Vibration-temperature correlation fault early warning method based on multi-sensor fusion

Abstract

The invention provides a vibration-temperature correlation fault early warning method based on multi-sensor fusion, and relates to the technical field of rotating machine fault diagnosis, and the method achieves the efficient diagnosis of the rotating machine fault through the three-field coupling analysis of vibration, temperature and torque. Multi-dimensional health indexes are introduced, sensor data of different dimensions are normalized into a monitorable numerical value, and the health state of equipment is reflected in real time through a dynamically optimized weight coefficient. A core depth time sequence neural network model is combined with an attention mechanism and a Transform coding layer, the problem of long-term dependence is effectively solved, and the fault recognition precision is improved. According to the method, future innovation directions such as cross-device collaborative diagnosis, quantum computing acceleration and digital twinborn visualization are also exhibited, and the method has important industrial application value.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of rotating machinery fault diagnosis, and particularly relates to a vibration-temperature correlation fault early warning method based on multi-sensor fusion. BACKGROUND

[0002] In modern industrial production, rotating machinery equipment is widely used in various production scenes, and its stable and reliable operation is crucial to ensure production efficiency and product quality. However, rotating machinery is prone to failure due to various factors during long-term operation, such as bearing wear, gear damage, and lubrication failure. If these faults are not discovered and addressed in a timely manner, it will lead to equipment downtime, production interruption, and even safety accidents, causing significant economic losses.

[0003] Traditional rotating machinery fault diagnosis methods mainly rely on single sensor data, such as monitoring equipment status only through vibration signals or temperature signals. Vibration analysis can reflect the dynamic characteristics of mechanical components, such as impact vibration caused by bearing spalling and gear tooth breakage. However, relying solely on vibration signals makes it difficult to accurately determine the root cause of the fault, and in low-speed operating conditions, the signal-to-noise ratio of the vibration signal is low, making the fault characteristics less obvious and increasing the difficulty of diagnosis. Temperature monitoring can reflect the thermal state of the equipment, such as the frictional heating caused by lubrication failure. However, temperature changes alone cannot accurately locate the fault position and type, and the response speed of temperature sensors to sudden failures is relatively slow. In addition, traditional fault diagnosis systems mostly use static threshold alarm methods, which trigger an alarm once the sensor data exceeds the pre-set threshold. This method is prone to false alarms or missed alarms when the operating conditions of the equipment change, and it cannot effectively predict the development trend of the fault, making it difficult to meet the high precision and high reliability requirements of modern industry for equipment fault diagnosis.

[0004] With the rapid development of industrial automation and informatization, multi-sensor fusion technology has gradually emerged, bringing new opportunities for rotating machinery fault diagnosis. By fusing data from multiple different types of sensors, the equipment status can be comprehensively perceived from multiple angles, and complex fault patterns can be more accurately identified in theory. However, some existing multi-sensor fusion fault diagnosis systems still have many shortcomings. On the one hand, the sampling rate, data format, and other differences between different sensors lead to feature drift problems during feature extraction and fusion, affecting the diagnosis accuracy. On the other hand, the analysis of multi-physical field coupling effects is not deep enough, and the internal relationship between different physical quantities is not fully explored, making it difficult to build a complete fault transmission chain model and achieve early warning and accurate positioning of faults.

[0005] In actual industrial scenarios, the fault diagnosis of rotating machinery equipment faces many challenges. For example, in the fault diagnosis of wind turbine gearboxes, due to the complex operating environment and frequent load changes, traditional single-physical-field diagnostic methods often fail to work. When early wear occurs in the bearings, it may be difficult to capture the weak fault features through vibration signals alone, and the change in temperature signals is relatively lagging. When gear tooth breakage occurs, torque fluctuations are obvious, but if only torque data is used, it is difficult to accurately determine whether the fault is in the gear itself or other related components. Therefore, there is an urgent need for an innovative fault diagnosis method that can break through the limitations of traditional technology.

[0006] In view of this, the present application is proposed. SUMMARY

[0007] The present application provides a vibration-temperature correlation fault warning method based on multi-sensor fusion, which can at least partially improve the above problems.

[0008] To achieve the above object, the present application adopts the following technical solutions: A vibration-temperature correlation fault warning method based on multi-sensor fusion, comprising: Obtaining a multi-source data set collected by a preset sensor component, preprocessing the multi-source data set, and aligning the preprocessed multi-source data set to obtain a JSON data package; According to the MHI index, the JSON data package is fused and processed to extract a feature vector, and sensor data of different dimensions are normalized into a health index; A pre-trained DTNN model is used to infer the feature vector to generate a fault probability and a fault type classification; The fault probability is judged and processed to generate a judgment result, and the judgment result is used for alarm triggering or updating processing.

