A fault diagnosis method, device, and storage medium based on event-triggered data-driven methods.

Abstract

This invention discloses a fault diagnosis method, device, and storage medium based on event-triggered data-driven methods. The steps are as follows: Step 1: Acquire data from various sensors of the target product; Step 2: Based on the acquired sensor data, perform triggered data transmission using an event-triggered data-driven mechanism; Step 3: Reconstruct the data based on the extracted trigger data and time information to obtain reconstructed data; Step 4: Divide the reconstructed data into datasets and input the divided datasets into a network model for training, obtaining a trained network model; Step 5: Diagnose the fault type of the target product based on the trained model. This invention provides a fault diagnosis method, device, and storage medium based on event-triggered data-driven methods, which can reduce data volume and communication congestion during actual product engineering communication processes, while improving or maintaining diagnostic accuracy.

Description

TECHNICAL FIELD

[0001] The application relates to an event-triggered data-driven fault diagnosis method, device and storage medium, and belongs to the technical field of rolling bearing fault diagnosis. BACKGROUND

[0002] At present, the research on data-driven fault diagnosis is mostly carried out according to a time-triggered mechanism. The product data at each sampling time needs to be sent to the diagnosis control system, but in engineering application, a large amount of data with little difference will be transmitted by using the time-triggered mechanism, and the frequent transmission of data will also cause great pressure on network communication. Especially under the background of intelligent manufacturing, engineering systems are more integrated and complex, and the communication pressure of detection equipment and the whole system is multiplied, so it is particularly important to reasonably use these hardware resources.

[0003] Traditional data-driven fault diagnosis methods such as methods based on analytical models are not suitable for the current complex and precise product engineering system environment with the characteristics of big data. Because for more complex and coupled product engineering systems, the traditional data-driven fault diagnosis method often cannot establish an effective physical model for fault analysis, detection and diagnosis.

[0004] For example, rolling bearings are widely used in rotating machinery and are common basic core components, and they also play a very important role in the safe and stable operation of mechanical equipment. Especially under the current environment of intelligent manufacturing, rotating machinery equipment based on rolling bearings requires more advanced and safe. Therefore, the processing of rolling bearing operation data and fault diagnosis has great research significance.

[0005] The event-triggered data-driven method is a "on-demand execution" triggering mechanism, which does not rely on periodic sampling, but only causes the sampling of communication equipment by judging whether the information meets certain triggering conditions. Most event-triggered technologies currently ignore the integrity of the original data information due to the non-uniform nature of the sampling mode.

[0006] Therefore, the research on the event-triggered data-driven rolling bearing fault diagnosis method has important practical significance. SUMMARY

[0007] Purpose: In order to overcome the deficiencies in the prior art, the application provides an event-triggered data-driven fault diagnosis method, device and storage medium, which effectively uses a neural network fault diagnosis model to diagnose faults, solves the problems of waste of network communication resources and hardware resources in the data transmission process of fault diagnosis and detection, and the loss of part of the system information caused by the non-uniform sampling of the event-triggered mechanism.

[0008] Technical solution: To solve the above technical problems, the technical solution adopted by the present application is:

[0009] In a first aspect, a fault diagnosis method based on event-triggered data driving includes the following steps:

[0010] Step 1: Obtain data of various sensors of the target product.

[0011] Step 2: Based on the obtained data of various sensors of the target product, trigger data transmission based on an event-triggered data driving mechanism.

[0012] Step 3: Reconstruct data based on the extracted trigger data and time information to obtain reconstructed data.

[0013] Step 4: Divide the reconstructed data into data sets and input the divided data sets into a network model for network model training to obtain a trained network model.

[0014] Step 5: Diagnose the fault type of the target product based on the trained model.

[0015] As a preferred solution, the step 2 includes:

[0016] Respectively obtain sample data of various sensors of the target product, the sample data including normal data and fault data.

[0017] Respectively extract trigger data from the obtained normal data and fault data using a first-order backward difference method.

