occurrence time and fault type.

[0046] Multi-modal fusion of training data During fine-tuning, we organize the input features required by the multi-head attention mechanism (such as vibration, power, electricity, and temperature) into continuous time-series samples; Each time-series sample contains both time-domain features and key indicators obtained through frequency-domain or statistical analysis. Through multi-modal fusion, the model can learn to associate the coupling relationships between different sensor data during training.

[0047] Optimization strategies for model parameters Learning rate tuning: Usually set a relatively small initial learning rate to ensure that the Transformer does not experience gradient explosion or fall into local overfitting when facing the complexity of industrial data; Initialization of Attention Head weights: Set relatively high initial attention weights for data dimensions related to faults (such as electrical features or vibration features) to help the model focus on key features in the early learning stage; Fault classification and prediction: During fine-tuning, by comparing the fault probabilities output by the model with historical fault labels, continuously adjust the parameters of each layer of the network to gradually improve its recognition rate and discrimination ability for different fault modes.

[0048] Recognition of multiple types of fault modes After completing the above training, the model can better identify and distinguish mechanical, thermal, and electrical faults. For example: Mechanical faults: Abnormal peaks appear in the spectrum, and the vibration amplitude increases significantly; Thermal faults: A rapid temperature rise curve appears in the temperature sensor data, or exceeds a specific high-temperature threshold; Electrical faults: One or more sensor channels of current, voltage, and resistance continuously deviate from the normal range.

[0049] Operating scenarios and examples In the actual deployment scenario of this embodiment, the system continuously receives the preprocessed data from the data processing module described in Embodiment 1 and shows a significant improvement in the following situations: Early capture of subtle fault signs When there are slight electrical fluctuations or small offsets in the vibration frequency of the oscillator, a single sensor may not be able to clearly identify them. However, since the multi-head attention mechanism includes both electrical and vibration features in the "important attention objects", the system can issue an early warning as soon as possible to avoid the deterioration of the fault.

[0050] Differentiation of multi-fault coupling scenarios In some cases, mechanical and electrical failures may occur simultaneously, or mechanical vibration anomalies may be caused by thermal failures. Through the multi-head attention mechanism, the model can distinguish whether it is a single failure mode or a superposition of multiple failures, and provide a more accurate judgment of the failure type to subsequent application modules or maintenance personnel.

[0051] Model adaptability When the operating environment, process parameters, or oscillator load change, the multi-head attention mechanism can dynamically adjust the focus of attention, reduce the impact of environmental changes on the prediction accuracy, and thus maintain a high failure detection sensitivity under different production conditions.

[0052] Interaction with other modules Docking with the data processing module: The inputs required by the model analysis module in this embodiment all come from the feature extraction results described in Embodiment 1. Therefore, the data processing module should ensure that it provides a data sequence with a unified format and aligned timestamps.

[0053] Linkage with the application module: After obtaining the failure probability, failure type, and remaining service life, the multi-head attention mechanism will give the feature dimension with the "highest attention"; the application module can use these results to provide targeted maintenance suggestions for the operation and maintenance personnel.

[0054] Closed-loop with the feedback optimization module: In the subsequent deployment stage, if there is a deviation between the actual maintenance result and the model prediction, the system will collect this information in a timely manner, retrain or adjust the attention weight allocation to achieve continuous optimization of the closed-loop feedback.

[0055] This embodiment details: How the multi-head attention mechanism focuses on vibration, electrical, temperature, power, and environmental data in separate modules and realizes dynamic adjustment of weights; How the pre-trained Transformer model is fine-tuned in combination with historical data to accurately identify mechanical, thermal, and electrical failures; The key links in the actual deployment process, including early capture of small failures, multi-failure coupling analysis, and adaptability to environmental changes, etc.

[0056] Through the multi-head attention mechanism and model fine-tuning training of this embodiment, the accuracy and stability of the entire system in failure identification can be significantly improved, and high-reliability prediction results and analysis bases are provided for subsequent application modules and feedback optimization modules. In this way, no matter what type of failure mode is faced, the system can identify it in a timely manner and give reasonable operation and maintenance suggestions, minimizing equipment downtime losses and extending the service life of ultrasonic oscillators to the greatest extent.