[0009] In summary, the vibration-temperature correlation fault warning method based on multi-sensor fusion realizes cross-physical-field coupling analysis of vibration, temperature and torque data through multi-sensor fusion technology, effectively improving the accuracy and reliability of fault diagnosis. The core of this method is its unique multi-dimensional health index (MHI), which normalizes sensor data of different dimensions into a unified quantitative index by fusing time-frequency domain features and temperature gradients, providing an intuitive evaluation of the health status of the equipment. In addition, a deep time series neural network (DTNN) based model is used, which introduces an attention mechanism and a Transformer encoding layer, effectively handling long-term dependencies and improving the accuracy and efficiency of fault identification.

[0010] In terms of technical implementation, the method realizes microsecond-level time alignment of multi-sensor data through the PTP protocol, solving the feature drift problem caused by differences in sampling rates in traditional systems. The DTNN model can not only analyze the morphological features of MHI curves, but also update the fault mode library in real time through adaptive learning, further improving adaptability and robustness. Through these innovative technologies, the method significantly improves the speed and accuracy of fault diagnosis while reducing false positive rates, providing strong technical support for predictive maintenance of industrial equipment. BRIEF DESCRIPTION OF DRAWINGS

[0011] Figure 1 is a flowchart of a vibration-temperature correlation fault early warning method based on multi-sensor fusion provided by an embodiment of the present application; Figure 2 is a flowchart of a vibration-temperature correlation fault early warning method based on multi-sensor fusion provided by an embodiment of the present application; Figure 3 is a flowchart of a DTNN model provided by an embodiment of the present application; Figure 4 is a digital twin visualization diagram provided by an embodiment of the present application. DETAILED DESCRIPTION

[0012] To make the objectives, technical solutions and advantages of the present application clearer, further detailed descriptions will be given below with reference to embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0013] Referring to Figure 1 , Figure 2 The first embodiment of the present application discloses a vibration-temperature correlation fault early warning method based on multi-sensor fusion, which can be executed by a vibration-temperature correlation fault early warning device based on multi-sensor fusion (hereinafter referred to as an early warning device), and in particular, by one or more processors in the early warning device to implement the following method: S1, obtaining a multi-source data set collected by a preset sensor component, preprocessing the multi-source data set, and performing alignment processing on the preprocessed multi-source data set to obtain a JSON data packet; Preferably, the sensor component includes a vibration sensor, a temperature component, and a torque sensor, the temperature component includes a temperature sensor and a digital conversion module, the vibration sensor is a three-axis accelerometer, the temperature sensor is a PT100 thermistor, the model of the digital conversion module is MAX31865, and the torque sensor is an HBM T40B torque flange.

[0014] Specifically, step S1 further comprises: receiving vibration data transmitted by the vibration sensor through the SPI interface, receiving temperature data transmitted by the temperature sensor through the I2C interface, and receiving torque data transmitted by the torque sensor through the CAN bus, to obtain a multi-source data set; The vibration data is subjected to Butterworth low-pass filtering to remove high-frequency noise, the temperature data is subjected to sliding average filtering, and the transmission delay of the torque data is compensated based on linear interpolation. Based on the IEEE 1588 precision time protocol, the timestamps of all sensors in the sensor assembly are synchronized to obtain a time-aligned multi-dimensional array, and the multi-dimensional array is encapsulated into a JSON data packet according to the TCP / IP technology.

[0015] In the embodiment, first, the data acquisition and preprocessing step is performed. This step involves obtaining a multi-source data set from preset sensor assemblies, including a vibration sensor, a temperature assembly, and a torque sensor. The vibration sensor uses a three-axis accelerometer (such as ADI ADXL355) to accurately measure the vibration of the device in three directions, i.e., raw acceleration data (±40g range, 16-bit resolution, 10kHz sampling rate); the temperature assembly uses a PT100 thermal resistor as a temperature sensor, and a digital conversion module with model MAX31865 to accurately collect temperature data (-40℃~150℃, accuracy ±0.5℃); and the torque sensor uses a HBM T40B torque flange to accurately measure the torque change and rotational speed data of the device (±5000Nm, accuracy 0.05%FS).

[0016] During the data acquisition process, vibration data transmitted by the vibration sensor is received through the SPI interface, temperature data transmitted by the temperature sensor is received through the I2C interface, and torque data transmitted by the torque sensor is received through the CAN bus (J1939 protocol), thereby forming a multi-source data set. This multi-source data set ensures the diversity and accuracy of the data source, and provides a rich information base for subsequent fault diagnosis.