[0018] As a preferred solution, the method for extracting trigger data using the first-order backward difference method includes:

[0019] Obtain the continuous event-triggered deviation e(t) of the normal data and the fault data, and the calculation formula is as follows:

[0020] e(t) = x(t) - x(t k ), t ∈ [t k , t k+1 ) (1)

[0021] Where: t k is the time when the sensor obtains data from the controlled object and transmits the data to the controller, t k = 0; x(t k ) and x(t) are the state of the system at the last transmission sampling time and the current state, respectively.

[0022] When the event-triggering condition is met, trigger sampling is directly performed to obtain trigger data.

[0023] The event triggering condition comprises a fixed threshold event triggering condition or a relative threshold event triggering condition.

[0024] The fixed threshold event triggering condition triggers the transmission of data when the absolute value of the continuous event triggering deviation e(t) exceeds a preset fixed threshold delta.

[0025] The relative threshold event triggering condition triggers the transmission of data when the absolute value of the continuous event triggering deviation e(t) exceeds a preset relative threshold eta relative to the absolute value of the current state x(t).

[0026] As a preferred solution, the step 3 comprises:

[0027] The data information is extracted at the triggering moment, and a zero filling operation is taken at the non-triggering moment to obtain the reconstructed data.

[0028] As a preferred solution, the network model adopts a CNN model.

[0029] As a preferred solution, the network model adopts a CNN+ResNet model.

[0030] As a preferred solution, delta=0.07.

[0031] As a preferred solution, the target product is set to a rolling bearing.

[0032] In a second aspect, a computer readable storage medium has a computer program stored thereon, and the computer program is executed by a processor to implement the event-triggered data-driven fault diagnosis method according to any one of the first aspect.

[0033] In a third aspect, a computer device comprises:

[0034] A memory is configured to store instructions.

[0035] A processor is configured to execute the instructions, so that the computer device executes the event-triggered data-driven fault diagnosis method according to any one of the first aspect.

[0036] Beneficial effects: the event-triggered data-driven fault diagnosis method, device and storage medium provided by the present application establish an event-triggered data-driven fault diagnosis model, selectively transmit data based on an event triggering mechanism, and combine a convolutional neural network and a residual network convolutional neural network with event-triggered data driving respectively, so as to realize the fusion of event-triggered data driving and deep learning network.

[0037] This invention proposes applying an event-triggered mechanism to data-driven processes. This mechanism selects which product data to transmit, ensuring only the necessary data is transmitted and guaranteeing the transmission of fault data characteristics. In actual product engineering communication, this reduces data volume, alleviates communication congestion, and simultaneously improves or maintains diagnostic accuracy. This has been thoroughly validated in applications such as rolling bearing fault diagnosis. Attached Figure Description

[0038] Figure 1 This is a schematic diagram illustrating the working principle of the fault diagnosis method based on event-triggered data-driven methods of the present invention.

[0039] Figure 2 This is a flowchart illustrating the implementation of the event-triggered data-driven fault diagnosis method of the present invention.

[0040] Figure 3 This is a comparison chart of the reconstructed data and the original data in this invention.

[0041] Figure 4 This invention is based on CNN scoring results driven by raw data.

[0042] Figure 5 This invention is based on CNN scoring results driven by event-triggered data.

[0043] Figure 6 This invention is based on the CNN+ResNet score results driven by event-triggered data. Detailed Implementation

[0044] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0045] The present invention will be further described below with reference to specific embodiments.

[0046] Example 1:

[0047] This embodiment introduces a fault diagnosis method based on event-triggered data, including the following steps:

[0048] Step 1: Obtain data from various sensors of the target product (rolling bearing).

[0049] Step 2: Based on the data from various sensors of the target product, trigger data transmission using an event-triggered data-driven mechanism.

[0050] Step 3: Reconstruct the data based on the extracted trigger data and time information to obtain the reconstructed data.

[0051] Step 4: Divide the reconstructed data into datasets and input the divided datasets into the network model for training to obtain the trained network model.

[0052] Step 5: Diagnose the fault type of the target product based on the trained model.

