WO2026007341A1 - 一种高频数据传输与精度自适应故障诊断的协同设计方法 - Google Patents

一种高频数据传输与精度自适应故障诊断的协同设计方法

Info

Publication number
WO2026007341A1
WO2026007341A1 PCT/CN2024/139533 CN2024139533W WO2026007341A1 WO 2026007341 A1 WO2026007341 A1 WO 2026007341A1 CN 2024139533 W CN2024139533 W CN 2024139533W WO 2026007341 A1 WO2026007341 A1 WO 2026007341A1
Authority
WO
WIPO (PCT)
Prior art keywords
frequency data
cloud server
fault diagnosis
data
edge gateway
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
PCT/CN2024/139533
Other languages
English (en)
French (fr)
Inventor
杨博
吴雨
朱大风
陈彩莲
关新平
王勇
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shanghai Electric Wind Power Group Co Ltd
Shanghai Jiao Tong University
Original Assignee
Shanghai Electric Wind Power Group Co Ltd
Shanghai Jiao Tong University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Shanghai Electric Wind Power Group Co Ltd, Shanghai Jiao Tong University filed Critical Shanghai Electric Wind Power Group Co Ltd
Priority to US19/220,787 priority Critical patent/US20260009673A1/en
Publication of WO2026007341A1 publication Critical patent/WO2026007341A1/zh
Pending legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B13/00—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion
    • G05B13/02—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric
    • G05B13/04—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric involving the use of models or simulators
    • G05B13/042—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric involving the use of models or simulators in which a parameter or coefficient is automatically adjusted to optimise the performance
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01H—MEASUREMENT OF MECHANICAL VIBRATIONS OR ULTRASONIC, SONIC OR INFRASONIC WAVES
    • G01H17/00—Measuring mechanical vibrations or ultrasonic, sonic or infrasonic waves, not provided for in the other groups of this subclass
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Definitions

  • This application relates to the field of online fault diagnosis, and in particular to a collaborative design method for high-frequency data transmission and accuracy-adaptive fault diagnosis.
  • Dual Prediction Schemes can reduce real-time data transmission while ensuring a deterministic range of data accuracy.
  • DPS based on deep learning (DL) prediction models can handle more complex data, achieving higher data prediction accuracy and transmission reduction rates, and they are becoming mainstream.
  • Yu et al. proposed a DL model based on transfer learning for diagnosing mechanical faults from unlabeled samples without a known label space.
  • Chai et al. proposed a multi-source simplified transfer network to address fault diagnosis problems with inconsistent domains and categories.
  • Jiao et al. proposed a passive adaptive diagnosis for rotating machinery, which can continuously optimize and update the fault diagnosis model using unlabeled data.
  • Shao et al. proposed a dual-threshold attention-guided generative adversarial network, providing a portable data generation solution for fault diagnosis with limited samples.
  • these methods still have certain limitations, as they do not consider the impact of data transmission on online fault diagnosis.
  • Chinese invention patent application CN114640695A discloses a method for efficient transmission of high-frequency time-series data in smart factories based on long sequence dual prediction and Informer.
  • This method first establishes a cloud-edge collaborative long sequence dual prediction architecture, then deploys a trained long sequence prediction model on the edge gateway and cloud server within this architecture, and finally employs the Long Sequence Dual Prediction Scheme (L-DPS) to reduce the amount of high-frequency data transmitted online while maintaining data accuracy.
  • L-DPS Long Sequence Dual Prediction Scheme
  • It improves the structure of traditional dual prediction methods by reducing the number of inference iterations of the prediction model through long sequence prediction, thereby significantly increasing the applicability of traditional methods and making it possible to reduce the amount of high-frequency data transmitted in smart manufacturing processes.
  • Simultaneously it introduces and combines the latest deep learning model Informer to solve the problems of gradient vanishing and increased model inference time caused by long sequence prediction, further improving the reduction rate and applicability of the proposed long sequence dual prediction method.
  • the technical problem to be solved by this application is how to achieve accurate and real-time fault diagnosis under limited bandwidth.
  • this application provides a collaborative design method for high-frequency data transmission and accuracy-adaptive fault diagnosis, the method comprising the following steps:
  • the edge gateway collects high-frequency data, preprocesses the high-frequency data, writes the preprocessed high-frequency data into a dataset, and sends it to the cloud server.
  • the cloud server receives the dataset sent by the edge gateway and uses the dataset to train the Informer long sequence prediction model
  • S103 The cloud server and the edge gateway synchronously load the trained Informer long sequence prediction model
  • the edge gateway runs a long-sequence dual prediction method to predict the high-frequency data, calculate the deviation between the predicted value and the actual value, and send the actual value whose deviation exceeds a threshold to the cloud server.
  • the cloud server runs the long sequence dual prediction method to predict the high-frequency data and restore the high-frequency data based on the received data;
  • the cloud server inputs the restored high-frequency data and the DPS length corresponding to the high-frequency data into the accuracy adaptive fault diagnosis model, and the accuracy adaptive fault diagnosis model outputs the fault diagnosis result online.