[0017] It should be noted that in the embodiment, in addition to these types of sensors, to further perform multi-physical field coupling modeling, the types of sensors can be expanded: acoustic emission sensors (to detect bearing cracks) and current sensors (to monitor motor winding abnormalities) can be added to build a vibration-temperature-acoustic-current four-dimensional fusion model. In addition, simulation-driven optimization can be performed to generate virtual sensor data through ANSYS Mechanical, thereby enhancing the robustness of the model in small sample scenarios.

[0018] Next, the collected multi-source data is preprocessed. For vibration data, a Butterworth low-pass filter is used (cutoff frequency 1 kHz) to effectively remove high-frequency noise, retain low-frequency signals that are important for fault diagnosis, improve data quality, and enable more accurate capture of equipment vibration characteristics in subsequent analysis. For temperature data, a moving average filter is used (window length 50 ms) to smooth the temperature curve and reduce data fluctuations caused by measurement errors or environmental interference, thereby more clearly reflecting the temperature trend of the equipment. At the same time, the transmission delay of torque data is compensated based on linear interpolation to ensure the consistency of torque data with other data in time, providing an accurate time reference for subsequent data fusion and analysis.

[0019] After preprocessing, the embedded processor (such as TI AM5708) sends the preprocessed data via Ethernet; based on IEEE 1588 Precision Time Protocol (PTP), the timestamps of all sensors in the sensor component are synchronized with an alignment accuracy of ≤1us. In simple terms, the PTP protocol is used to achieve us-level time alignment of vibration, temperature, and torque data, solving the feature drift problem caused by differences in sampling rates in traditional systems. This synchronization step is crucial as it ensures the consistency of data collected by different sensors in time, allowing data from different sensors to accurately correspond to the same time point, providing a reliable time basis for subsequent multi-physical field coupling analysis. After synchronization, a multi-dimensional array is obtained (TCP / IP transmission), which is a multi-dimensional feature vector constructed through Timestamp Synchronization; this array contains data from multiple dimensions such as vibration, temperature, and torque, fully reflecting the running state of the equipment under different physical fields. The transmission protocol used is: TCP / IP encapsulates JSON data packets preprocessed digital signals (SPI / I2C protocol).

[0020] Finally, the time-aligned multi-dimensional array is encapsulated into a JSON data packet based on TCP / IP technology. The use of JSON data packets not only facilitates data transmission and storage, but also allows data to be presented in a structured manner, making subsequent processing and analysis easier. By this encapsulation method, the dimensional data in the multi-source data set is integrated together to form a complete and easily processed data unit, providing a standardized input for subsequent fault diagnosis models.

[0021] S2, according to the MHI index, the JSON data packet is fused and processed to extract the feature vector, and the sensor data of different dimensions is normalized into a health index; Preferably, the calculation formula of the MHI index is: wherein, health index, represents the vibration energy spectrum density, represents the temperature change rate, represents the torque fluctuation standard deviation.

[0022] Specifically, step S2 further comprises: extracting the vibration frequency spectrum features in the JSON data packet by using the FFT algorithm, and performing wavelet transform according to the extracted vibration frequency spectrum features to detect the transient impact signal, so as to obtain a 120-dimensional frequency domain / time domain feature vector in JSON format and generate a health index; Wherein, the 120-dimensional frequency domain / time domain feature vector contains the spectrum peak value, energy entropy and temperature gradient in JSON format, the range of the health index MHI is [0, 1], when the health index MHI is greater than 0.8, it represents the health state, when the health index MHI is greater than 0.5 and less than or equal to 0.8, it represents the early warning state, and when the health index MHI is less than or equal to 0.5, it represents the fault state.

[0023] In this embodiment, the frequency domain (vibration spectrum), time domain (temperature trend) and energy domain (torque fluctuation) features are combined to construct a multi-dimensional health index (MHI); that is, the preprocessed and time-aligned JSON data packet is fused by using MHI (multi-dimensional health index), wherein the weight coefficient in the MHI index is optimized by a genetic algorithm. This process aims to convert the sensor data of different dimensions, i.e. vibration, temperature and torque data, into a unified health index through a specific calculation formula, so as to provide a quantitative evaluation of the overall health status of the equipment.

[0024] In short, the MHI index plays a key role in normalizing sensor data of different dimensions into a monitorable value. For example, vibration spectrum reflects mechanical impact, temperature change reflects lubrication state, and torque fluctuation reflects load abnormality. MHI integrates these symptoms into a "health score". When MHI continues to decline, it triggers an early warning, which is more reliable than looking at a single parameter.