[0053] Furthermore, step 2 includes:

[0054] Sample data of various sensors of the target product are obtained respectively. The sample data includes normal data and fault data. The fault data includes inner ring fault data, outer ring fault data and rolling element fault data.

[0055] Trigger data is extracted from the acquired normal data and fault data using a first-order backward differential method.

[0056] Furthermore, the method for extracting trigger data using the first-order backward differential method includes:

[0057] The continuous event triggering deviation e(t) between normal data and fault data is calculated using the following formula:

[0058] e(t)=x(t)-x(t) k ),t∈[t k ,t k+1 (1)

[0059] Where: t k t is the time when the sensor acquires data from the controlled object and transmits the data to the controller. k =0; x(t) k x(t) and x(t) represent the state of the system at the previous transmission sampling time and the current state, respectively.

[0060] When the event triggering conditions are met, trigger sampling is performed directly to obtain trigger data.

[0061] The event triggering conditions include: fixed threshold event triggering conditions or relative threshold event triggering conditions.

[0062] The fixed threshold event triggering condition is based on a set fixed threshold δ. Specifically, data transmission is triggered when the absolute value of the continuous event triggering deviation e(t) exceeds the preset fixed threshold δ.

[0063] The relative threshold event triggering condition: This triggering condition is based on the degree of change relative to the current state to determine event triggering. Specifically, when the absolute value of the continuous event triggering deviation e(t) relative to the absolute value of the current state x(t) exceeds a preset relative threshold η, data transmission is triggered.

[0064] Where δ>0 is a given constant.

[0065] Furthermore, step 3 includes: extracting data information at the trigger time and performing zero-padding operation at the non-trigger time to obtain the reconstructed data.

[0066] Furthermore, the network model used is a CNN model.

[0067] Furthermore, the network model adopted is a CNN+ResNet model.

[0068] Furthermore, δ = 0.07.

[0069] Example 2:

[0070] This embodiment discloses a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements a fault diagnosis method based on event-triggered data-driven methods as described in any of Embodiment 1.

[0071] Example 3:

[0072] This embodiment describes a computer device, including:

[0073] Memory is used to store instructions.

[0074] A processor is configured to execute the instructions, causing the computer device to perform an event-triggered data-driven fault diagnosis method as described in any of Embodiment 1.

[0075] Example 4:

[0076] This embodiment uses vibration data of deep groove ball bearing 6205-RS JEM SKF from the publicly available dataset of Case Western Reserve University for verification and analysis. The data includes a CPU processor, a 1.5kW motor, a torque sensor, a decoder, a power tester, and an acceleration sensor, with the acceleration sensor mounted on the base housing.

[0077] like Figure 1As shown, this invention provides a fault diagnosis method based on event-triggered data-driven approaches, consisting of an event-triggered data-driven module and a neural network fault diagnosis module. An event-triggered data acquisition mechanism is established, allowing for on-demand sampling, reducing the amount of data transmitted and saving network resources. Data after the event is triggered is used, and the data is reconstructed based on the acquisition time information for verification of roller fault data. Then, based on deep learning, fault diagnosis models are established using a convolutional neural network (CNN) and a CNN+ResNet fault diagnosis model. The CNN consists of an input layer, convolutional layers, activation functions, pooling layers, and fully connected layers, used for feature classification. Introducing a residual network increases the number of network layers and solves the gradient explosion problem, improving fault diagnosis accuracy. After event-triggered data acquisition, the traditional CNN model reduces the amount of transmitted data and improves diagnostic accuracy. Compared to CNN, the model combined with a residual network (CNN+ResNet) reduces data transmission while maintaining high accuracy in fault diagnosis after event-triggered data acquisition. The specific process is as follows... Figure 2 As shown.

[0078] The method of the invention includes the following steps:

[0079] Step 1: According to Figure 1 The model shown is an event-triggered, data-driven fault diagnosis model. This model includes two modules: event triggering and neural network fault diagnosis.

