Collaborative fault diagnosis design method capable of realizing high-frequency data transmission and adaptive precision

Abstract

The present application relates to a collaborative fault diagnosis design method capable of realizing high-frequency data transmission and adaptive precision. The method comprises: an edge gateway collecting data and performing prediction on the data, and sending to a cloud server an actual value, the deviation of which from a predicted value exceeds a threshold value; and the cloud server restoring the data on the basis of the actual value and a predicted value from the cloud server, and outputting a fault diagnosis result on the basis of the restored data and a DPS length. The present application ensures the accuracy of data transmission and fault diagnosis, thus implementing accurate and real-time fault diagnosis with limited bandwidth resources.

Description

A collaborative design method of high-frequency data transmission and precision adaptive fault diagnosis TECHNICAL FIELD

[0001] The present application relates to the field of online fault diagnosis, and in particular to a collaborative design method of high-frequency data transmission and precision adaptive fault diagnosis. BACKGROUND

[0002] With the increasing requirements for equipment safety and economy, online fault diagnosis technology has gradually become an indispensable technology in modern industry. In recent years, with the vigorous development of big data and artificial intelligence, fault diagnosis technology has shifted from expert-oriented to data-driven. However, data-driven online fault diagnosis technology requires continuous large-scale high-frequency data transmission. The sampling frequency of high-frequency data such as vibration signals is at least 1kHz, and their real-time transmission will consume unbearable bandwidth resources, resulting in serious network congestion and high network delay. At the same time, the quality of data will directly affect the accuracy of fault diagnosis. How to effectively reduce the real-time high-frequency data transmission volume without reducing the accuracy of fault diagnosis is becoming a bottleneck for fault diagnosis technology in industry practice.

[0003] In recent years, people have made a lot of efforts to reduce online data transmission, and most of these works can only be applied to simple data, and the availability of data cannot be guaranteed. The dual prediction scheme (DPS) can reduce the real-time data transmission volume while ensuring the deterministic range of data accuracy. Compared with the DPS based on traditional prediction models, the DPS based on deep learning (DL) prediction models can deal with more complex data and obtain higher data prediction accuracy and transmission reduction rate, and they are becoming mainstream.

[0004] Recently, many works have applied DL to the field of fault diagnosis and focused on solving practical problems in industrial scenarios. Yu et al. proposed a DL model based on transfer learning for diagnosing mechanical faults of unlabeled samples without known label space. Chai et al. proposed a multi-source compact transfer network for fault diagnosis problems with inconsistent fields and categories. Jiao et al. proposed a passive adaptive diagnosis for rotating machinery, which can use unmarked data to continuously optimize and update the fault diagnosis model. Shao et al. proposed a dual-threshold attention-guided generative adversarial network, which provides a portable data generation solution for fault diagnosis with limited samples. These methods still have certain deficiencies and do not consider the impact of data transmission on online fault diagnosis.

[0005] In addition, in order to achieve a balance between data transmission volume and fault diagnosis accuracy, application-oriented design to improve transmission efficiency has become a new direction of communication research. Currently, application-oriented design to improve transmission efficiency has become a new direction of communication research. Semantic-oriented, task-oriented and goal-oriented communication have attracted widespread attention from academia and industry. Feng et al. proposed a goal-oriented bandwidth allocation framework. By optimizing the information utility gain of application transmission data, the performance of typical applications in networked physical systems is effectively improved. Wang et al. proposed a semantic communication framework for text data transmission. By extracting and transmitting semantic information modeled by a knowledge graph, the data transmission volume is effectively reduced while ensuring the semantic similarity of text data. Yang et al. proposed an edge-driven semantic extraction scheme that utilizes edge intelligence to meet the computing, storage and communication requirements of semantic communication.

[0006] A Chinese invention patent application with publication number CN114640695A discloses an intelligent factory high-frequency time series data effective transmission method based on long sequence dual prediction and Informer. The method first builds a cloud-edge collaborative long sequence dual prediction architecture, then deploys the trained long sequence prediction model on the edge gateway and cloud server of the architecture, and finally uses the long sequence dual prediction method (L-DPS) to reduce the transmission volume of high-frequency data online while ensuring data accuracy. The structure of the traditional dual prediction method is improved, and the inference times of the prediction model are reduced through long sequence prediction, thereby greatly improving the applicability frequency of the traditional method, making it possible to reduce the transmission volume of high-frequency data required in the intelligent manufacturing process. At the same time, the latest deep learning model Informer is introduced and combined to solve the problems of gradient vanishing and model inference time explosion caused by long sequence prediction, thereby further improving the transmission volume reduction ratio and applicability frequency of the proposed long sequence dual prediction method.

[0007] Existing online fault diagnosis methods require real-time and reliable transmission of high-frequency signals such as vibration. How to achieve accurate and real-time fault diagnosis under limited bandwidth is a practical, important and key problem in the field of fault diagnosis.

[0008] Therefore, those skilled in the art are committed to developing a high-frequency data transmission and precision adaptive fault diagnosis collaborative design method. SUMMARY

[0009] In view of the above defects of the prior art, the technical problem to be solved by the present application is how to achieve accurate and real-time fault diagnosis under limited bandwidth.