[0025] Specifically, based on the calculation formula of MHI index. By weighted average of the data in these three dimensions, the health index MHI obtained can comprehensively reflect the health status of the equipment under the vibration, temperature and torque three key physical fields. The parameter explanation of the MHI index calculation formula is shown in Table 1.

[0026] Table 1

[0027] Specifically, in step S2, the vibration data in the JSON data packet is first processed using the FFT (Fast Fourier Transform) algorithm to extract the vibration spectral features (0-5 kHz, resolution 1 Hz). The FFT algorithm can convert time-domain signals into frequency-domain signals, clearly showing the energy distribution of the vibration signal at different frequencies. Subsequently, wavelet transform (Daubechies 4) is performed based on the extracted vibration spectral features, which is an effective method for detecting transient impacts in signals. Through wavelet transform, short-term mutation characteristics in the vibration signal can be further captured, which is particularly important for identifying mechanical faults. Finally, these processing results are normalized into a 120-dimensional frequency / time domain feature vector, which is stored in JSON format. This 120-dimensional feature vector covers key information such as spectral peak, energy entropy, and temperature gradient, providing a rich data foundation for subsequent health index calculation.

[0028] Based on the above feature vector, the calculated health index MHI is limited to the range [0, 1]. This quantitative range makes the evaluation of the device health status intuitive and easy to understand. When the health index MHI is greater than 0.8, it indicates that the device is in a healthy state and can operate normally; when the health index MHI is greater than 0.5 and less than or equal to 0.8, it indicates that the device is in a warning state, prompting the operator to pay attention to the device condition and take measures in advance; and when the health index MHI is less than or equal to 0.5, it clearly indicates that the device is in a fault state and needs to be repaired or replaced parts in time.

[0029] Referring to Figure 3 , S3, a pre-trained DTNN model is used to infer the feature vector to generate a fault probability and a fault type classification; Specifically, step S3 further includes: inputting the 120-dimensional feature vector into the Attention-LSTM layer to capture the spectral envelope feature, temperature gradient feature, and torque entropy feature. The calculation formula of the Attention-LSTM layer is: , Q is the query vector, K is the key vector, V is the value vector, is the scaling factor, is the normalization function, wherein the DTNN model is composed of an Attention-LSTM layer, a Transformer encoding layer, and a fully connected layer; The spectrum envelope feature, temperature gradient feature and torque entropy feature are input to the Transformer encoding layer, and the cross-physical field correlation is established through the self-attention mechanism, and the formula is: vibration_feat = encoder(vib_data), temp_feat = encoder(temp_data), cross_attention = MultiHeadAttention(vibration_feat, temp_feat), wherein vibration_feat / temp_feat is a feature vector output by the encoder, and the dimension is usually [d_model, seq_len], wherein: d_model is the feature dimension (such as 512); seq_len is the time step. The encoder refers to a feature encoder (such as CNN or Transformer), which is used to convert the original sensor data into a high-dimensional feature representation: the vibration encoder can extract frequency domain features (such as wavelet coefficients); the temperature encoder can capture spatial heat distribution patterns, vib_data is the time series data collected by the original vibration / temperature sensor, which is usually a multi-dimensional time series matrix, which can include acceleration, frequency and other vibration features or temperature values of different temperature measurement points, temp_feat, temp_data, cross_attention, MultiHeadAttention are core components of the multi-head attention mechanism, and the calculation process includes (vibration_feat, temp_feat); The full connection layer is connected, the Sigmoid activation function outputs the fault probability, and the Softmax layer outputs the fault type classification, wherein the calculation formula of the fault probability is: , is the fault probability, and the value is [0, 1], is the base of the natural logarithm, is the input value of the full connection layer.

[0030] In this embodiment, a pre-trained deep time series neural network (DTNN) model is used to infer the feature vectors to generate the failure probability (0-1) and the failure type classification; this step introduces an attention mechanism to design a deep learning model based on LSTM, inputs multi-dimensional feature vectors, and outputs failure probability and type (such as bearing wear, motor overheating), supporting early warning (≥30 min in advance). This process is the core of the entire fault diagnosis method, and its purpose is to analyze the fused feature vectors through advanced deep learning technology, thereby achieving accurate prediction and classification of equipment failure. Among them, the deep time series neural network (DTNN) is composed of Attention-LSTM layer (i.e. LSTM network, including 3 layers, 256 neurons), Transformer encoding layer and fully connected layer.

[0031] In simple terms, the innovation of the DTNN model lies in: 1) The input layer is specially designed to process MHI time series 2) Add Transformer encoding layer to capture long-term features 3) The output layer simultaneously generates failure probability and type classification. Specifically, the model will first analyze the morphological features of the MHI curve (such as mutation points, trend slope), and then match the failure mode library. For example, bearing failure usually shows MHI stepwise decline, while gear damage appears MHI pulse peak.