  • step S101 the edge gateway preprocesses the high-frequency data by performing time series transformation and normalization.
  • time series conversion involves adding a timestamp online to each sample value of the high-frequency data to mark the acquisition time of the high-frequency data.
  • the timestamp includes year, month, day, hour, minute, second, and millisecond information
  • the timestamp format is "yyyy-MM-dd HH:mm:ss.SSS".
  • the normalization is to regularize the values of the high-frequency data to values between -1 and 1.
  • xi is the normalized value
  • x is the original value
  • step S102 after receiving the dataset sent by the edge gateway, the cloud server divides the dataset into a training set, a validation set, and a test set.
  • step S102 the cloud server trains the Informer long sequence prediction model using the dataset, which includes the following sub-steps:
  • S1021 The cloud server uses the training set to train the Informer long sequence prediction model
  • the cloud server uses the validation set to adjust the hyperparameters of the Informer long sequence prediction model
  • S1023 The cloud server verifies the effect on the test set and obtains the optimal prediction model
  • S1024 The cloud server saves the trained optimal model as a file.
  • step S103 the cloud server sends the optimal prediction model of the Informer long sequence prediction model to the edge gateway, and loads the optimal prediction model synchronously with the edge gateway.
  • step S104 includes the following sub-steps:
  • the edge gateway collects the high-frequency data and performs the preprocessing on the high-frequency data to obtain the actual value of the high-frequency data;
  • the edge gateway uses the Informer long sequence prediction model to predict the high-frequency data and obtains the first predicted value of the high-frequency data;
  • the edge gateway calculates the deviation between the first predicted value and the actual value, and determines whether the deviation exceeds the threshold;
  • the edge gateway uses the actual value and the first predicted value to reconstruct the high-frequency data.
  • step S105 includes the following sub-steps:
  • the cloud server uses the Informer long sequence prediction model to predict the high-frequency data and obtains a second predicted value for the high-frequency data;
  • the cloud server receives the actual value of the high-frequency data sent by the edge gateway;
  • S1053 The cloud server uses the actual value and the second predicted value to reconstruct the high-frequency data.
  • the accuracy-adaptive fault diagnosis model is based on any one of the following architectures: one-dimensional convolutional neural network, transformer, recurrent neural network, and multilayer perceptron.
  • the accuracy-adaptive fault diagnosis model is based on the one-dimensional convolutional neural network architecture, which includes block stacking, global average pooling layers, and fully connected layers, and is configured to extract local and global features of vibration signals.
  • the block stack is composed of multiple blocks stacked on top of each other, and each of the multiple blocks includes two convolutional layers and a max pooling layer.
  • the two convolutional layers are used to extract the local features of the vibration signal, and the convolutional layers are defined as follows:
  • W ⁇ sub> i ⁇ /sub> is a tensor
  • W ⁇ sub>i ⁇ /sub> is a convolutional filter
  • b ⁇ sub>i ⁇ /sub> is the first bias term
  • ReLU( ⁇ ) is the first activation function
  • i and i+1 are the layer numbers
  • t is the time step
  • Conv1d is the convolutional layer.
  • the max-pooling layer is used to extract high-level features of the vibration signal, and the max-pooling layer is defined as follows:
  • MaxPool is the max pooling layer
  • max means taking the maximum value
  • i and i+1 are the layer numbers
  • t is the time step. It is a tensor.
  • the global average pooling layer is used to receive the high-level features output by the block stack and extract the global features of the vibration signal.
  • the global average pooling layer is defined as follows:
  • GlobalAvgPool is the global average pooling layer
  • j is the index
  • n is the length of the model input data
  • i and i+1 are the layer indices
  • t is the time step. It is a tensor.
  • the fully connected layer is used to receive the global features output by the global average pooling layer, and the fully connected layer is configured to output fault labels.
  • the fully connected layer is defined as follows:
  • FC is a fully connected layer
  • Wi is the weight matrix
  • bi is the second bias term
  • Softmax( ⁇ ) is the second activation function
  • i and i+1 are the layer numbers
  • t is the time step. It is a tensor.
  • step S106 includes the following sub-steps:
  • the cloud server inputs the restored high-frequency data and the DPS length corresponding to the high-frequency data into the accuracy adaptive fault diagnosis model;
  • the block stack extracts the local features and the high-level features of the vibration signal
  • the global average pooling layer extracts the global features of the vibration signal based on the high-level features
  • the fully connected layer obtains the fault label of the vibration signal based on the global features and outputs the fault diagnosis result.
  • the DPS length corresponding to the high-frequency data is configured according to the frequency of the high-frequency data.
  • This application achieves a balance between reducing data transmission volume and ensuring fault diagnosis accuracy by coordinating the design of a high-frequency data real-time compression transmission algorithm and its supported online fault diagnosis application. While compressing and transmitting high-frequency data in real time to meet the limited bandwidth constraints, it also coordinates the accuracy of data transmission and the accuracy of fault diagnosis, thereby achieving accurate and real-time fault diagnosis under limited bandwidth resources.