[0010] To achieve the above object, the application provides a collaborative design method of high-frequency data transmission and precision adaptive fault diagnosis, which comprises the following steps:

[0011] S101: The edge gateway collects high-frequency data, pre-processes the high-frequency data, writes the pre-processed high-frequency data into a data set, and sends the data set to a cloud server;

[0012] S102: The cloud server receives the data set sent by the edge gateway, and trains an Informer long sequence prediction model using the data set;

[0013] S103: The cloud server and the edge gateway synchronously load the trained Informer long sequence prediction model;

[0014] S104: The edge gateway runs a long sequence double prediction method, predicts the high-frequency data, calculates the deviation of the predicted value and the actual value, and sends the actual value whose deviation exceeds a threshold to the cloud server;

[0015] S105: The cloud server runs the long sequence double prediction method, predicts the high-frequency data, and restores the high-frequency data according to the received data;

[0016] S106: The cloud server inputs the restored high-frequency data and the corresponding DPS length of the high-frequency data into a precision adaptive fault diagnosis model, and the precision adaptive fault diagnosis model outputs a fault diagnosis result online.

[0017] Further, in the step S101, the pre-processing of the high-frequency data by the edge gateway comprises time sequence conversion and normalization.

[0018] Further, the time sequence conversion is to add a time stamp online for each sampling value of the high-frequency data to mark the collection time of the high-frequency data.

[0019] Further, the time stamp comprises year, month, day, hour, minute, second and millisecond information, and the format of the time stamp is “yyyy-MM-dd HH:mm:ss.SSS”.

[0020] Further, the normalization is to normalize the numerical value of the high-frequency data to a value between -1 and 1.

[0021] Further, the normalization is completed by using an inverse tangent function normalization method, and the inverse tangent function normalization is processed using the following function:

[0022] In the formula, x iFor normalized values, x is the original value.

[0023] Further, in the step S102, the cloud server divides the data set into a training set, a validation set and a test set after receiving the data set sent by the edge gateway.

[0024] Further, in the step S102, the cloud server training the Informer long sequence prediction model with the data set includes the following sub-steps:

[0025] S1021: The cloud server trains the Informer long sequence prediction model with the training set;

[0026] S1022: The cloud server adjusts the hyperparameters of the Informer long sequence prediction model with the validation set;

[0027] S1023: The cloud server verifies the effect on the test set to obtain an optimal prediction model;

[0028] S1024: The cloud server saves the optimal model obtained by training as a file.

[0029] Further, in the step S103, the cloud server sends the optimal prediction model of the Informer long sequence prediction model to the edge gateway, and synchronously loads the optimal prediction model with the edge gateway.

[0030] Further, the step S104 includes the following sub-steps:

[0031] S1041: The edge gateway collects the high-frequency data and pre-processes the high-frequency data to obtain the actual value of the high-frequency data;

[0032] S1042: The edge gateway predicts the high-frequency data with the Informer long sequence prediction model to obtain a first predicted value of the high-frequency data;

[0033] S1043: The edge gateway calculates the deviation between the first predicted value and the actual value, and determines whether the deviation exceeds the threshold value;

[0034] S1044: If the deviation exceeds the threshold value, the edge gateway sends the actual value of the high-frequency data;

[0035] S1045: The edge gateway restores the high-frequency data with the actual value and the first predicted value.

[0036] Further, the step S105 comprises the following sub-steps:

[0037] S1051: The cloud server uses the Informer long sequence prediction model to predict the high-frequency data, to obtain a second predicted value of the high-frequency data;

[0038] S1052: The cloud server receives the actual value of the high-frequency data sent by the edge gateway;

[0039] S1053: The cloud server restores the high-frequency data using the actual value and the second predicted value.

[0040] Further, in the step S106, the precision adaptive fault diagnosis model is based on any one of the architectures of one-dimensional convolutional neural network, transformer, recurrent neural network and multi-layer perception.

[0041] Further, the precision adaptive fault diagnosis model is based on the one-dimensional convolutional neural network architecture, which comprises block stacking, global average pooling layer and full connection layer, and is configured to be able to complete extraction of local features and global features of a vibration signal.

[0042] Further, the block stacking is stacked by a plurality of blocks, each of the plurality of blocks comprising two convolutional layers and a maximum pooling layer.

[0043] Further, the two convolutional layers are used to extract the local features of the vibration signal, and the convolutional layer is defined as:

[0044] In the formula, is a tensor, W i is a convolutional filter, b i is a first bias term, ReLU(·) is a first activation function, i, i+1 is a layer number, t is a time step, and Conv1d is a convolutional layer.

[0045] Further, the maximum pooling layer is used to extract high-level features of the vibration signal, and the maximum pooling layer is defined as follows:

[0046] In the formula, MaxPool is a maximum pooling layer, max is a maximum value, i, i+1 is a layer number, t is a time step, is a tensor.

[0047] Further, the global average pooling layer is used to receive the high-level features output by the block stacking to extract the global features of the vibration signal, and the global average pooling layer is defined as follows:

[0048] In the formula, GlobalAvgPool is a global average pooling layer, j is a serial number, n is the length of model input data, i and i+1 are layer serial numbers, t is a time step, is a tensor.