[0057] Embodiment 3: Such as Figure 1 and Figure 2As shown, the fault warning sub-module In actual industrial production, abnormal equipment often requires rapid and clear feedback so that maintenance personnel can intervene immediately and avoid the expansion of accidents or losses. Therefore, in this embodiment, the "fault warning sub-module" is set as one of the most front-end functions in the application module. Its main responsibility is to automatically send an alarm message to the maintenance personnel when the predicted fault probability exceeds the safety threshold set by the system.

[0058] Setting of the fault probability threshold When the system is deployed, one or more fault probability thresholds (such as 50%, 70%, 90%, etc.) will be preset according to specific process characteristics, equipment importance, and past fault statistics data. When the fault probability given by the model analysis module (see Embodiment 2) exceeds a certain threshold, the fault warning sub-module will immediately trigger an alarm at the corresponding level. The threshold can be adjusted in the setting interface of the application module to meet the requirements of different devices for fault sensitivity at different operating stages.

[0059] Alarm prompt and information recording Once the fault probability exceeds the limit, the system will automatically send an alarm message to the maintenance personnel, usually including various forms such as pop-up windows, text messages, emails, or on-site audible and visual alarms, to ensure that it is noticed immediately. At the same time, the system will record data such as the alarm time, oscillator position information (such as equipment number, production line number, etc.), and fault probability in a log file or database for subsequent review and analysis. If the actual situation confirms that a fault has occurred or is about to occur, the fault warning sub-module will also synchronize this information to the maintenance suggestion sub-module and the historical data query sub-module to provide a basis for further maintenance decisions.

[0060] User-defined function The system interface allows operation and maintenance personnel to customize alarm actions and processing priorities under different thresholds according to experience or on-site needs. For example, a general reminder can be given when the low fault probability exceeds the limit, and an emergency response program can be immediately started when the high fault probability exceeds the limit.

[0061] Through the above mechanism, the fault warning sub-module can play the functions of real-time monitoring and rapid warning in a complex industrial site, greatly reducing production downtime or safety accidents caused by fault delays.

[0062] Maintenance suggestion sub-module After the fault warning sub-module issues a warning, maintenance personnel often need to quickly obtain clear and practical maintenance plans. In this embodiment, the output results of the model analysis module (including fault types, remaining service life, key feature attention levels, etc.) are combined with historical experience to refine the core functions provided by the maintenance advice sub-module.

[0063] Guidance based on fault type and remaining service life When the system predicts faults, it will give the possible fault types (such as mechanical faults, electrical faults, thermal faults) of the oscillator and the corresponding remaining service life (RUL). The maintenance advice sub-module automatically matches the pre-established "fault countermeasures" knowledge base according to different fault types, and presents the corresponding maintenance methods, spare part replacement processes, or production process adjustment plans to the user. For oscillators with a relatively short RUL, it can prompt for early shutdown or preventive maintenance; for cases where the RUL is still relatively long but the fault probability increases, relatively mild maintenance strategies (such as partial detection or increased monitoring frequency) will be proposed.

[0064] Differentiated handling of electrical and mechanical faults Electrical abnormal scenario: When the current, voltage, or resistance exceeds the limit or fluctuates violently, the system will prompt the maintenance personnel to check whether there are problems such as poor contact, aging, or short circuit in the electrical circuit, and suggest testing the electrical insulation and measuring the contact resistance, etc.; at the same time, it may suggest replacing the fuse or adjusting the power supply voltage stabilizing equipment. Mechanical abnormal scenario: For situations such as excessive vibration frequency or amplitude, and abnormal power consumption, the system will provide operation guidelines such as replacing mechanical components, tightening bolts, and lubrication maintenance, or suggest checking for jamming or structural component damage. Process parameter adjustment suggestions If the system finds that the fault is closely related to process conditions (such as excessive load, too high working temperature), it will suggest appropriately reducing the load, slowing down the working frequency of the oscillator, or enhancing cooling measures to avoid further deterioration of the problem. In some cases, the application module can even be linked with the upper control system to automatically issue commands to fine-tune the production process, such as reducing the power output or shortening the ultrasonic working duration.