[0032] Specifically, first, the 120-dimensional feature vector is input into the Attention-LSTM layer of the DTNN model. The design of this layer aims to capture key information such as spectral envelope features, temperature gradient features, and torque entropy features. The Attention-LSTM layer dynamically weights key time steps through its unique attention mechanism: focusing on feature mutations 10s before failure occurs; it can dynamically focus on the most information-rich part of the feature vector, thereby improving the model's sensitivity to fault features. The calculation formula of this layer involves query vectors (Q), key vectors (K), and value vectors (V), where the scaling factor and normalization function are used to adjust and optimize the calculation process, ensuring the stability and accuracy of the model. In this way, the Attention-LSTM layer can effectively extract key features related to failure from complex feature vectors Subsequently, the extracted spectral envelope features, temperature gradient features, and torque entropy features are input into the Transformer encoding layer. This layer establishes cross-physical field associations through self-attention mechanisms, further enhancing the model's understanding of the relationships between different physical field features. Through the MultiHeadAttention function, the model can calculate cross_attention, i.e. cross-physical field attention features, which helps the model better understand the manifestation of fault features in different physical fields.

[0033] Finally, the output of the Transformer encoding layer is connected to a fully connected layer. In the fully connected layer, the failure probability is output by the Sigmoid activation function, and the Softmax layer outputs the probability distribution of 6 types of failures, as shown in Table 2. The value range of the failure probability is [0, 1], and e in the formula is the base of the natural logarithm (approximately equal to 2.71828), and z is the input value of the fully connected layer (which can be any real number). The Sigmoid activation function maps the input value to the [0, 1] interval, thereby obtaining the probability of device failure. At the same time, the Softmax layer is responsible for converting the input value into a probability distribution of fault type classification, so that the model can predict the specific fault type. Among them, the edge computing node (NVIDIA Jetson AGX Xavier) has a delay of ≤20ms. Failure probability (floating point number, MQTT protocol).

[0034] Table 2

[0035] For edge-cloud collaborative inference, its hierarchical processing architecture includes: edge layer, lightweight model performs real-time fault detection (delay ≤50ms); cloud, deploy high-precision models (such as 3D CNN) for root cause analysis and feedback to optimize edge models; 5G-MEC integration, use 5G ultra-low latency (URLLC) to transmit key data and support cross-device collaborative diagnosis. It also involves quantum computing enhancement technology, quantum feature encoding, which maps 120-dimensional features to quantum states (Qubit) and accelerates feature selection through quantum parallelism; hybrid quantum-classical training uses D-Wave quantum annealing machine to optimize LSTM hyperparameters (such as number of layers, learning rate).

[0036] For example, when the following features are detected: FFT appears sideband at 3.2kHz, temperature rises by 1.2℃ per minute, torque fluctuation entropy value >0.8; model output: failure probability =0.93, failure type = bearing outer ring damage (confidence 92%).

[0037] S4, judging the failure probability, generating a judgment result, and triggering an alarm or updating processing according to the judgment result.

[0038] Specifically, step S4 further includes: judging whether the failure probability is greater than or equal to a preset threshold, wherein the preset threshold is dynamically calculated based on a historical failure data sliding window; If yes, trigger a three-level alarm; If not, when the threshold change exceeds 5%, trigger the model fine-tuning mode for online learning and update the adaptive threshold.

[0039] In this embodiment, the fault probability output by the DTNN model is judged. The core of this judgment process is to compare the fault probability with a preset threshold, and the threshold comparison result is a Boolean signal. The preset threshold is not fixed, but is dynamically calculated based on the historical fault data sliding window (30 days) (mean + 3). This dynamic calculation method fully considers the changes of the device operating conditions and the statistical characteristics of the historical fault data, so that the threshold can adapt to the device state under different operating stages and environments, thereby improving the accuracy and reliability of the judgment. In this way, various complex operating conditions can be more flexibly coped with, and the false alarm or missed alarm problems caused by the fixed threshold can be avoided.

[0040] Among them, the adaptive threshold adjustment means that the threshold is automatically adjusted according to the device load cycle (such as start-stop and peak operation) to avoid false alarms caused by static thresholds. After the alarm event is triggered, the fault data is automatically collected and the training set is updated to realize model self-evolution.