[0049] Further, the full connection layer is configured to receive the global features output by the global average pooling layer, and the full connection layer is configured to output fault labels, and the full connection layer is defined as follows:

[0050] In the formula, FC is a full connection layer, W i is a weight matrix, b i is a second bias term, Softmax(·) is a second activation function, i and i+1 are layer serial numbers, t is a time step, is a tensor.

[0051] Further, the step S106 includes the following sub-steps:

[0052] S1061: The cloud server inputs the restored high-frequency data and the DPS length corresponding to the high-frequency data to the precision adaptive fault diagnosis model.

[0053] S1062: The block stacking extracts the local features and the high-level features of the vibration signal.

[0054] S1063: The global average pooling layer extracts the global features of the vibration signal according to the high-level features.

[0055] S1064: The full connection layer obtains the fault labels of the vibration signal according to the global features, and outputs a fault diagnosis result.

[0056] Further, the DPS length corresponding to the high-frequency data is configured according to the frequency of the high-frequency data.

[0057] Compared with the prior art, the present application has the following beneficial effects:

[0058] 1. Through the collaborative design of the real-time compression transmission algorithm of high-frequency data and the online fault diagnosis application supported thereby, the present application can achieve a balance between reducing data transmission amount and ensuring fault diagnosis accuracy, and can simultaneously ensure the accuracy of data transmission and the accuracy of fault diagnosis while realizing real-time fault diagnosis under limited bandwidth resources.

[0059] 2、The application uses the prior information of the DPS length as the precision reference of the transmission restoration data, and the model can adapt to the transmission restoration data of different precision, thereby improving the accuracy of fault diagnosis.

[0060] 3、The application balances the relative difference between the high-frequency vibration signal values through the non-linear arctangent function normalization processing, so that the samples with relatively small values can also be effectively trained, thereby effectively improving the accuracy of time series prediction in the real-time compression transmission algorithm of high-frequency data.

[0061] The concept, specific structure and technical effects of the application will be further described below in combination with the drawings, so as to fully understand the purpose, features and effects of the application. BRIEF DESCRIPTION OF DRAWINGS

[0062] Fig. 1 is a flowchart of the collaborative design method of the embodiment of the application;

[0063] Fig. 2 is a schematic diagram of the data transmission and fault diagnosis collaborative design architecture of the embodiment of the application;

[0064] Fig. 3 is a schematic diagram of the precision adaptive fault diagnosis model of the embodiment of the application;

[0065] Fig. 4 is a schematic diagram of the restoration data precision when the DPS length is 10 in the embodiment of the application;

[0066] Fig. 5 is a schematic diagram of the restoration data precision when the DPS length is 50 in the embodiment of the application. DETAILED DESCRIPTION

[0067] The following describes the preferred embodiments of the application with reference to the drawings, so that the technical content of the application is clearer and easier to understand. The application can be embodied in many different forms, and the protection scope of the application is not limited to the embodiments mentioned herein.

[0068] In the drawings, the same components have the same reference numerals, and components with similar structures or functions have similar reference numerals. The size and thickness of each component shown in the drawings are arbitrarily shown, and the size and thickness of each component are not limited in the application. In order to make the drawing clearer, the thickness of some components is appropriately exaggerated in some places.

[0069] The existing online fault diagnosis method needs real-time and reliable transmission of high-frequency signals such as vibration, but limited bandwidth resources are difficult to meet the requirements of high frequency and real-time. In order to achieve the balance between data transmission amount and fault diagnosis accuracy, the Chinese invention patent application with publication number CN114640695A discloses an effective transmission method for high-frequency time series data of intelligent factory based on long sequence double prediction L-DPS and Informer. The method first builds a cloud-edge collaborative long sequence double prediction architecture, then deploys the trained long sequence prediction model on the edge gateway and cloud server of the architecture, and finally uses the long sequence double prediction method L-DPS to reduce the transmission amount of high-frequency data online and ensure the accuracy of the data. The structure of the traditional double prediction method is improved, and the inference times of the prediction model are reduced through long sequence prediction, thereby greatly improving the applicable frequency of the traditional method, making it possible to reduce the transmission amount of high-frequency data required in the intelligent manufacturing process. At the same time, the latest deep learning model Informer is introduced and combined to solve the problems of gradient disappearance and model inference time explosion caused by long sequence prediction, thereby further improving the transmission amount reduction ratio and applicable frequency of the proposed long sequence double prediction method.

[0070] In view of the fact that the existing online fault diagnosis method cannot balance the transmission amount of high-frequency data and the accuracy of fault diagnosis at the same time under limited bandwidth resources, the present application proposes a collaborative design method for high-frequency data transmission and precision adaptive fault diagnosis. The method collaboratively designs online fault diagnosis applications and high-frequency data transmission algorithms, and realizes accurate and real-time fault diagnosis under limited bandwidth resources.

[0071] As shown in FIG. 1, the collaborative design method for high-frequency data transmission and precision adaptive fault diagnosis proposed by the embodiments of the present application includes the following steps:

[0072] Step 1: The edge gateway collects high-frequency data, pre-processes the high-frequency data, writes the pre-processed high-frequency data into a data set, and sends it to the cloud server.