[0080] The event-triggered module selectively transmits collected data according to the event-triggered mechanism, reducing the amount of data that needs to be transmitted. This avoids communication overhead during actual communication. Furthermore, in the experimental simulation, the data following the event-triggered module is reconstructed to include the original event information. This avoids the data loss issues caused by the non-uniform sampling mode of event-triggered systems.

[0081] The neural network fault diagnosis module, based on the neural network fault diagnosis method, can accurately handle fault diagnosis problems without the need for a precise mathematical model. This is because it learns itself according to set criteria by mapping the input and output relationships of data samples.

[0082] Step 2: Acquire roller vibration data from vibration sensors and execute an event-triggered mechanism before inputting it into the convolutional neural network model. Useful data is transmitted on demand using a first-order backward difference method.

[0083] In event-triggered data-driven processing, the first-order backward difference method is applied to the processing of roller vibration data. Specifically, by acquiring data from vibration sensors, a continuous time series of roller vibration signals can be obtained. Then, the first-order backward difference method is used to calculate the signal changes between adjacent time points. This allows the extraction of the changing trends and characteristics of the roller vibration signals, thereby determining whether the event triggering conditions are met.

[0084] By using the first-order backward difference method, continuous vibration data can be transformed into a discrete sequence of changes, which can then be used to determine the event triggering mechanism. In the event triggering conditions, certain thresholds or conditions can be set; when the change in the roller vibration signal exceeds or meets specific conditions, the transmission or processing of the event is triggered.

[0085] The continuous event triggering deviation e(t) is defined as follows:

[0086] e(t)=x(t)-x(t) k ),t∈[t k ,t k+1 (1)

[0087] Where: t k t is the time when the sensor acquires data from the controlled object and transmits the data to the controller. k =0; x(t) k x(t) and x(t) represent the state of the system at the previous transmission sampling time and the current state, respectively.

[0088] Based on the real-time measured state x(t), e(t) can be calculated online, and when the event triggering condition is met, sampling can be directly triggered.

[0089] In event triggering, triggering conditions are used to determine when to trigger the transmission and processing of an event. Fixed threshold event triggering conditions and relative threshold event triggering conditions are two common types of triggering conditions, as shown in the following formula.

[0090] Fixed threshold event trigger condition: This trigger condition is based on a set fixed threshold δ. Data transmission is triggered when the absolute value of the continuous event trigger deviation e(t) exceeds the preset fixed threshold.

[0091] Relative threshold event triggering condition: This triggering condition is based on the degree of change relative to the current state to determine event triggering. Specifically, when the absolute value of the continuous event triggering deviation e(t) relative to the absolute value of the current state x(t) exceeds a preset relative threshold η, data transmission is triggered.

[0092] Where δ > 0 is a given constant. Whether the current data can be transmitted is determined by the event triggering scheme.

[0093] By combining the first-order backward difference method with a fixed threshold event triggering condition or a relative threshold event triggering condition, an event triggering mechanism based on state change can be implemented. The first-order backward difference method is used to calculate the amount of state change, while the triggering condition is used to determine whether this change satisfies the condition for triggering an event.

[0094] This combination can effectively reduce the amount of data transmitted when an event is triggered, and only trigger the transmission and processing of the event when the state change exceeds a set threshold, thereby greatly saving communication resources.

[0095] In the event triggering mechanism, it is essential to ensure that the minimum time interval between any two event triggering moments is strictly greater than zero during transmission. If the minimum event interval is zero, the Zeno phenomenon, where an infinite number of events occur within a finite time, will occur.

[0096] The Zeno phenomenon refers to the phenomenon where, in a continuous event triggering mechanism, if no minimum event interval is set, an infinite number of events may occur within a finite time period. This can cause the system to be unable to effectively process and transmit large amounts of event data, resulting in system overload and resource waste.

[0097] By using a first-order backward difference method combined with event triggering conditions, the conditions for event triggering can be restricted, ensuring that event transmission and processing are only triggered when the state change exceeds a set threshold. This restriction ensures a minimum time interval between events, thereby preventing Zeno's phenomenon from occurring.