[0065] User-friendly interface design To facilitate on-site maintenance personnel to quickly understand and execute the suggestions, the system will present the maintenance plan in an intuitive way such as charts, texts, and animations on the interface. When it comes to more complex operations, it will also provide interactive step-by-step breakdowns, enabling novice or temporary responsible persons to complete most basic maintenance operations following the prompts.

[0066] Historical data query and visualization In addition to real-time fault warnings and maintenance decisions, many enterprises also hope to review and summarize past fault situations to improve subsequent maintenance strategies or optimize production processes. Therefore, in this embodiment, a historical data query sub-module is specifically established within the application module, which mainly realizes the following functions: Multi-dimensional data visualization Maintenance personnel can view historical curves or comparison charts of parameters such as vibration, temperature, current, voltage, resistance, and power at different time periods on the system interface; The system allows users to focus on a specific time window by dragging or zooming in to observe the dynamic change process before and after the fault. For a specific oscillator or a group of oscillators, information such as the cumulative number of faults and the warning occurrence frequency can also be comprehensively presented in various forms such as line charts, bar charts, and pie charts.

[0067] Fault records and annotations Each fault warning or real fault leaves a detailed record in the database, including the occurrence time, fault type, handling method, and repair time. Users can also add text descriptions or attachments (such as maintenance reports and on-site photos). This information not only helps on-site maintenance personnel quickly review the handling process of past faults but also provides a quantitative basis for managers to evaluate equipment reliability and estimate maintenance costs.

[0068] Trend analysis and data mining By inductively analyzing historical data, the system can help users discover whether certain specific types of faults are more likely to occur during seasonal changes, production load changes, or specific process stages; The analysis results of this part can also be further fed back to the model analysis module and the feedback optimization module for strengthening the online learning of the model and parameter adaptive adjustment.

[0069] Connection and value-added functions with other modules Receiving model analysis results In Embodiment 2, the Transformer model with a multi-head attention mechanism outputs the fault probability, remaining service life, and main suspicious features of each oscillator. The application module reads these results to achieve external information display and decision support. Information interaction and data return flow After the maintenance personnel complete the maintenance, they can fill back information such as the actual maintenance results, fault causes, and repair situations to the record interface of the application module, and then the feedback optimization module collects and uses it for the next round of model optimization. In this way, a data closed-loop is formed: real-time monitoring → model analysis → maintenance guidance → on-site implementation → result feedback → continuous optimization.

[0070] Practical value Warning: Ensure a warning is issued before or at the beginning of a fault, reducing the risk of sudden shutdown or safety accidents; Suggestion: Through a deep learning model that integrates historical experience and real-time data, quickly provide executable maintenance plans or suggestions for production process adjustment; Data query and analysis: Provide comprehensive and visual data support for fault review, trend prediction, and management decision-making.

[0071] In this embodiment, the application module is not only an information window for the system to the outside world but also a complete operation and maintenance assistance platform. It combines functions such as fault warning, maintenance plan recommendation, and historical data visualization, enabling maintenance personnel to always grasp the equipment status, quickly respond to potential faults, and continuously improve maintenance strategies. All technical implementations are based on the data collection and model analysis results of the previous two embodiments, forming a closely connected and mutually supportive whole.

[0072] As the system is deployed and run in the actual industrial environment, maintenance personnel will continuously accumulate new data and fault cases, further enriching the system's knowledge base, making the fault warning and maintenance suggestions more in line with the on-site reality, and maximizing the safety, efficiency, and predictability of the ultrasonic oscillator group operation and maintenance. Through the detailed introduction of this embodiment, the application method of the present invention in production practice and its profound impact on fault monitoring and preventive maintenance work can be better understood.