[0073] Online fault diagnosis needs to transmit high-frequency signals such as vibration in real time. These analog signals in the physical world are converted into a specific range of electrical levels by sensors, and then converted into digital signals by high-speed ADC modules. These digital signals are transmitted to the edge gateway through the corresponding interface of the edge gateway, such as parallel port, via a wired high-speed cable. The gateway obtains the original data by decoding through the corresponding driver and transmission protocol. After obtaining the original data, the original data is pre-processed online. The pre-processing can include time series conversion and arctangent function normalization. Specifically, it includes:

[0074] 1. Time series conversion: a timestamp is added to each sample value online to mark its collection time. The format of the timestamp is "yyyy-MM-dd HH:mm:ss.SSS", that is, "year-month-day-hour-minute-second-millisecond".

[0075] 2. Normalization of arctangent function: due to the large variation of sample values, samples with small values cannot be effectively trained, which will affect the accuracy of time series prediction. Traditional normalization methods cannot reduce the relative difference between data values. Therefore, we use the nonlinear arctangent function normalization method to balance the relative difference between values, so that samples with relatively small values can also be effectively trained. The arctangent function normalization can normalize the values between-1 and 1. At the same time, the relative difference between values is also effectively reduced.

[0076] The arctangent function normalization is processed using the following function:

[0077] In the formula, x i is the normalized value, and x is the original value.

[0078] After the edge gateway completes the processing of the high-frequency data, the edge gateway writes the preprocessed high-frequency data into a dataset file. When the dataset file reaches a sufficient number, the edge gateway sends the dataset to the cloud server.

[0079] Step 2: The cloud server receives the dataset sent by the edge gateway and uses the dataset to train the Informer long sequence prediction model.

[0080] After receiving the dataset, the cloud server divides it into a training set, a validation set, and a test set. The training set is used to train the Informer long sequence prediction model, then the model hyperparameters are adjusted through the validation set, and finally the effect is verified on the test set to obtain the optimal prediction model, and the optimal prediction model is saved as a *.pkl file.

[0081] The role of the Informer long sequence prediction model is to provide accurate and fast long sequence time series prediction for the long sequence double prediction method (L-DPS). Historical time series data is input into the Informer long sequence prediction model, and through the prediction reasoning of the Informer long sequence prediction model, the predicted value of future time series data is output.

[0082] When training the Informer long sequence prediction model, the cloud server includes the following sub-steps:

[0083] Step 2.1: The cloud server uses the training set to train the Informer long sequence prediction model;

[0084] Step 2.2: The cloud server adjusts the hyperparameters of the Informer long sequence prediction model using the validation set;

[0085] Step 2.3: The cloud server verifies the effect on the test set and obtains the optimal prediction model;

[0086] Step 2.4: The cloud server saves the trained optimal model as a file.

[0087] Step 3: The cloud server and the edge gateway synchronously load the trained Informer long sequence prediction model.

[0088] The cloud server sends the model file of the optimal prediction model to the edge gateway and synchronously loads the optimal prediction model with the edge gateway, which can realize the prediction of high-frequency data.

[0089] Step 4: The edge gateway runs the long sequence double prediction method to predict high-frequency data, calculates the deviation of the predicted value and the actual value, and sends the actual value whose deviation exceeds the threshold to the cloud server.

[0090] Before the edge gateway and the cloud server run the long sequence double prediction method to predict high-frequency data, the DPS length parameter in L-DPS needs to be configured in advance. This parameter will affect the transmission reduction ratio and the data accuracy requirement for restoration, and needs to be manually configured. The determination of the DPS length parameter can be combined with the frequency of high-frequency data. The DPS length parameter is a specific numerical value, such as 10 or 50. Different DPS lengths will result in certain differences in data restoration accuracy. Generally speaking, the smaller the DPS length, the higher the data restoration accuracy. After the edge gateway and the cloud server successfully load the same prediction model, the edge gateway collects and pre-processes the original data to obtain initial values, and first sends the model input length n initial values to the cloud server as the initial input of the prediction model. Then, the edge gateway can use L-DPS to compress and transmit the collected and pre-processed initial values to the cloud server in real time.

[0091] In the high-frequency data real-time compression transmission algorithm L-DPS, the L-DPS compression algorithm includes three mechanisms of double prediction, confirmation and restoration, wherein the double prediction mechanism uses the Informer long sequence prediction model to perform prediction, and the edge gateway organizes the uploaded data according to the prediction result; the confirmation mechanism screens and judges the part of the real value whose prediction value deviation is greater than the threshold value, if the prediction result is inaccurate, the data points in each sampling value are all greater than the threshold value, then all the real values, that is, the sampling values of the high-frequency data, need to be transmitted, and the L-DPS does not compress any high-frequency data; if the prediction is very accurate, all the data points in the sampling value are less than the threshold value, then no sampling value needs to be uploaded, and the L-DPS has the highest compression efficiency. For specific implementation details of the L-DPS algorithm, refer to the technical solution disclosed in CN114640695A. The L-DPS algorithm is not optimized in this application.