[0098] Therefore, by using the first-order backward difference method in combination with fixed threshold event triggering conditions or relative threshold event triggering conditions, Zeno's phenomenon can be effectively avoided, and the system can be ensured to trigger events and transmit data under reasonable conditions.

[0099] Step 3: Reconstruct the data output from Step 2 by combining it with time information to generate data with enhanced features and complete time information. This data will be input into the convolutional neural network model.

[0100] Traditional event-triggered mechanisms suffer from information loss due to non-uniform sampling patterns during data transmission. This leads to discrepancies between data at non-event times and the actual data, rendering fault diagnosis methods ineffective. To address this, the transmitted trigger data can be reconstructed. Since this reconstruction incorporates the time information of the original data, the reconstructed data uses the non-trigger time data as the origin, while the trigger time data also includes the time information of the original data.

[0101] like Figure 3As shown, the reconstructed data is compared with the original data. Data information is extracted at the trigger time, and zero-padding is performed at non-trigger time. Combined with the time information of the original data, the time information of the non-trigger time is preserved in the reconstructed data, thus ensuring that the reconstructed data retains the information of the original system. This also makes the data features more prominent, facilitating neural network models to learn from them and improving the accuracy of fault diagnosis.

[0102] Step 4: Perform data preprocessing, dividing the data into a 70% training set and a 30% test set. After normalizing the data, input it into a convolutional neural network model. The model is trained through convolutional layers for feature extraction, pooling layers for dimensionality reduction, and fully connected layers for classification, resulting in a model with optimal parameters.

[0103] Convolutional Neural Networks (CNNs), as representative networks of deep learning, are extensions of perceptrons with more hidden layers, hence they are also called Multi-Layer Perceptrons (MLPs).

[0104] Neural networks use the sum of squared errors and derivatives to adjust weights and thresholds to minimize error. The training process begins by computing the output of each layer until the network's output vector is obtained. The error vector is obtained by calculating the difference between the target vector and the network output. Then, the multilayer perceptron (MLP) is trained using gradient descent and backpropagation algorithms.

[0105] Gradient descent is a classic optimization algorithm in neural networks, also known as the steepest descent method, and is widely used in various optimization problems. In gradient descent, the gradient of the function f(θ) with respect to the independent variable θ is calculated. This is used to update the independent variable in order to gradually approach the minimum point of the function.

[0106]

[0107] The first-order Taylor expansion of f(θ) is shown below:

[0108]

[0109] The formula is then rewritten as:

[0110]

[0111] Where θ = (x, y) and θ0 = (x0, y0) are denoted as (θ - θ0) and denoted as μv, where μ is a positive number and v represents the unit vector of θ - θ0.

[0112] then:

[0113]

[0114] Since μ is a positive number, it is necessary to Right now:

[0115]

[0116] ||v||and The vector magnitude is positive, therefore cosθ(α) < 0.

[0117] Right now

[0118]

[0119] Will Let η be the equation, then the equation is updated as follows:

[0120]

[0121] Where η is the step size.

[0122] Step size (learning rate) is an important hyperparameter. Choosing too large a step size may cause the algorithm to diverge and fail to converge; while choosing too small a step size may result in a slow convergence speed, requiring more iterations to reach the optimal solution.

[0123] Gradient descent algorithms require a trade-off in step size selection, which typically necessitates tuning based on the specific problem and dataset characteristics. A common approach is to experiment with different step size values, observe the algorithm's convergence and performance, and ultimately choose an appropriate step size to achieve better results.

[0124] Backpropagation is a commonly used method in training deep neural networks. It calculates the gradient of the loss function corresponding to the weights and biases of all neurons in the network, and then uses optimization algorithms (such as gradient descent) to update the weights of the neural network, thereby minimizing the objective function.

[0125] The formula for updating parameters using gradient descent is as follows:

[0126]

[0127]

[0128] Among them, W (l) and b (l) Let represent the matrices of weights and biases for the l-th layer, respectively. As can be seen from formulas (9) and (10), the key to backpropagation is to calculate the partial derivatives of the weights and biases of each layer with respect to the loss function.