[0073] Embodiment 4: As Figure 1 and Figure 2 shown, in this embodiment, we connect the various functional modules of the system in series to form a closed-loop fault prediction and maintenance process, which mainly includes the following steps: 1. Data collection (S1) 2. Data transmission (S2) 3. Data preprocessing (S3) 4. Model analysis (S4) 5. Prediction result and maintenance suggestion (S5) 6. Feedback and model update (S6) These six major steps correspond to the overall method process of the system, which also includes several key features, such as dynamic adjustment of the sensor sampling frequency, dynamic parameter adjustment of the multi-head attention mechanism, and specific guidance plans for maintenance suggestions. Below, the implementation of each technical feature will be described in detail in combination with each step.

[0074] Data collection (S1) Multi-source sensor layout According to the introduction of Embodiment 1, each ultrasonic oscillator is equipped with vibration sensors, temperature sensors, power sensors, and electrical sensors such as current, voltage, and resistance; environmental sensors (such as humidity and pressure) can also be added if necessary. During the production process, these sensors will collect corresponding physical quantities in real time, forming a multi-dimensional monitoring of the operating state of the oscillator.

[0075] Sampling Frequency and Its Dynamic Adjustment To balance data quality and system load, the conventional sampling frequency is generally set between 10Hz and 100Hz. However, when encountering certain special situations, such as sharp fluctuations in current or voltage that last for more than a preset threshold, the system will automatically increase the sampling frequency of the oscillator (e.g., from 50Hz to 100Hz) to more accurately capture possible high-frequency short-term faults.

[0076] Case Example: If it is detected that the current rises significantly within just a few seconds, but the temperature has not changed significantly yet, the system will automatically increase the sampling rate of the current and voltage sensors to obtain more transient data. These additional data can help the model analysis module judge more quickly and accurately whether there are potential electrical short circuit or overload hazards.

[0077] Data Transmission (S2) In this embodiment, to ensure real-time performance, data is transmitted through the Industrial Internet of Things (IIoT) network, or in a smaller-scale factory environment, it can be transmitted through a local area network (LAN). After the data collected by the sensors is packed and appended with a timestamp, it is sent to the data center or the cloud according to a preset protocol.

[0078] If there are a large number of on-site devices, the system usually adopts a hierarchical transmission architecture: first, perform preliminary integration at the local controller or edge computing node, and then upload the combined data to the central server. During the data transmission process, factors such as network latency and bandwidth limitations are considered. For data peaks within a short period, a caching mechanism or batch uploading can be used.

[0079] Data Preprocessing (S3) After the raw data is received at the data center or the cloud, it needs to go through a series of preprocessing operations (refer to the description of the data processing module in Embodiment 1) to ensure the accuracy and efficiency of subsequent model analysis: Denoising and Outlier Removal For transient outliers caused by sensor jitter, external interference, or network transmission packet loss, the system can use algorithms such as moving average, filtering, or interpolation repair to correct or remove them. Normalization Processing Convert the numerical values of different dimensions such as vibration, temperature, and electricity to a relatively unified numerical range (such as [0, 1]) to prevent certain large numerical features from having an excessive impact on subsequent model training. Key feature extraction Vibration features (time-domain and frequency-domain indicators, such as RMS, peak factor, main frequency, etc.); Power fluctuation features (average power, peak power, fluctuation range); Electrical parameter features (transient change rates of current, voltage, and resistance, average offset, etc.); Temperature and environmental features (optional data such as temperature change rate, humidity, air pressure, etc.), After being processed in this stage, the data is packed into a sequence of feature vectors and prepared to be input into the model analysis module.

[0080] Model analysis (S4) After the above preprocessing, the data will be input into a pre-trained model based on the Transformer architecture (see Embodiment 2 for details). The most crucial part is the multi-head attention mechanism, that is, the model can apply different attention weights to vibration features, electrical features, temperature features, etc.