  • This application utilizes prior information about the DPS length as a precision reference for the transmission restoration data.
  • the proposed model can adapt to transmission restoration data of different precisions, thereby improving the accuracy of fault diagnosis.
  • This application uses nonlinear arctangent function normalization to balance the relative differences between high-frequency vibration signal values, so that even samples with relatively small values can be effectively trained, thereby effectively improving the accuracy of time series prediction in high-frequency data real-time compression and transmission algorithms.
  • Figure 1 is a flowchart illustrating the collaborative design method according to an embodiment of this application.
  • Figure 2 is a schematic diagram of the collaborative design architecture for data transmission and fault diagnosis in an embodiment of this application
  • Figure 3 is a schematic diagram of the accuracy adaptive fault diagnosis model according to an embodiment of this application.
  • Figure 4 is a schematic diagram of the restored data accuracy when the DPS length is 10 in an embodiment of this application;
  • Chinese invention patent application CN114640695A discloses an effective high-frequency time-series data transmission method for smart factories based on long-sequence dual prediction (L-DPS) and Informer. This method first establishes a cloud-edge collaborative long-sequence dual prediction architecture, then deploys a trained long-sequence prediction model on the edge gateway and cloud server within this architecture, and finally uses the long-sequence dual prediction method (L-DPS) to reduce the transmission volume of high-frequency data online while maintaining data accuracy.
  • L-DPS long-sequence dual prediction
  • This method improves the structure of traditional dual prediction methods by reducing the number of inference iterations of the prediction model through long-sequence prediction, thereby significantly increasing the applicability of traditional methods and making it possible to reduce the transmission volume of high-frequency data required in smart manufacturing processes.
  • the introduction and integration of the latest deep learning model Informer solves the problems of gradient vanishing and increased model inference time caused by long-sequence prediction, further improving the transmission volume reduction ratio and applicability of the proposed long-sequence dual prediction method.
  • this application proposes a collaborative design method for high-frequency data transmission and accuracy-adaptive fault diagnosis.
  • This method collaboratively designs online fault diagnosis applications and high-frequency data transmission algorithms, achieving accurate and real-time fault diagnosis under limited bandwidth resources.
  • the collaborative design method for high-frequency data transmission and accuracy-adaptive fault diagnosis proposed in this application includes the following steps:
  • Step 1 The edge gateway collects high-frequency data, preprocesses the high-frequency data, writes the preprocessed high-frequency data into the dataset, and sends it to the cloud server.
  • Online fault diagnosis requires real-time transmission of high-frequency signals such as vibration. These analog signals from the physical world are converted into voltage levels within a specific range by sensors, and then converted into digital signals by a high-speed ADC module. These digital signals are transmitted to the edge gateway via a wired high-speed cable through the corresponding interface, such as a parallel port. The gateway decodes the data using the appropriate driver and transmission protocol to obtain the raw data. After obtaining the raw data, online preprocessing is performed. Preprocessing may include time series transformation and arctangent function normalization. Specifically, it includes:
  • Time Series Conversion Add a timestamp online to each sampled value to mark its collection time.
  • the timestamp format is "yyyy-MM-dd HH:mm:ss.SSS”, which is "year-month-day-hour-minute-second-millisecond”.
  • Arctangent Function Normalization Due to the large variation in sampled 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 differences between data values. Therefore, we use a non-linear arctangent function normalization method to balance the relative differences between values, so that even samples with relatively small values can be effectively trained. Arctangent function normalization can regularize the values to between -1 and 1. At the same time, the relative differences between values are also effectively reduced.
  • Arctangent function normalization is performed using the following function:
  • xi is the normalized value
  • x is the original value
  • the edge gateway After completing the processing of high-frequency data, the edge gateway writes the preprocessed high-frequency data into a dataset file. Once the dataset file reaches a sufficient number of records, the edge gateway sends the dataset to the cloud server.
  • 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.
  • the cloud server After receiving the dataset, the cloud server divides it into training, validation, and test sets.
  • the Informer long sequence prediction model is trained using the training set, the model's hyperparameters are then adjusted using the validation set, and finally, the performance is validated on the test set to obtain the optimal prediction model, which is then saved as a *.pkl file.
  • the Informer long series forecasting model provides accurate and fast long series time series forecasts for the Long Series Dual Forecasting Method (L-DPS).
  • Historical time series data is input into the Informer long series forecasting model, and through its predictive inference, it outputs predicted values for future time series data.
  • the cloud server trains the Informer long sequence prediction model, including the following sub-steps:
  • Step 2.1 The cloud server trains the Informer long sequence prediction model using the training set
  • Step 2.2 The cloud server uses the validation set to adjust the hyperparameters of the Informer long sequence prediction model
  • Step 2.3 The cloud server verifies the effect on the test set to obtain the optimal prediction model
  • Step 2.4 The cloud server saves the optimal model obtained from training as a file.
  • Step 3 The cloud server and edge gateway synchronously load the trained Informer long sequence prediction model.