[0129] The backpropagation algorithm propagates the error from the network's output layer backward, calculating the gradient layer by layer, and then using the chain rule to pass the gradient back to each layer. This allows for efficient calculation of the gradient of each parameter with respect to the loss function, thus enabling parameter updates.

[0130] Through backpropagation, neural networks can iterate repeatedly based on training data, continuously adjusting weights and biases to minimize the loss function, thereby improving the network's performance and accuracy.

[0131] When designing neural networks for fault diagnosis, choosing the appropriate network type is crucial and should be based on the characteristics of the problem and the system. We chose a Convolutional Neural Network (CNN) as our fault diagnosis model, as it can automatically extract features and effectively process input data through convolutional and pooling layers.

[0132] In the early stages of the fault diagnosis task, data preprocessing is performed to normalize the input data and scale it to a specific range to ensure that the numerical differences between different features do not affect the performance of the model.

[0133] Next is the feature extraction part, which is the core of the entire fault diagnosis model. We use a deep learning model to extract features from the data collected by the event-triggered module. This data has been reconstructed and contains the temporal information of the original data. Through the stacking of convolutional neural network layers, we are able to automatically learn and extract effective features from the input data. Convolutional layers can capture local patterns and features in the data, while pooling layers help reduce the dimensionality of the data and retain the most important features.

[0134] Finally, we use the SoftMax classifier to classify and diagnose faults based on the extracted features. The SoftMax classifier maps feature signals to probability distributions for different fault categories. By interpreting these probability distributions, we can determine which fault category the input data belongs to and arrive at the final diagnostic result.

[0135] Step 5: Evaluate the model obtained in Step 4 using the test set. For fault diagnosis, use accuracy and F1-Score metrics to analyze the model's actual performance. The F1-Score output will be displayed row by row by category, providing the accuracy, recall, F1 score, and support for each category. Additionally, macro-averaging and weighted-averaging metrics will be calculated and displayed to summarize the performance evaluation across all categories. For event-triggered data-driven CNN score results... Figure 5 As shown. For ease of comparison,Figure 4 This is the CNN score result based on the raw data.

[0136] Figure 4 As shown, the CNN model's fault diagnosis accuracy for the original data is 83.3%, and the accuracy for diagnosing the four operating states is as follows: normal state 100%, inner ring fault 71%, rolling element fault 76%, and outer ring fault 89%.

[0137] Figure 5 As shown, the CNN score results based on event-triggered data-driven analysis show an accuracy score of 90.5%, with the following accuracy rates for diagnosing four operating states: normal state 98%, inner race fault 85%, rolling element fault 86%, and outer race fault 93%.

[0138] As can be seen, the overall fault diagnosis accuracy increased by 7.2% in the results of the event-triggered data-driven CNN. Furthermore, the diagnostic accuracy increased for all three different fault types.

[0139] Step Six: Modify the event triggering method, try different differential values, and then repeat steps Three, Four, and Five. Finally, select the event triggering method that performs best as the final solution for fault diagnosis.

[0140] This invention tested eight different first-order backward differential triggering methods with a fixed threshold δ. Each different first-order backward differential triggering method corresponds to a different amount of data, a reduced data rate, and the accuracy of the CNN at that time. The details are shown in Table 1.

[0141] Table 1: CNN accuracy and data reduction results for different triggering methods

[0142]

[0143] As shown in Table 1, as the first-order backward difference increases from 0.01 to 0.08, the amount of data after the event is triggered gradually decreases, while the data reduction rate during communication transmission gradually increases to 62.36%. The fault diagnosis accuracy of the CNN model does not show a significant change. Considering both the data reduction rate and the CNN accuracy, we choose a first-order backward difference of 0.07 as the event triggering method.

[0144] Step 7: Introduce a residual network into the existing convolutional neural network model to perform fault diagnosis on the original data and obtain the results. Then, execute the above steps under the optimal event triggering mode to obtain and analyze the results.

[0145] For the CNN model using the first-order backward difference obtained in step six, we also analyzed the results of fault diagnosis using both the original data and the data after event triggering and data reconstruction. The results also correspond to information such as data volume and data reduction rate.