[0081] Attention head dynamic parameter adjustment When the system detects high load or rising ambient temperature, the model will automatically increase the attention to temperature, current, and voltage features; When abnormal peaks or frequency drifts appear in the vibration signal, the attention to vibration features will be enhanced; Through this dynamically adjusted weight allocation mechanism, the model can maintain high prediction accuracy and sensitivity to fault symptoms under various complex working conditions.

[0082] Prediction results Fault probability: Give the possibility of the oscillator failing in the future for a certain period of time; Remaining useful life (RUL): Estimate the life of the oscillator under the current working conditions based on time series analysis; Fault type: Indicate the most likely fault category, such as electrical fault, mechanical fault, thermal fault, etc., These prediction results will be further utilized in the next step "S5. Prediction Results and Maintenance Suggestions".

[0083] Prediction Results and Maintenance Suggestions (S5) When the model analysis result shows that the fault probability of a certain oscillator exceeds the set threshold (see the description of the warning threshold in Embodiment 3 for details), the system will send a warning to the maintenance personnel through the application module and at the same time provide corresponding maintenance suggestions, specifically including: Electrical circuit detection If electrical characteristics (current, voltage, resistance) are abnormal, it is recommended to inspect or replace relevant circuits, terminal blocks, and fuses to eliminate risks such as poor contact and overheating. Mechanical structure reinforcement When the vibration amplitude increases or the main frequency shows an abnormal peak, the system prompts to check structural components such as the ultrasonic oscillator's frame, fasteners, and coupling, and perform necessary reinforcement or lubrication maintenance. Cooling system maintenance When the temperature sensor reports a high heating rate or the ambient temperature is too high, it is recommended to check the cooling fan and water cooling pipeline to ensure the temperature remains within a safe range. Replacement or upgrade of key components For oscillator components that have been in operation for a long time and whose remaining service life approaches the lower limit, the system provides reference suggestions on whether to scrap in advance, replace, or upgrade to avoid unexpected downtime or safety accidents. Handling of abnormal power consumption When the power consumption increases significantly and does not match the expected load, corresponding adjustments can be made to the production process or operating environment according to the system prompt, such as reducing the workload, replacing worn parts, or improving the process flow.

[0084] Through these maintenance suggestions, maintenance personnel can not only handle potential faults in a timely manner but also plan the maintenance timing based on the remaining service life indicators provided by the system, reducing maintenance costs and equipment downtime risks.

[0085] Feedback and model update (S6) Finally, after completing the actual maintenance or repair operations, maintenance personnel will fill back data such as the faulty parts, fault causes, repair information, and equipment operating status they found into the system (see the "Historical Data Query and Visualization" function in Embodiment 3). The system then hands over this information to the feedback optimization module for processing and analysis, thereby continuously improving the parameter settings and attention weight allocation strategies of the pre-trained Transformer model.

[0086] Model parameter correction If it is found during several actual maintenance processes that the model's fault prediction has missed alarms or false alarms, the system will automatically collect these cases and perform offline or online fine-tuning training to update hyperparameters such as the weight initialization or learning rate of the attention mechanism. Optimization of data acquisition strategy If in certain types of faults, high-frequency sampling significantly improves the detection rate, the system will consider triggering an increase in the sampling frequency earlier under the corresponding working conditions; if the data redundancy of some sensors is too high, the sampling frequency may be appropriately reduced to relieve network and storage pressure. Improvement of maintenance manual Meanwhile, the knowledge base in the maintenance advice sub-module can also be continuously enriched and corrected, making the operation and maintenance advice more in line with the actual needs of the enterprise site.

[0087] Through this closed-loop of feedback and model update, the system will continuously learn and evolve during long-term operation, gradually forming a mature "intelligent predictive maintenance solution for ultrasonic oscillator groups" exclusive to the enterprise.