  • the cloud server sends the model file of the optimal prediction model to the edge gateway and loads the optimal prediction model synchronously with the edge gateway, thus enabling the prediction of high-frequency data.
  • Step 4 The edge gateway runs a long-sequence dual prediction method to predict high-frequency data, calculates the deviation between the predicted value and the actual value, and sends the actual value with a deviation exceeding the threshold to the cloud server.
  • the DPS length parameter in L-DPS needs to be configured in advance. This parameter affects the reduction ratio of transmission volume and the accuracy requirements of the restored data.
  • the DPS length parameter needs to be manually configured.
  • the DPS length parameter can be determined based on the frequency of the high-frequency data.
  • the DPS length parameter is a specific value, such as 10 or 50. Different DPS lengths will result in different accuracy of the restored data. Generally speaking, the smaller the DPS length, the higher the data restoration accuracy.
  • the edge gateway collects and preprocesses the raw data in real time to obtain initial values, and first sends n initial values of model input length to the cloud server as the initial input of the prediction model. Then, the edge gateway can use L-DPS to compress the collected and preprocessed initial values in real time and transmit them to the cloud server.
  • the L-DPS compression algorithm includes three mechanisms: dual prediction, confirmation, and restoration.
  • the dual prediction mechanism uses the Informer long sequence prediction model for prediction, and the edge gateway organizes the uploaded data based on the prediction results.
  • the confirmation mechanism filters and judges the actual values whose prediction deviations are greater than a threshold. If the prediction result is inaccurate, and the prediction deviation of each data point in the sample value is greater than the threshold, then all actual values, i.e., the sample values of high-frequency data, need to be transmitted, and L-DPS does not compress any high-frequency data.
  • L-DPS achieves the highest compression efficiency.
  • L-DPS algorithm please refer to the technical solution disclosed in CN114640695A. This application does not optimize the L-DPS algorithm.
  • This step includes the following sub-steps:
  • Step 4.1 The edge gateway collects high-frequency data and preprocesses the high-frequency data to obtain sampled values of the high-frequency data;
  • the edge gateway preprocesses high-frequency data in the same way as the preprocessing method in step 1, mainly including time series transformation and arctangent function normalization. After preprocessing, the initial high-frequency data becomes the sampled value of the high-frequency data.
  • Step 4.2 The edge gateway uses the Informer long sequence prediction model to predict high-frequency data and obtain the first predicted value of the high-frequency data;
  • Step 4.3 The edge gateway calculates the deviation between the first predicted value and the sampled value, and determines whether the deviation exceeds the threshold;
  • Step 4.4 If the deviation exceeds the threshold, the edge gateway sends the actual value of the high-frequency data
  • Step 4.5 The edge gateway uses the actual and predicted values to reconstruct high-frequency data.
  • the predicted value obtained by the edge gateway from predicting high-frequency data may have a certain deviation from the sampled value. If the deviation does not exceed the threshold, the edge gateway does not need to send the actual sampled value. Otherwise, the edge gateway needs to send the actual sampled value, and the cloud server will reconstruct the high-frequency data based on this actual sampled value.
  • Step 5 The cloud server runs a long-sequence dual prediction method to predict high-frequency data and restore the high-frequency data based on the received data.
  • the cloud server and edge gateway both load the same Informer long-sequence prediction model, the cloud server's prediction results for high-frequency data should be largely consistent with those of the edge gateway. Therefore, the cloud server can use L-DPS to receive data sent by the edge gateway and restore the compressed data. When the edge gateway's prediction result matches the actual sampled value, the edge gateway does not need to report the actual value. Therefore, in this case, the cloud server directly uses its own prediction value to restore the data.
  • This step includes the following sub-steps:
  • Step 5.1 The cloud server uses the Informer long sequence prediction model to predict high-frequency data and obtains the second predicted value of the high-frequency data;
  • Step 5.2 The cloud server receives the actual value of the high-frequency data sent by the edge gateway;
  • Step 5.3 The cloud server uses the actual value and the second predicted value to reconstruct the high-frequency data.
  • Step 6 The cloud server inputs the restored high-frequency data and the corresponding DPS length into the accuracy adaptive fault diagnosis model, and the accuracy adaptive fault diagnosis model outputs the fault diagnosis results online.
  • the accuracy-adaptive fault diagnosis model is based on a one-dimensional convolutional neural network architecture, including block stacking, global average pooling layers, and fully connected layers, to extract local and global features of vibration signals.
  • Block stacking consists of multiple blocks stacked on top of each other, with each block including two convolutional layers and one max pooling layer.
  • This step includes the following sub-steps:
  • Step 6.1 The cloud server inputs the restored high-frequency data and the corresponding DPS length into the accuracy adaptive fault diagnosis model
  • Step 6.2 Extract local and high-level features of vibration signals by stacking blocks
  • Step 6.3 The global average pooling layer extracts global features of the vibration signal based on high-level features
  • Step 6.4 The fully connected layer obtains the fault labels of the vibration signal based on the global features and outputs the fault diagnosis results.