[0146] Table 2 shows the accuracy results of the CNN and CNN+ResNet models before and after event triggering.

[0147] Table 2: Accuracy Results of CNN and CNN+ResNet Models Before and After Event Triggering

[0148]

[0149] As shown in Table 2, the amount of data that needs to be transmitted during communication by the Convolutional Neural Network (CNN) decreased from 12,130,000 to 4,886,962, a reduction of 59.71% in communication data transmission. Simultaneously, the accuracy of fault diagnosis increased from 83.3% to 90.5%, an improvement of 7.2%.

[0150] The convolutional neural network with residual network (CNN+ResNet) achieved a fault diagnosis accuracy of 98.9% on the original dataset, while maintaining the same 98.9% performance on the event-triggered data-driven mode. However, the amount of data to be transmitted during communication was reduced from 12,130,000 to 4,886,962, a reduction of 59.71%.

[0151] As can be seen, even after reducing the amount of data requiring communication transmission by 59.71%, the convolutional neural network (CNN+ResNet) with residual network still maintains a very high accuracy. Simultaneously, combined with... Figure 6 The CNN+ResNet score results based on event-triggered data-driven analysis showed an accuracy of 98.9%, with the following accuracy rates for diagnosing four operating states: normal state 100%, inner race fault 100%, rolling element fault 96%, and outer race fault 100%.

[0152] This is due to the introduction of residual networks as an enhancement to convolutional neural networks. Residual networks can effectively solve the problems of vanishing gradients and model degradation, enhancing the model's expressive and generalization abilities. By incorporating residual networks, the model may be able to learn and extract features better, thereby improving the accuracy of fault diagnosis.

[0153] This paper analyzes the effects of the event-triggered mechanism before and after its implementation, and explains the reasons for these effects. Firstly, the differential approach of the event-triggered mechanism, during the on-demand transmission of useful data, utilizes the local properties and correlations of signals, retaining only the characteristic information useful for fault diagnosis, thereby reducing the transmission of redundant data. Furthermore, the output data is reconstructed by combining it with time information. This process, through appropriate timestamp marking and data resampling techniques, makes the generated data more complete and has a higher information density. Therefore, while reducing the amount of data, it retains important time information, which helps improve the accuracy of fault diagnosis.

[0154] This invention discloses a fault diagnosis method based on event-triggered data-driven approach. It is a fault diagnosis method that combines a deep learning model based on convolutional neural networks and residual networks with an event-triggered data-driven mechanism, and has been verified in the fault diagnosis of rolling bearings.

[0155] An event-triggered data-driven mechanism is constructed, leveraging the "on-demand sampling" characteristic of event triggering to simulate the fault diagnosis process in actual engineering. During data sampling, data is collected not according to a fixed time period but based on the event-triggered mechanism. This effectively reduces the amount of data transmitted during communication in actual engineering, saving network communication resources. In the verification process using roller fault data, the data after the event trigger is reconstructed based on the collection time information and fed into the neural network model for fault diagnosis.

[0156] Next, a deep learning-based neural network fault diagnosis model, namely a convolutional neural network (CNN), is established after the event-triggered mechanism. The CNN consists of an input layer, convolutional layers, activation functions, pooling layers, and fully connected layers. Fault classification is based on the classification results of the fully connected layers on the convolutional features. Then, a residual network is introduced into the CNN model (CNN+ResNet), which deepens the network layers and solves the gradient explosion problem, thus achieving higher accuracy in fault diagnosis. For traditional CNNs, using event-triggered data-driven methods reduces the amount of data transmitted while improving fault diagnosis accuracy. For the model combining CNNs and residual networks (CNN+ResNet), using event-triggered data-driven methods reduces the amount of data transmitted while maintaining almost the same high accuracy in fault diagnosis.