[0092] This step includes the following sub-steps:

[0093] Step 4.1: The edge gateway collects high-frequency data and pre-processes the high-frequency data to obtain sampling values of the high-frequency data.

[0094] The pre-processing of the high-frequency data by the edge gateway is consistent with the pre-processing method in step 1, mainly including time series conversion and arctangent function normalization processing. The initial high-frequency data after pre-processing obtains the sampling values of the high-frequency data.

[0095] Step 4.2: The edge gateway uses the Informer long sequence prediction model to predict the high-frequency data to obtain the first prediction value of the high-frequency data.

[0096] Step 4.3: The edge gateway calculates the deviation between the first prediction value and the sampling value, and judges whether the deviation exceeds the threshold value.

[0097] Step 4.4: If the deviation exceeds the threshold value, the edge gateway sends the actual value of the high-frequency data.

[0098] Step 4.5: The edge gateway restores the high-frequency data using the actual value and the prediction value.

[0099] The prediction value obtained by the edge gateway predicting the high-frequency data may have a certain deviation from the sampling value. If the deviation does not exceed the threshold value, the edge gateway does not need to send the actual sampling value, otherwise the edge gateway needs to send the actual sampling value, and the cloud server restores the high-frequency data according to the actual sampling value.

[0100] Step 5: The cloud server runs a long sequence double prediction method to predict the high-frequency data, and restores the high-frequency data according to the received data.

[0101] Since the cloud server and the edge gateway load the same Informer long sequence prediction model, the prediction result of the cloud server on the high-frequency data should be basically consistent with the prediction result of the edge gateway, therefore, the cloud server can use the L-DPS to receive the data sent by the edge gateway and restore the compressed data. When the prediction result of the edge gateway is consistent with the actual sampling value, the edge gateway does not need to report the actual value, therefore, in this case, the cloud server directly uses the prediction value of the cloud server to restore the data when restoring the data.

[0102] This step includes the following sub-steps:

[0103] Step 5.1: The cloud server uses the Informer long sequence prediction model to predict the high-frequency data to obtain a second prediction value of the high-frequency data.

[0104] Step 5.2: The cloud server receives the actual value of the high-frequency data sent by the edge gateway.

[0105] Step 5.3: The cloud server restores the high-frequency data using the actual value and the second prediction value.

[0106] Step 6: The cloud server inputs the restored high-frequency data and the corresponding DPS length into the precision adaptive fault diagnosis model, and the precision adaptive fault diagnosis model outputs the fault diagnosis result online.

[0107] The precision adaptive fault diagnosis model is based on a one-dimensional convolutional neural network architecture, including block stacking, a global average pooling layer and a fully connected layer, to complete the extraction of local features and global features of the vibration signal. The block stacking is stacked by multiple blocks, and each block includes two convolutional layers and a maximum pooling layer.

[0108] This step includes the following sub-steps:

[0109] Step 6.1: The cloud server inputs the restored high-frequency data and the corresponding DPS length into the precision adaptive fault diagnosis model.

[0110] Step 6.2: The block stacking extracts the local features and high-level features of the vibration signal.

[0111] Step 6.3: The global average pooling layer extracts the global features of the vibration signal according to the high-level features.

[0112] Step 6.4: The fully connected layer obtains the fault label of the vibration signal according to the global features, and outputs the fault diagnosis result.

[0113] The embodiment provides a cooperative high-frequency data real-time compression transmission algorithm and an online fault diagnosis application supported by transmission restored data, so as to realize high transmission reduction rate and fault diagnosis accuracy at the same time. The embodiment provides a cooperative design architecture of transmission and diagnosis, so as to balance data transmission reduction and fault diagnosis accuracy, thereby realizing accurate and real-time fault diagnosis under limited bandwidth resources. According to the accuracy difference of transmission restored data, an accuracy adaptive fault diagnosis model is provided, which can use the prior information of the DPS length in the high-frequency data real-time compression transmission algorithm L-DPS as an accuracy reference, and adapt to transmission restored data with different accuracies, thereby effectively improving the accuracy of fault diagnosis. According to the problem that the sampling value of high-frequency vibration signal changes sharply, so that the time sequence prediction model in L-DPS cannot be effectively trained, an inverse tangent function is used for nonlinear normalization preprocessing of data. The relative difference between the values of the high-frequency vibration signal is balanced through the nonlinear inverse tangent function normalization, so that the sample with relatively small value can also be effectively trained, the accuracy of the time sequence prediction in L-DPS is effectively improved, and the transmission reduction rate is further improved. The application solves the practical, important and key problem of accurate and real-time fault diagnosis under limited bandwidth resources, and has practicability, advancement and effectiveness.

[0114] As shown in FIG. 2, the embodiment of the application provides a cooperative design architecture of data transmission and fault diagnosis, which cooperates a data transmission algorithm and a fault diagnosis application to realize the balance between reducing high-frequency data transmission and ensuring the accuracy of online fault diagnosis. At the hardware level, the cloud edge cooperation is adopted, the prediction model is trained on the cloud server, the cloud resources are effectively used for training the prediction model, and after the training of the prediction model is completed, the trained model is loaded synchronously with the edge gateway. The edge gateway collects high-frequency data generated by the equipment in real time, and pre-processes the collected high-frequency data on the edge side in real time, and predicts through the loaded prediction model. The DPS length connects the transmission algorithm and the fault diagnosis application. The accuracy reference in the accuracy adaptive fault diagnosis model can adapt to restored data with different accuracies, thereby effectively improving the accuracy of fault diagnosis. Through the cooperation between transmission and diagnosis, the architecture can realize high transmission reduction rate and fault diagnosis accuracy at the same time.