  • This embodiment provides a collaborative high-frequency data real-time compression and transmission algorithm and its transmitted and restored data-supported online fault diagnosis application, achieving both high transmission reduction rate and fault diagnosis accuracy.
  • This embodiment proposes a collaborative design architecture for transmission and diagnosis to balance data transmission reduction with fault diagnosis accuracy assurance, thereby achieving accurate and real-time fault diagnosis under limited bandwidth resources.
  • an accuracy-adaptive fault diagnosis model is proposed. This model utilizes prior information about the DPS length in the L-DPS algorithm as an accuracy reference to adaptively adapt to transmitted and restored data with different accuracies, effectively improving fault diagnosis accuracy.
  • Nonlinear arctangent function normalization balances the relative differences between high-frequency vibration signal values, allowing even relatively small samples to be effectively trained, effectively improving the accuracy of time-series prediction in L-DPS and further increasing the transmission reduction rate.
  • this application provides a collaborative design architecture for data transmission and fault diagnosis.
  • This architecture coordinates data transmission algorithms and fault diagnosis applications to achieve a balance between reducing high-frequency data transmission volume and ensuring online fault diagnosis accuracy.
  • a cloud-edge collaborative approach is adopted.
  • the prediction model is trained on a cloud server, effectively utilizing cloud resources for training.
  • the trained model is synchronously loaded with the edge gateway.
  • the edge gateway collects high-frequency data generated by the device in real time, performs preprocessing on the collected high-frequency data at the edge, and makes predictions using the loaded prediction model.
  • the DPS length connects the transmission algorithm and the fault diagnosis application.
  • the accuracy-adaptive fault diagnosis model can adapt to reconstruction data of different accuracies, effectively improving the accuracy of fault diagnosis.
  • this architecture can simultaneously achieve a high transmission reduction rate and a high fault diagnosis accuracy.
  • the edge gateway is responsible for collecting high-frequency data from the device side and preprocessing the collected high-frequency data. This preprocessing includes time series transformation and arctangent function normalization.
  • the preprocessed data is then sent to the cloud server as the dataset for training the Informer long-sequence prediction model.
  • the cloud server is responsible for training the Informer long-sequence prediction model.
  • the cloud server completes the training of the Informer long-sequence prediction model and sends the trained Informer long-sequence prediction model back to the edge gateway.
  • both the cloud server and the edge gateway simultaneously load the trained Informer long-sequence prediction model to predict high-frequency data.
  • the cloud server and the edge gateway When predicting high-frequency data, the cloud server and the edge gateway simultaneously run the high-frequency data real-time compression and transmission algorithm L-DPS, each predicting the high-frequency data and restoring the data based on the prediction results.
  • the cloud server After the cloud server reconstructs the high-frequency data, it inputs the reconstructed high-frequency data and the corresponding DPS length into the accuracy-adaptive fault diagnosis model, which then outputs the fault diagnosis results online.
  • this application proposes an accuracy-adaptive fault diagnosis model to address the accuracy discrepancy problem in transmitted and reconstructed data caused by L-DPS.
  • This model is based on a one-dimensional convolutional neural network (CNN) architecture. Unlike traditional CNNs, this application considers the accuracy differences of the model input during data transmission. By utilizing the DPS length in L-DPS to connect data transmission and diagnosis, it provides a strong reference for the accuracy of the model input data. This model can adapt to reconstructed data of different accuracies, thereby reducing data transmission while effectively improving the accuracy of fault diagnosis.
  • This accuracy-adaptive fault diagnosis model includes convolutional layers, max-pooling layers, global average pooling layers, and fully connected layers.
  • the main part of the model is block stacking; each block consists of two convolutional layers and one max-pooling layer, with multiple blocks stacked on top of each other.
  • the model mines the prior information about the accuracy of the vibration signal contained in the DPS length, extracting local and global features of the vibration signal while identifying the accuracy state of the input vibration value.
  • the DPS length is a user-defined parameter in the L-DPS algorithm, which affects the accuracy of the vibration signal reconstructed by the L-DPS algorithm.
  • the accuracy-adaptive fault diagnosis model can learn and explore the relationship between DPS length and L-DPS reconstructed data accuracy through training, and utilize the prior information on reconstructed data accuracy provided by the DPS length during inference for more accurate fault diagnosis.
  • the accuracy-adaptive fault diagnosis model is based on a one-dimensional convolutional neural network architecture, including block stacking, global average pooling layers, and fully connected layers, to extract local and global features of vibration signals.
  • the block stacking consists of multiple stacked blocks, with each block comprising two convolutional layers and one max-pooling layer.
  • the convolutional layers are used to extract local features of the vibration signal; a convolutional layer is defined as follows:
  • W ⁇ sub> i ⁇ /sub> is a tensor
  • W ⁇ sub>i ⁇ /sub> is a convolutional filter
  • b ⁇ sub>i ⁇ /sub> is the first bias term
  • ReLU( ⁇ ) is the first activation function
  • i and i+1 are the layer numbers
  • t is the time step
  • Conv1d is the convolutional layer.
  • Max pooling layers are used to extract high-level features of vibration signals.