[0157] This method solves the problem of transmitting unnecessary information during fault diagnosis data transmission. It is a rolling bearing fault diagnosis method that improves resource utilization and effectiveness under limited hardware and network resources. It includes: establishing a deep learning model based on a neural network, a convolutional neural network (CNN), automatically mining features from the data, transforming them layer by layer, and classifying the output sequence to achieve fault diagnosis; simultaneously considering the limited communication resources and the existence of a large amount of redundant data in the fault diagnosis system, an event-triggered mechanism is applied to the acquisition and transmission process, greatly alleviating the system communication pressure and improving system operating efficiency without changing channel resources. This invention utilizes an event-triggered mechanism combined with artificial intelligence methods to achieve the effect of improving the model's fault diagnosis accuracy while reducing data transmission volume.

[0158] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An event-triggered data-driven fault diagnosis method, characterized in that: Step 1: obtaining data of various sensors of a target product; Step 2: transmitting data based on an event-triggered data-driven mechanism according to the obtained data of various sensors of the target product; Step 3: reconstructing data according to the extracted trigger data and time information to obtain reconstructed data; Step 4: dividing the reconstructed data into data sets and inputting the divided data sets into a network model for network model training to obtain a trained network model; Step 5: diagnosing the fault type of the target product according to the trained model; the step 2 comprises: obtaining sample data of various sensors of the target product, the sample data including normal data and fault data, and the fault data including inner ring fault data, outer ring fault data and rolling element fault data; extracting trigger data from the obtained normal data and fault data by using a first-order backward difference method; the first-order backward difference method for extracting trigger data comprises: obtaining a continuous event-triggered deviation e(t) of the normal data and the fault data, and the calculation formula is as follows: when the event-triggering condition is met, trigger data is directly obtained by triggering sampling; the event-triggering condition includes a fixed threshold event-triggering condition or a relative threshold event-triggering condition; the fixed threshold event-triggering condition is that when the absolute value of the continuous event-triggered deviation e(t) exceeds a preset fixed threshold δ, the transmission of the trigger data is triggered; the relative threshold event-triggering condition is that when the absolute value of the continuous event-triggered deviation e(t) exceeds a preset relative threshold η relative to the absolute value of the current state x(t), the transmission of the trigger data is triggered; the network model adopts a CNN+ResNet model; the network model training specifically comprises: dividing the trigger data into 70% training sets and 30% test sets; after normalizing the trigger data, inputting the normalized trigger data into a convolutional neural network model; training the convolutional neural network model through the processes of feature extraction by a convolutional layer, dimension reduction by a pooling layer and classification by a fully connected layer to obtain a convolutional neural network model with optimal parameters; the convolutional neural network model uses error sum of squares and derivatives to adjust weights and thresholds to minimize errors; the training process starts from calculating the output of each layer until the output vector of the network is obtained; by calculating the difference between the target vector and the network output, an error vector is obtained; then, the multilayer perceptron is trained using the gradient descent and back propagation algorithms; the test set is used to evaluate the convolutional neural network model, and the accuracy and F1-Score evaluation indicators are used in fault diagnosis to analyze the actual effect of the convolutional neural network model; the step 3 comprises: extracting data information at the triggering time and adopting a zero-filling operation at the non-triggering time to obtain the reconstructed data; the δ=0.07; a computer program is stored thereon, which is executed by a processor to implement the event-triggered data-driven fault diagnosis method according to any one of claims 1-3; comprising: ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ e(t) = x(t) - x(t k ), t e [t k ,t k+1 ) (1) where: t k is the time at which the sensor obtains data from the plant and transmits the data to the controller, t k = 0; x(t k ) and x(t) are the state of the system at the last transmission sample time and the current state, respectively. ​ ​ ​ ​ ​ ​ ​ ​ ​ 2. The event-triggered data-driven fault diagnosis method according to claim 1, characterized in that: ​ ​ 3. The event-triggered data-driven fault diagnosis method according to claim 2, wherein: ​ 4. A computer-readable storage medium, characterized in that: ​ 5. A computer device, characterized by: ​ a memory for storing instructions; a processor for executing the instructions to cause the computer device to perform an event-triggered data-driven fault diagnosis method as claimed in any one of claims 1-3.

Images