[0115] In the embodiment, the edge gateway is responsible for collecting high-frequency data on the device side, and pre-processing the collected high-frequency data, which includes time series conversion and arctangent function normalization processing, and sending the pre-processed data to the cloud server as a data set for training the Informer long sequence prediction model. The cloud server is responsible for training the Informer long sequence prediction model, and through the test data set sent by the edge gateway, the cloud server completes the training of the Informer long sequence prediction model, and sends the trained Informer long sequence prediction model to the edge gateway. After completing the training of the Informer long sequence prediction model, the cloud server and the edge gateway simultaneously load the trained Informer long sequence prediction model to predict high-frequency data.

[0116] When predicting high-frequency data, the cloud server and the edge gateway simultaneously run the high-frequency data real-time compression transmission algorithm L-DPS, each predicts high-frequency data, and restores data according to the prediction result.

[0117] After restoring the high-frequency data, the cloud server inputs the restored high-frequency data and the corresponding DPS length to the precision adaptive fault diagnosis model, and the precision adaptive fault diagnosis model outputs the fault diagnosis result online.

[0118] The existing online fault diagnosis method needs real-time and reliable transmission of high-frequency signals such as vibration, which is difficult to meet the requirements of high frequency and real-time in limited bandwidth resources. However, reducing data transmission to relieve bandwidth pressure will cause accuracy deviation of transmitted data and reduce the accuracy of fault diagnosis. The collaborative design architecture of data transmission and fault diagnosis proposed in the application realizes high transmission reduction rate and fault diagnosis accuracy rate at the same time through the collaborative design of the high-frequency data real-time compression transmission algorithm L-DPS and the online fault diagnosis application supported thereby. Through the collaborative design of the high-frequency data real-time compression transmission algorithm and the online fault diagnosis application supported thereby, the balance between reducing data transmission and ensuring fault diagnosis accuracy can be achieved. While real-time compression transmission of high-frequency data meets the limited bandwidth restriction, the accuracy of data transmission and the accuracy of fault diagnosis are collaboratively ensured.

[0119] As shown in FIG. 3, the precision adaptive fault diagnosis model proposed in the embodiments of the present application is used to solve the problem of precision difference of restored data caused by L-DPS, and is based on a one-dimensional convolutional neural network (CNN) architecture. Unlike traditional CNN, the present application considers the precision difference of model input during data transmission. The DPS length in L-DPS is used to connect data transmission and diagnosis, providing a strong reference for the precision of model input data. The model can adapt to restored data of different precisions, thereby reducing data transmission while effectively improving the accuracy of fault diagnosis. The precision adaptive fault diagnosis model includes a convolutional layer, a max-pooling layer, a global average pooling layer, and a fully connected layer. The main part of the model is block stacking, and a block is composed of two convolutional layers and a max-pooling layer. Multiple blocks are stacked together. The model mines the prior information about the precision of the vibration signal contained in the DPS length, and extracts the local and global features of the vibration signal while identifying the precision state of the input vibration value. The DPS length is a parameter set by the user in the L-DPS algorithm, and this parameter will affect the precision of the restored vibration signal by L-DPS. The precision adaptive fault diagnosis model can mine and learn the relationship between the DPS length and the precision of the restored data by L-DPS through training, and use the prior information about the precision of the restored data provided by the DPS length to perform more accurate fault diagnosis during inference. The model effectively reduces the dimension and noise of the data, and enhances the representativeness of the signal. At the same time, it can also learn complex fault patterns through multi-layer nonlinear transformation, thereby improving the accuracy and robustness of diagnosis. As shown in FIGS. 4 and 5, the precision of the restored data is different for different DPS lengths. It can be seen that the two curves are different. In FIG. 4, when the DPS length is 10, the restored data is more accurate; while in FIG. 5, when the DPS length is 50, the curve is relatively flat and inaccurate, and many features are smoothed out. Therefore, in the present application, considering that the precision of the restored data is obviously different for different DPS lengths, the present application needs to adapt to the precision, which is the motivation for considering precision adaptation in the present application.

[0120] In the embodiments, the precision adaptive fault diagnosis model is based on a one-dimensional convolutional neural network architecture, including block stacking, global average pooling layer, and fully connected layer, to complete the extraction of local and global features of the vibration signal.

[0121] The block stacking is composed of multiple blocks stacked together, and a block includes two convolutional layers and a max-pooling layer. The convolutional layer is used to extract the local features of the vibration signal, and the convolutional layer is defined as:

[0122] In the formula, is a tensor, W iis a convolution filter, b i is a first bias term, ReLU(·) is a first activation function, i, i+1 is a layer number, t is a time step, and Conv1d is a convolution layer.