[0041] In the judgment process, if the fault probability is greater than or equal to the preset threshold, a three-level alarm will be triggered. The three-level alarm mechanism is a hierarchical warning strategy, which divides the alarm into different levels according to the severity and urgency of the fault, so that the operator can take appropriate measures according to the alarm level. For example, a first-level alarm may indicate that the device has a potential fault risk and needs further inspection; a second-level alarm may indicate that the device has a high probability of failure and needs to take measures as soon as possible; and a third-level alarm indicates that the device has failed and needs to be shut down for repair immediately. Through this hierarchical alarm mechanism, more specific guidance can be provided to the operator, which helps to timely discover and handle faults, reduces equipment downtime, and improves production efficiency.

[0042] On the contrary, if the fault probability is less than the preset threshold, it will be further judged whether the threshold change exceeds 5%. When it is judged that the threshold change exceeds 5%, it indicates that the performance may have deviated to a certain extent, and needs to be adjusted and optimized accordingly. Therefore, the model fine-tuning mode is triggered to perform online learning and update the adaptive threshold. Online learning is a learning method that can automatically adjust and optimize the parameters of the model according to new data, which enables the model to continuously learn and adapt to new operating conditions and fault modes during operation, thereby maintaining high performance and high reliability. In this way, performance deviations caused by changes in operating conditions or model aging can be corrected in time to ensure long-term stable operation.

[0043] Please refer to Figure 4 , preferably, further comprising: based on the fault probability, the fault type classification and the judgment result, constructing a device digital twin through the Unity engine to display the fault location in real time.

[0044] Specifically, in this embodiment, after the fault probability judgment process is completed and the corresponding judgment result is generated, the key information will be further utilized for digital twin visualization; a digital twin corresponding to the actual device is built through the Unity engine. As a powerful 3D development tool, the Unity engine can realize highly realistic virtual environment and model construction, providing a solid technical foundation for the creation of device digital twins. In this process, the fault information will be accurately mapped to the corresponding position (such as bearings, gearboxes) and severity (thermal map rendering) of the digital twin according to the fault probability, fault type classification, and judgment result. For example, if it is determined that there is a risk of failure at a certain bearing position, the bearing position will be highlighted with a specific color or identifier on the digital twin, and may be accompanied by a specific value of the fault probability and a brief description of the fault type.

[0045] This real-time display of fault location provides an intuitive and efficient fault diagnosis means for operators. Operators do not need to delve into complex sensor data and diagnostic model internals, but can quickly locate the device parts that may have problems by observing the intuitive display on the digital twin. For example: high-frequency vibration + low temperature rise → bearing failure, medium-frequency vibration + high temperature rise → gear wear, life prediction basis: MHI attenuation slope is strongly correlated with remaining useful life (RUL) (R 2 >0.91). This not only greatly shortens the fault troubleshooting time, improves maintenance efficiency, but also reduces the requirement for professional skills of operators, so that more frontline workers can effectively participate in the daily maintenance work of the device.

[0046] In addition, the fault display based on the digital twin can also be combined with augmented reality (AR) technology for AR-assisted maintenance, further improving the convenience and accuracy of fault diagnosis and maintenance work. For example, through AR devices such as Microsoft HoloLens, fault diagnosis guidelines can be superimposed to guide on-site engineers to quickly locate problems; on-site engineers can directly see the digital twin fault information superimposed on the device in the physical environment of the actual device. This virtual-real combined fault diagnosis method enables engineers to more accurately locate fault locations, and also provides detailed fault analysis and maintenance suggestions, so that engineers can more efficiently complete maintenance tasks, reduce device downtime, and improve production efficiency.

[0047] Preferably, it further comprises: storing the alarm record triggered each time through the Hyperledger Fabric module to ensure that the audit trace is tamper-proof, wherein the alarm record includes time, sensor data, and processing result.

[0048] Specifically, in this embodiment, after completing the fault probability judgment process and generating the corresponding judgment result, the alarm record triggered each time is further stored through the Hyperledger Fabric module; that is, the fault event is chained. Hyperledger Fabric is an enterprise-level blockchain platform known for its high performance, scalability, and strong privacy protection features. By using the Hyperledger Fabric module, the integrity and tamper resistance of the alarm record can be ensured, which is crucial for subsequent audits and fault analysis.

[0049] The alarm record includes time, sensor data, processing results, and other key information. The timestamp records the exact time when the alarm occurs, the sensor data provides the running state of the device when the alarm is triggered, and the processing result contains the fault probability, fault type classification, and judgment result, etc. These information are integrated into an unalterable record and stored on the blockchain. Due to the distributed ledger characteristics of the blockchain, these records cannot be modified or deleted once written, thus ensuring the authenticity and reliability of the data.