  • the definition of a max pooling layer is as follows:
  • MaxPool is the max pooling layer
  • max means taking the maximum value
  • i and i+1 are the layer numbers
  • t is the time step. It is a tensor.
  • the global average pooling layer receives high-level features from the stacked receive blocks and extracts global features from the vibration signal.
  • the global average pooling layer is defined as follows:
  • GlobalAvgPool is the global average pooling layer
  • j is the index
  • n is the length of the model input data
  • i and i+1 are the layer indices
  • t is the time step. It is a tensor.
  • the fully connected layer receives global features from the global average pooling layer and outputs fault labels.
  • the fully connected layer is defined as follows:
  • FC is a fully connected layer
  • Wi is the weight matrix
  • bi is the second bias term
  • Softmax( ⁇ ) is the second activation function
  • i and i+1 are the layer numbers
  • t is the time step. It is a tensor.
  • the model can adapt to restoration data of different accuracies by mining the accuracy reference within it. Therefore, although the CNN architecture is currently the most effective fault diagnosis architecture, this accuracy-adaptive fault diagnosis model is not limited to the CNN architecture, and can also be based on architectures such as Transformer, Recurrent Neural Networks (RNN), and Multilayer Perceptron (MLP).
  • RNN Recurrent Neural Networks
  • MLP Multilayer Perceptron
  • the accuracy-adaptive fault diagnosis model provided in this application embodiment can be matched with the transmission algorithm to achieve high fault diagnosis accuracy for transmission restoration data with different accuracies.
  • the collaborative design method for high-frequency data transmission and accuracy-adaptive fault diagnosis provided in this application has the following advantages:

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • General Engineering & Computer Science (AREA)
  • Biophysics (AREA)
  • Data Mining & Analysis (AREA)
  • Biomedical Technology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Automation & Control Theory (AREA)
  • Telephonic Communication Services (AREA)
  • Data Exchanges In Wide-Area Networks (AREA)
  • Maintenance And Management Of Digital Transmission (AREA)

Abstract

本申请涉及一种高频数据传输与精度自适应故障诊断的协同设计方法，边缘网关采集数据并对数据进行预测，将预测值偏差超过阈值的实际值发给云服务器；云服务器根据实际值和自己预测值还原数据，根据还原数据和DPS长度输出故障诊断结果。本申请兼顾数据传输与故障诊断准确性，实现了有限带宽资源下准确实时的故障诊断。

Description

一种高频数据传输与精度自适应故障诊断的协同设计方法 技术领域