[0088] Deployment and Benefits in Industrial Environments Combined with the process description of this embodiment, the following steps can be followed when deploying this system in an actual industrial site: 1. Hardware installation: Install necessary sensors on the ultrasonic oscillators; 2. Software configuration: Integrate data transmission, preprocessing, model analysis modules, etc. into the factory network and servers; 3. Parameter setting: Set the sampling frequency range and fault warning threshold according to production requirements; 4. Go live and run: Continuously monitor the oscillator status, perform inspections regularly or irregularly, and fill back the results to the system; 5. Iterative optimization: The model continuously improves accuracy as the amount of on-site data increases, and the maintenance advice becomes more accurate.

[0089] During this process, the enterprise can often significantly reduce the equipment unexpected shutdown rate, reduce the maintenance labor cost, and extend the service life of the oscillator components, thereby improving the overall production efficiency and economic benefits.

[0090] Through this embodiment, the present invention provides a complete and feasible fault prediction and maintenance method, covering the full life cycle management of data collection and dynamic adjustment of sampling frequency (S1), data preprocessing (S3), model analysis based on the multi-head attention mechanism (S4), maintenance advice (S5), and feedback and model update (S6). It integrates the system modules, algorithm mechanisms, and application functions described in the previous embodiments 1, 2, and 3, and truly realizes the intelligent prediction and proactive maintenance of ultrasonic oscillator groups in the industrial production environment.

[0091] The implementation of this embodiment can help enterprises predict equipment anomalies in a timely manner, avoid the expansion of faults, and enable the system to continuously self-learn and self-optimize through dynamic sampling, attention adjustment, and continuous feedback. Thus, it can still maintain a good fault detection rate and low false alarm rate under complex and changeable working conditions, promoting the development of industrial maintenance towards high efficiency, precision, and intelligence.

[0092] The above are only the preferred embodiments of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein, and should not be regarded as excluding other embodiments. Instead, it can be used in various other combinations, modifications, and environments, and can be changed within the scope of the concept described herein through the above teachings or the techniques or knowledge in related fields. As long as the changes and variations made by those skilled in the art do not depart from the spirit and scope of the present invention, they should all be within the protection scope of the appended claims of the present invention.

Claims

1. An ultrasonic vibrator group intelligent predictive maintenance system, characterized in that: include: The data acquisition module is used to collect data related to the operating status of each ultrasonic vibrator using various sensors deployed on the vibrator, including: The vibration sensor collects the vibration frequency and amplitude; The temperature sensor collects the working temperature of the oscillator; The power sensor collects the power consumption information of the oscillator; The electrical sensor collects the current, voltage and resistance during the operation of the vibrator; Environmental sensors can optionally collect environmental parameters such as humidity and pressure; A data transmission module, used to transmit the multi-source data acquired by the data acquisition module to a data center through an industrial Internet of Things network; The data processing module is used to perform data cleaning, normalization and key feature extraction on the received multi-source data; A model analysis module, which is used to predict and analyze the key features based on a pre-trained Transformer architecture model. The model combines a multi-head attention mechanism to focus on changes in vibration characteristics, current, resistance, voltage, temperature, power, and environmental data, and outputs the failure probability, remaining service life, and possible failure type of each vibrator; Application module, used to present fault prediction results and maintenance suggestions through the user interface, including real-time status monitoring, fault warning, maintenance guidance and historical data query; The feedback optimization module is used to obtain the actual maintenance results of maintenance personnel, and update and adjust the parameters of the pre-trained model based on the accuracy of fault diagnosis, so as to continuously improve the accuracy of fault prediction and the rationality of maintenance recommendations.

2. The ultrasonic vibrator group intelligent predictive maintenance system according to claim 1, characterized in that: The sampling frequency of the data acquisition module is set between 10 Hz and 100 Hz according to the equipment operation characteristics and fault response requirements, and can be dynamically adjusted according to the abnormal fluctuation of the oscillator.

3. The ultrasonic vibrator group intelligent predictive maintenance system according to claim 2, characterized in that: The data processing module comprises: Data cleaning unit, used to remove noise data, missing values ​​and transient outliers; Normalization units are used to convert data of different dimensions and ranges to the same standard range; The feature extraction unit is used to extract the frequency domain features, time domain features, temperature change rate, power fluctuation range, and change rate or transient response features of the vibration signal.