[0050] This way of storing alarm records on the blockchain greatly facilitates the maintenance and audit of devices. First, it ensures the tamper resistance of the alarm record, which is crucial for responsibility tracing and accident investigation. When a device failure or safety accident occurs, maintenance personnel and auditors can rely on the alarm records stored on the blockchain to quickly and accurately determine the time, cause, and responsible person of the failure. Second, the distributed ledger characteristics of the blockchain allow the alarm record to be shared and accessed by multiple relevant parties without worrying about data consistency and integrity issues. This not only improves information transparency, but also promotes collaboration and communication between different departments.

[0051] In addition, storing alarm records through the Hyperledger Fabric module can also be combined with smart contract technology to further improve automation and intelligence. For example, when the alarm record is written to the blockchain, the smart contract can automatically trigger a series of preset operations, such as notifying the maintenance team, generating a repair work order, starting the spare parts procurement process, etc. These automated operations not only improve response speed, but also reduce human error, ensuring that the device can be maintained in a timely and effective manner.

[0052] Specifically, in this embodiment, the vibration-temperature correlation fault warning method based on multi-sensor fusion needs to emphasize three core differences: first, the cross-physical field coupling analysis of vibration-temperature-torque, the three-field coupling diagnosis paradigm of this method breaks through the traditional single-physical field limitation and establishes a fault transmission chain model of vibration-temperature-torque. Second, the MHI index fuses time-frequency domain features and temperature gradient to form an interpretable quantitative index; it realizes the interpretable fusion of multi-source heterogeneous data and provides a quantitative benchmark for predictive maintenance. Finally, the DTNN model introduces an attention mechanism to handle long-term dependencies, and the Attention-LSTM+Transformer hybrid model solves the long-term dependency problem, with a fault recognition accuracy of 98.7%. In addition, it has strong engineering practicability, edge computing deployment (<50ms delay), and has verified a false alarm rate of <1.2% in a wind turbine gearbox.

[0053] In summary, the vibration-temperature correlation fault warning method based on multi-sensor fusion innovatively integrates vibration, temperature, and torque data from three physical fields, and through advanced data processing and analysis techniques, it realizes accurate assessment of equipment health status and early warning of faults. The key technology lies in its unique data fusion method and deep learning model, which can effectively improve the accuracy and efficiency of diagnosis and provide a scientific basis for the maintenance of industrial equipment.

[0054] In terms of data processing, through preprocessing and time alignment techniques, the consistency and usability of multi-source data are ensured. Further, the MHI index is used to normalize sensor data of different dimensions, forming a unified health assessment standard. This index not only reflects the health status of the equipment in real time, but also adapts to the changes in fault characteristics under different working conditions through dynamically optimized weight coefficients.

[0055] In terms of fault diagnosis models, the DTNN-based architecture is adopted, which combines attention mechanisms and Transformer encoding layers to effectively address the shortcomings of traditional models in handling long-term dependencies. Through this model, key features in MHI time series can be automatically analyzed, and through matching with the fault pattern library, rapid identification of fault types and accurate calculation of fault probabilities can be achieved.

[0056] In addition, digital twin technology is introduced, which displays fault locations in real time through the Unity engine, providing intuitive fault diagnosis information for operators. At the same time, blockchain technology is used to store alarm records, ensuring data immutability and traceability, and providing solid technical support for equipment maintenance and auditing.

[0057] The above is the preferred embodiment of the present application, it should be pointed out that, for those skilled in the art, without departing from the principles of the present application, can also make a number of improvements and refinements, these improvements and refinements are also considered to be within the scope of the present application.

Claims

1. A vibration-temperature correlation fault early warning method based on multi-sensor fusion, characterized in that, include: Obtain the multi-source dataset collected by the preset sensor components, preprocess the multi-source dataset, and align it with the processed multi-source dataset to obtain a JSON data packet; The JSON data packets are fused based on the MHI index, feature vectors are extracted, and sensor data of different dimensions are normalized into a health index. A pre-trained DTNN model is used to perform inference processing on the feature vectors to generate fault probability and fault type classification. The probability of failure is judged, a judgment result is generated, and an alarm is triggered or updated based on the judgment result.

2. The vibration-temperature correlation fault early warning method based on multi-sensor fusion according to claim 1, characterized in that, The sensor assembly includes a vibration sensor, a temperature assembly, and a torque sensor. The temperature assembly includes a temperature sensor and a digital conversion module. The vibration sensor is a triaxial accelerometer, the temperature sensor is a PT100 resistance temperature detector, the digital conversion module is a MAX31865, and the torque sensor is an HBM T40B torque flange.