本申请涉及在线故障诊断领域，尤其涉及一种高频数据传输与精度自适应故障诊断的协同设计方法。
背景技术
随着对设备安全性和经济性要求的不断提高，在线故障诊断技术已逐渐成为现代工业中不可或缺的技术。近年来，随着大数据和人工智能的蓬勃发展，故障诊断技术已从专家导向转向数据驱动。然而，基于数据驱动的在线故障诊断技术需要持续的大规模高频数据传输。振动信号等高频数据的采样频率至少为1kHz，它们的实时传输将消耗难以承受的带宽资源，从而导致严重的网络拥塞和高网络延迟。同时，数据质量将直接影响故障诊断的准确性。如何在不降低故障诊断准确性的前提下，有效减少实时高频数据传输量，正成为故障诊断技术在行业实践中的瓶颈。
近年来，人们在减少在线数据传输量方面做了很多努力，这些工作大多只能应用于简单的数据，而且数据的可用性无法保证。双预测方案(Dual Prediction Scheme，DPS)可以减少实时数据传输量，同时确保数据精度的确定性范围。其中，与基于传统预测模型的DPS相比，基于深度学习(Deep Learn，DL)预测模型的DPS可以应对更复杂的数据，获得更高的数据预测精度和传输减少率，它们正在成为主流。
近来，许多著作将DL应用于故障诊断领域，并着重解决工业场景中的实际问题。Yu等人提出了一种基于迁移学习的DL模型，用于诊断无已知标签空间的无标签样本的机械故障。Chai等人针对领域和类别不一致的故障诊断问题提出了一种多源精简转移网络。Jiao等人提出了旋转机械的无源自适应诊断，它可以使用无标记数据不断优化和更新故障诊断模型。Shao等人提出了一种双阈值注意力引导的生成对抗网络，为样本有限的故障诊断提供了一种可移植的数据生成解决方案。这些方法还存在一定的不足，并没有考虑数据传输对在线故障诊断的影响。
另外，为了实现数据传输量与故障诊断准确性之间的平衡，面向应用的设计提高传输效率已成为通信研究的新方向。当前，面向应用设计以提高传输效率已成为通信研究的新方向。语义导向、任务导向和目标导向通信引起了学术界和产业界的广泛关注。Feng等人提出了一种面向目标的带宽分配框架。通过优化应用传输数据的信息效用增益，有效提高了网络物理系统中典型应用的性能。Wang等人提出了一种文本数据传输的语义通信框架。通过提取和传输以知识图谱为模型的语义信息，有效地减少了数据传输量并保证了文本数据的语义相似性。Yang等人提出了一种边缘驱动的语义提取方案，利用边缘智能满足语义通信的计算、存储和通信要求。
公开号为CN114640695A的中国发明专利申请公开了一种智能工厂基于长序列双预测和Informer的高频时序数据有效传输方法，该方法首先搭建云边协同长序列双预测架构，然后在该架构的边缘网关和云服务器部署训练好的长序列预测模型，最后采用长序列双预测方法(Long Sequence Dual Prediction Scheme，L-DPS)在线地减少高频数据的传输量，并保证数据的精度。改进了传统双预测方法的结构，通过长序列预测降低了预测模型的推理次数，从而大大提高了传统方法的适用频率，使其用于降低智能制造过程中所需要高频数据的传输量成为可能。同时，引入并结合最新深度学习模型Informer解决了长序列预测带来的梯度消失和模型推理时间激增的问题，从而进一步提高了所提长序列双预测方法的传输量减少比例和适用频率。
现有的在线故障诊断方法需要振动等高频信号的实时可靠传输，如何在有限带宽下，实现准确、实时的故障诊断，是故障诊断领域目前所面临的实际，重要和关键问题。
因此，本领域的技术人员致力于开发一种高频数据传输与精度自适应故障诊断的协同设计方法。
发明内容
有鉴于现有技术的上述缺陷，本申请所要解决的技术问题是如何在有限带宽下，实现准确和实时的故障诊断。
为实现上述目的，本申请提供了一种高频数据传输与精度自适应故障诊断的协同设计方法，所述方法包括以下步骤：
S101：边缘网关采集高频数据，对所述高频数据进行预处理，将经过预处理后的所述高频数据写入数据集，并发送给云服务器；
S102：所述云服务器接收所述边缘网关发送的数据集，并利用所述数据集训练Informer长序列预测模型；
S103：所述云服务器和所述边缘网关同步加载训练好的所述Informer长序列预测模型；
S104：所述边缘网关运行长序列双预测方法，对所述高频数据进行预测，计算预测值和实际值的偏差，将所述偏差超过阈值的所述实际值发送给所述云服务器；
S105：所述云服务器运行所述长序列双预测方法，对所述高频数据进行预测，并根据接收到的数据对所述高频数据进行还原；
S106：所述云服务器将还原后的所述高频数据和所述高频数据对应的DPS长度输入至精度自适应故障诊断模型，所述精度自适应故障诊断模型在线输出故障诊断结果。
进一步地，在所述步骤S101中，所述边缘网关对所述高频数据进行预处理包括：时间序列转换和归一化。
进一步地，所述时间序列转换是为所述高频数据的每个采样值在线添加时间戳，以标记所述高频数据的采集时间。
进一步地，所述时间戳包括年、月、日、时、分、秒和毫秒信息，所述时间戳的格式为“yyyy-MM-dd HH:mm:ss.SSS”。
进一步地，所述归一化是将所述高频数据的数值正则化到-1和1之间的数值。
进一步地，所述归一化利用反正切函数归一化方法来完成，所述反正切函数归一化使用如下函数进行处理：
式中，xi为归一化后的值，x为原始值。
进一步地，在所述步骤S102中，所述云服务器在接收到所述边缘网关发送的所述数据集后，将所述数据集划分为训练集、验证集和测试集。
进一步地，在所述步骤S102中，所述云服务器利用所述数据集训练所述Informer长序列预测模型包括如下子步骤：
S1021：所述云服务器利用所述训练集对所述Informer长序列预测模型进行训练；
S1022：所述云服务器利用所述验证集调整所述Informer长序列预测模型的超参数；
S1023：所述云服务器在所述测试集上验证效果，得到最优预测模型；