[0123] The max-pooling layer is used to extract high-level features of the vibration signal, and the max-pooling layer is defined as follows:

[0124] In the formula, MaxPool is the max-pooling layer, max is the maximum value, i, i+1 is a layer number, t is a time step, is a tensor.

[0125] The global average pooling layer receives high-level features output by the block stack, extracts global features of the vibration signal, and the global average pooling layer is defined as follows:

[0126] In the formula, GlobalAvgPool is the global average pooling layer, j is a serial number, n is the length of the model input data, i, i+1 is a layer number, t is a time step, is a tensor.

[0127] The fully connected layer receives global features output by the global average pooling layer, and the fully connected layer outputs a fault label, and the fully connected layer is defined as follows:

[0128] In the formula, FC is the fully connected layer, W i is a weight matrix, and b i is a second bias term, Softmax(·) is a second activation function, i, i+1 is a layer number, t is a time step, is a tensor.

[0129] In another embodiment of the present application, considering that the core of the precision adaptive fault diagnosis model is to input the corresponding DPS length of the restored data at the same time, so that the model can adapt to different precision restored data by mining the precision reference therein. Therefore, although the CNN architecture is the most effective fault diagnosis architecture at present, the precision adaptive fault diagnosis model is not limited to the CNN architecture, but can also be based on architectures such as transformer, recurrent neural network (RNN), multilayer perceptron (MLP), etc.

[0130] The precision adaptive fault diagnosis model provided by the embodiments of the present application can be matched with the transmission algorithm to realize high fault diagnosis accuracy for transmission restored data of different precision.

[0131] Compared with the prior art, the collaborative design method of high-frequency data transmission and precision adaptive fault diagnosis provided by the application has the following beneficial effects:

[0132] 1. The existing online fault diagnosis method needs real-time and reliable transmission of high-frequency signals such as vibration, and limited bandwidth resources are difficult to meet the requirements of high frequency and real-time. Reducing the amount of data transmission to relieve bandwidth pressure will cause precision deviation of the transmitted data and reduce the accuracy of fault diagnosis. In order to realize accurate and real-time fault diagnosis under limited bandwidth resources, the application proposes a collaborative design architecture of data transmission and fault diagnosis, which realizes high transmission reduction rate and fault diagnosis accuracy rate through the collaborative design of the high-frequency data real-time compression transmission algorithm L-DPS and the online fault diagnosis application supported thereby. Through the collaborative design of the high-frequency data real-time compression transmission algorithm and the online fault diagnosis application supported thereby, the balance between reducing the amount of data transmission and ensuring the accuracy of fault diagnosis can be achieved. While real-time compression transmission of high-frequency data meets the limitation of limited bandwidth, the accuracy of data transmission and the accuracy of fault diagnosis are collaboratively ensured, achieving the balance between reducing the amount of data transmission and ensuring the accuracy of fault diagnosis, so as to realize accurate and real-time fault diagnosis under limited bandwidth resources.

[0133] 2. The existing fault diagnosis model does not consider the precision difference of model input data, and cannot realize accurate fault diagnosis due to the lack of adaptability to the precision difference and loss of restored data after data compression transmission. The application proposes a precision adaptive fault diagnosis model for the precision difference of transmission restored data. By using the prior information of the length of DPS as the precision reference of transmission restored data, the proposed model can adapt to transmission restored data of different precision, thereby improving the accuracy of fault diagnosis.

[0134] 3. The existing scheme generally uses linear normalization method for data preprocessing, which cannot reduce the relative difference between data values. For high-frequency vibration signals, due to the large change of sampling values, samples with small values cannot be effectively trained, and accurate time series prediction is difficult to achieve. In view of the problem that the time series prediction model cannot be effectively trained due to the sharp change of high-frequency vibration signal sampling values, the application proposes to use the inverse tangent function for nonlinear normalization preprocessing of data. By balancing the relative difference between high-frequency vibration signal values through nonlinear inverse tangent function normalization, samples with relatively small values can also be effectively trained, effectively improving the accuracy of time series prediction in the high-frequency data real-time compression transmission algorithm.

[0135] The preferred embodiments of the application have been described above in detail. It should be understood that modifications and variations to the preferred embodiments could be made by those skilled in the art in light of the teachings above. It is therefore contemplated that the application can encompass other embodiments that fall within the scope of the claims.

Claims

1. A collaborative design method of high-frequency data transmission and precision adaptive fault diagnosis, the method comprising the following steps: S101: An edge gateway collects high-frequency data, pre-processes the high-frequency data, writes the pre-processed high-frequency data into a data set, and sends the data set to a cloud server; S102: The cloud server receives the data set sent by the edge gateway, and trains an Informer long sequence prediction model using the data set; S103: The cloud server and the edge gateway synchronously load the trained Informer long sequence prediction model; S104: The edge gateway runs a long sequence double prediction method to predict the high-frequency data, calculates the deviation of the predicted value and the actual value, and sends the actual value whose deviation exceeds a threshold to the cloud server; S105: The cloud server runs the long sequence double prediction method to predict the high-frequency data, and restores the high-frequency data according to the received data; S106: The cloud server inputs the restored high-frequency data and the corresponding DPS length of the high-frequency data into a precision adaptive fault diagnosis model, and the precision adaptive fault diagnosis model outputs a fault diagnosis result online.