3. The vibration-temperature correlation fault early warning method based on multi-sensor fusion according to claim 2, characterized in that, Obtain the multi-source dataset collected by the preset sensor components, preprocess the multi-source dataset, and align it with the preprocessed multi-source dataset to obtain a JSON data packet, specifically: Vibration data transmitted by vibration sensor is received through SPI interface, temperature data transmitted by temperature sensor is received through I2C interface, and torque data transmitted by torque sensor is received through CAN bus to obtain multi-source dataset. The vibration data is subjected to Butterworth low-pass filtering to remove high-frequency noise, the temperature data is subjected to moving average filtering, and the transmission delay of the torque data is compensated based on linear interpolation. Based on the IEEE 1588 precision time protocol, the timestamps of all sensors in the sensor assembly are synchronized to obtain a time-aligned multidimensional array, and the multidimensional array is encapsulated into a JSON data packet according to TCP / IP technology.

4. The vibration-temperature correlation fault early warning method based on multi-sensor fusion according to claim 1, characterized in that, The formula for calculating the MHI index is as follows: ,in, For health index, Represents the vibrational energy spectral density. Indicates the rate of temperature change. This represents the standard deviation of torque fluctuation.

5. The vibration-temperature correlation fault early warning method based on multi-sensor fusion according to claim 1, characterized in that, The JSON data packets are fused based on the MHI index, feature vectors are extracted, and sensor data of different dimensions are normalized into a single health index, specifically: The FFT algorithm is used to extract vibration spectrum features from JSON data packets. Wavelet transform is then performed based on the extracted vibration spectrum features to detect transient impact signals, resulting in a 120-dimensional frequency domain / time domain feature vector in JSON format, which is then used to generate a health index. Among them, the 120-dimensional frequency domain / time domain feature vector includes spectral peak, energy entropy, and temperature gradient in JSON format. The health index MHI ranges from [0,1]. When the health index MHI is greater than 0.8, it indicates a healthy state. When the health index MHI is greater than 0.5 and less than or equal to 0.8, it indicates a warning state. When the health index MHI is less than or equal to 0.5, it indicates a fault state.

6. The vibration-temperature correlation fault early warning method based on multi-sensor fusion according to claim 1, characterized in that, A pre-trained DTNN model is used to perform inference processing on the feature vectors to generate fault probabilities and fault type classifications, specifically: The 120-dimensional feature vector is input into the Attention-LSTM layer to capture spectral envelope features, temperature gradient features, and torque entropy features. The calculation formula for the Attention-LSTM layer is as follows: Q is the query vector, K is the key vector, and V is the value vector. Scaling factor The normalization function is defined as follows: the DTNN model consists of an Attention-LSTM layer, a Transformer encoding layer, and a fully connected layer. The spectral envelope features, temperature gradient features, and torque entropy features are input into the Transformer encoding layer. Cross-physical field correlation is established through a self-attention mechanism, with the formula: vibration_feat = encoder(vib_data), temp_feat = encoder(temp_data), cross_attention = MultiHeadAttention(vibration_feat,temp_feat), where vibration_feat is the feature vector output by the encoder, encoder is the feature encoder, vib_data is the time-series data collected by the original vibration / temperature sensor, and temp_feat, temp_data, cross_attention, and MultiHeadAttention are all core components of the multi-head attention mechanism. vibration_feat and temp_feat are the calculation process of the core components of the multi-head attention mechanism. The fully connected layer outputs the fault probability using the sigmoid activation function, and the softmax layer outputs the fault type classification. The formula for calculating the fault probability is: , The probability of failure is represented by a value in the range [0,1]. is the base of the natural logarithm. This is the input value for the fully connected layer.

7. The vibration-temperature correlation fault early warning method based on multi-sensor fusion according to claim 1, characterized in that, The probability of failure is assessed, a assessment result is generated, and an alarm is triggered or updated based on the assessment result, specifically as follows: Determine whether the failure probability is greater than or equal to a preset threshold, wherein the preset threshold is dynamically calculated based on a sliding window of historical failure data; If so, trigger a level three alarm; If not, when the threshold change is detected to exceed 5%, the model fine-tuning mode is triggered to perform online learning and update the adaptive threshold.

8. The vibration-temperature correlation fault early warning method based on multi-sensor fusion according to claim 1, characterized in that, Also includes: Based on fault probability, fault type classification, and judgment results, a digital twin of the device is built using the Unity engine to display the fault location in real time.

9. The vibration-temperature correlation fault early warning method based on multi-sensor fusion according to claim 1, characterized in that, Also includes: Each triggered alarm record is stored through the Hyperledger Fabric module to ensure that audit traceability is tamper-proof. The alarm record includes the time, sensor data, and processing result.

Images