4. The ultrasonic vibrator group intelligent predictive maintenance system according to claim 3, characterized in that: The multi-head attention mechanism in the model analysis module includes at least: The first attention head is used to focus on the vibration frequency and amplitude to capture the early signs of mechanical failure; A second attention head for focusing on changes in current, voltage, and resistance to identify electrical faults or abnormal energy loss; A third attention head for focusing on temperature changes to detect overheating or poor heat dissipation faults; at least one additional attention head for integrating power and environmental characteristics to analyze the impact of load changes and environmental factors on the oscillator operation; Among them, the weight of each attention head can be automatically adjusted dynamically according to the real-time operating conditions.

5. The ultrasonic vibrator group intelligent predictive maintenance system according to claim 2, characterized in that: The pre-trained model is combined with the historical current, resistance, voltage, vibration and temperature data of the ultrasonic vibrator for targeted training during the fine-tuning process, thereby improving the recognition and prediction capabilities of multiple types of fault modes.

6. The ultrasonic vibrator group intelligent predictive maintenance system according to claim 5, characterized in that: The application module includes: A fault warning submodule is used to automatically send an alarm to maintenance personnel when the predicted fault probability is higher than a preset threshold; The maintenance suggestion submodule is used to provide corresponding maintenance guidance plans based on the fault type, remaining service life and abnormal electrical / mechanical parameters, including early replacement of parts, maintenance of electrical circuits or adjustment of process parameters; The historical data query submodule is used to visualize historical fault records and operating parameter curves, making it easier for maintenance personnel to analyze abnormal causes and trends.

7. An ultrasonic vibrator group intelligent predictive maintenance method, characterized in that: The following steps are involved: S1. Data acquisition: Vibration, temperature, power, current, voltage, resistance and other data are acquired in real time through various sensors deployed on the ultrasonic vibrator, and the sampling frequency is adjusted according to the working conditions; S2, data transmission: the collected data is transmitted to the data center through the industrial Internet of Things network; S3, data preprocessing: denoising, normalizing and extracting multiple features of the received data, including vibration features, power fluctuation features, electrical parameter anomaly features and environmental parameter features; S4, model analysis: input the features into the pre-trained Transformer model, and obtain the failure probability, remaining service life and failure type of each vibrator through a multi-head attention mechanism; S5. Prediction results and maintenance suggestions: Set warning thresholds based on prediction results. When the probability of failure exceeds the threshold, send an alarm to maintenance personnel and give maintenance suggestions for possible fault locations or types. S6. Feedback and model update: After the fault is handled, collect the actual maintenance results, compare the predicted results with the actual situation, and update the pre-trained model and system parameters regularly or in real time to improve the overall prediction accuracy and maintenance decision-making level.

8. The method for intelligent predictive maintenance of ultrasonic vibrator groups according to claim 7, characterized in that: When collecting the current, voltage and resistance in step S1, the transient electrical waveform changes are further monitored. When the current or voltage fluctuates violently and the duration exceeds a preset threshold, the sampling frequency of the sensor is automatically increased.

9. The method for intelligent predictive maintenance of ultrasonic vibrator groups according to claim 7, characterized in that: The multi-head attention mechanism in step S4 automatically adjusts the weight distribution of each attention head according to the real-time operating environment of the vibrator, including increasing the attention to temperature, current, and voltage characteristics under high load or high temperature conditions, and increasing the attention to vibration characteristics when the vibration frequency is abnormal.

10. The method for intelligent predictive maintenance of ultrasonic vibrator groups according to claim 7, characterized in that: The maintenance suggestions in step S5 include specific guidance plans for electrical circuit inspection, mechanical structure tightening, cooling system maintenance, replacement or upgrading of key components, and adjustment of production processes or operating environments in the event of abnormal power consumption.

Images