2. The method of claim 1, wherein, In the S101, the pre-processing of the high-frequency data by the edge gateway comprises time series conversion and normalization.

3. The method of claim 2, wherein, The time series conversion is to add a timestamp online for each sampling value of the high-frequency data to mark the collection time of the high-frequency data.

4. The method of claim 3, wherein, The timestamp includes year, month, day, hour, minute, second and millisecond information, and the format of the timestamp is "yyyy-MM-dd HH:mm:ss.SSS".

5. The method of claim 4, wherein, The normalization is to normalize the numerical value of the high-frequency data to a value between -1 and 1.

6. The method of claim 5, wherein, The normalization is done using an arctangent function normalization method, which processes using the following function: In the formula, x i is a normalized value, and x is an original value.

7. The method of claim 6, wherein, In the S102, after receiving the data set sent by the edge gateway, the cloud server divides the data set into a training set, a validation set and a test set.

8. The method of claim 7, wherein, In the S102, the cloud server trains the Informer long sequence prediction model using the data set, comprising the following sub-steps: S1021: The cloud server trains the Informer long sequence prediction model using the training set; S1022: The cloud server adjusts the hyperparameters of the Informer long sequence prediction model using the validation set; S1023: The cloud server verifies the effect on the test set to obtain an optimal prediction model; S1024: The cloud server saves the obtained optimal model as a file.

9. The method of claim 8, wherein, In the S103, the cloud server sends the optimal prediction model of the Informer long sequence prediction model to the edge gateway, and synchronously loads the optimal prediction model with the edge gateway.

10. The method of claim 9, wherein, The S104 comprises the following sub-steps: S1041: The edge gateway collects the high-frequency data, and pre-processes the high-frequency data to obtain the actual value of the high-frequency data; S1042: The edge gateway uses the Informer long sequence prediction model to predict the high-frequency data, obtaining a first predicted value of the high-frequency data; S1043: The edge gateway calculates the deviation between the first predicted value and the actual value, and determines whether the deviation exceeds the threshold value; S1044: If the deviation exceeds the threshold value, the edge gateway sends the actual value of the high-frequency data; S1045: The edge gateway restores the high-frequency data using the actual value and the first predicted value.

11. The method of claim 10, wherein, The S105 includes the following sub-steps: S1051: The cloud server uses the Informer long sequence prediction model to predict the high-frequency data, obtaining a second predicted value of the high-frequency data; S1052: The cloud server receives the actual value of the high-frequency data sent by the edge gateway; S1053: The cloud server restores the high-frequency data using the actual value and the second predicted value.

12. The method of claim 11, wherein, In the S106, the precision adaptive fault diagnosis model is based on one of a one-dimensional convolutional neural network, a transformer, a recurrent neural network, and a multi-layer perceptron architecture.

13. The method of claim 12, wherein, The precision adaptive fault diagnosis model is based on the one-dimensional convolutional neural network architecture, which includes block stacking, global average pooling layer, and fully connected layer, and is configured to be able to complete the extraction of local features and global features of the vibration signal.

14. The method of claim 13, wherein, The block stacking is stacked by a plurality of blocks, each of which includes two convolutional layers and a maximum pooling layer.

15. The method of claim 14, wherein, The two convolutional layers are used to extract the local features of the vibration signal, and the convolutional layer is defined as: In the formulae, is a tensor, W i is a convolution filter, b i is a first bias term, ReLU(·) is a first activation function, i, i+1 are layer indices, t is a time step, and Conv1d is a convolution layer.

16. The method of claim 15, wherein, The max pooling layer is used to extract high-level features of the vibration signal, and is defined as follows: In the formula, MaxPool is a max-pooling layer, max is a maximum value, i and i+1 are layer numbers, and t is a time step. is a tensor.

17. The method of claim 16, wherein, The global average pooling layer is used to receive the high-level features of the block stack output, extract the global features of the vibration signal, and is defined as follows: wherein GlobalAvgPool is a global average pooling layer, j is a serial number, n is a length of model input data, i, i+1 are layer serial numbers, and t is a time step, is a tensor.

18. The method of claim 17, wherein, The fully connected layer is configured to receive the global features output by the global average pooling layer, and is configured to output a failure label. The fully connected layer is defined as follows: where FC is a fully connected layer, W i is a weight matrix, b i is a second bias term, Softmax(·) is a second activation function, i, i+1 are layer numbers, t is a time step, is a tensor.

19. The method of claim 18, wherein, The S106 includes the following sub-steps: S1061: The cloud server inputs the restored high-frequency data and the DPS length corresponding to the high-frequency data to the precision adaptive fault diagnosis model; S1062: The block stacking extracts the local features and high-level features of the vibration signal; S1063: The global average pooling layer extracts the global features of the vibration signal according to the high-level features; S1064: The fully connected layer obtains the fault label of the vibration signal according to the global features, and outputs a fault diagnosis result.

20. The method of claim 19, wherein, The DPS length corresponding to the high-frequency data is configured according to the frequency of the high-frequency data.

Images