S1024：所述云服务器将训练得到的所述最优模型保存为文件。
进一步地，在所述步骤S103中，所述云服务器将所述Informer长序列预测模型的所述最优预测模型发送给所述边缘网关，并与所述边缘网关同步加载所述最优预测模型。
进一步地，所述步骤S104包括如下子步骤：
S1041：所述边缘网关采集所述高频数据，并对所述高频数据进行所述预处理，得到所述高频数据的所述实际值；
S1042：所述边缘网关利用所述Informer长序列预测模型对所述高频数据进行预测，得到所述高频数据的第一预测值；
S1043：所述边缘网关计算所述第一预测值和所述实际值之间的所述偏差，并判断所述偏差是否超过所述阈值；
S1044：如果所述偏差超过所述阈值，则所述边缘网关发送所述高频数据的所述实际值；
S1045：所述边缘网关利用所述实际值和所述第一预测值，还原所述高频数据。
进一步地，所述步骤S105包括如下子步骤：
S1051：所述云服务器利用所述Informer长序列预测模型对所述高频数据进行预测，得到所述高频数据的第二预测值；
S1052：所述云服务器接收所述边缘网关发送的所述高频数据的所述实际值；
S1053：所述云服务器利用所述实际值和所述第二预测值，还原所述高频数据。
进一步地，在所述步骤S106中，所述精度自适应故障诊断模型是基于一维卷积神经网络、变换器、循环神经网络和多层感知机中的任意一种架构。
进一步地，所述精度自适应故障诊断模型是基于所述一维卷积神经网络架构，所述一维卷积神经网络架构包括块堆叠、全局平均池化层和全连接层，并且被配置为能够完成振动信号的局部特征和全局特征的提取。
进一步地，所述块堆叠是由多个区块相互堆叠而成，所述多个区块中的每一个区块包括两个卷积层和一个最大池化层。
进一步地，所述两个卷积层用来提取所述振动信号的所述局部特征，所述卷积层定义为：
式中，是张量，Wi是卷积滤波器，bi是第一偏置项，ReLU(·)是第一激活函数，i、i+1是层序号，t是时间步，Conv1d是卷积层。
进一步地，所述最大池化层用来提取所述振动信号的高层次特征，所述最大池化层定义如下：
式中，MaxPool是最大池化层，max为取最大值，i、i+1是层序号，t是时间步，是张量。
进一步地，所述全局平均池化层用来接收所述块堆叠输出的所述高层次特征，提取所述振动信号的所述全局特征，所述全局平均池化层定义如下：
式中，GlobalAvgPool是全局平均池化层，j是序号，n是模型输入数据的长度，i、i+1是层序号，t是时间步，是张量。
进一步地，所述全连接层用来接收所述全局平均池化层输出的所述全局特征，所述全连接层被配置为能够输出故障标签，所述全连接层定义如下：
式中，FC是全连接层，Wi是权重矩阵，bi是第二偏置项，Softmax(·)是第二激活函数，i、i+1是层序号，t是时间步，是张量。
进一步地，所述步骤S106包括如下子步骤：
S1061：所述云服务器将所述还原后的所述高频数据和所述高频数据对应的所述DPS长度，输入至所述精度自适应故障诊断模型；
S1062：所述块堆叠提取所述振动信号的所述局部特征和所述高层次特征；
S1063：所述全局平均池化层根据所述高层次特征提取所述振动信号的所述全局特征；
S1064：所述全连接层根据所述全局特征得到所述振动信号的所述故障标签，输出故障诊断结果。
进一步地，所述高频数据对应的所述DPS长度，根据所述高频数据的频度进行配置。
和现有技术相比，本申请具有如下有益效果：
1、本申请通过对高频数据实时压缩传输算法和其支持的在线故障诊断应用的协同设计，能够实现减少数据传输量和保证故障诊断准确率之间的平衡，在实时压缩传输高频数据以满足有限带宽限制的同时，协同保证数据传输的精度和故障诊断的准确率，从而实现在有限带宽资源下准确、实时的故障诊断。
2、本申请通过利用DPS长度的先验信息作为传输还原数据的精度参考，提出的模型可以对不同精度的传输还原数据进行自适应，从而提高故障诊断的准确率。
3、本申请通过非线性的反正切函数归一化处理，平衡高频振动信号数值间的相对差异，从而使数值相对较小的样本也能得到有效的训练，有效提高高频数据实时压缩传输算法中时序预测的准确性。
以下将结合附图对本申请的构思、具体结构及产生的技术效果作进一步说明，以充分地了解本申请的目的、特征和效果。
附图说明
图1是本申请实施例的协同设计方法的流程示意图；
图2是本申请实施例的数据传输和故障诊断协同设计架构示意图；
图3是本申请实施例的精度自适应故障诊断模型示意图；
图4是本申请实施例的DPS长度为10时的还原数据精度示意图；
图5是本申请实施例的DPS长度为50时的还原数据精度示意图。
具体实施方式
以下参考说明书附图介绍本申请的多个优选实施例，使其技术内容更加清楚和便于理解。本申请可以通过许多不同形式的实施例来得以体现，本申请的保护范围并非仅限于文中提到的实施例。
在附图中，结构相同的部件以相同数字标号表示，各处结构或功能相似的组件以相似数字标号表示。附图所示的每一组件的尺寸和厚度是任意示出的，本申请并没有限定每个组件的尺寸和厚度。为了使图示更清晰，附图中有些地方适当夸大了部件的厚度。
现有的在线故障诊断方法需要振动等高频信号的实时可靠传输，但有限的带宽资源难以满足高频率和实时性要求。为了实现数据传输量与故障诊断准确性之间的平衡，公开号为CN114640695A的中国发明专利申请公开了一种智能工厂基于长序列双预测L-DPS和Informer的高频时序数据有效传输方法，该方法首先搭建云边协同长序列双预测架构，然后在该架构的边缘网关和云服务器部署训练好的长序列预测模型，最后采用长序列双预测方法L-DPS在线地减少高频数据的传输量，并保证数据的精度。改进了传统双预测方法的结构，通过长序列预测降低了预测模型的推理次数，从而大大提高了传统方法的适用频率，使其用于降低智能制造过程中所需要高频数据的传输量成为可能。同时，引入并结合最新深度学习模型Informer解决了长序列预测带来的梯度消失和模型推理时间激增的问题，从而进一步提高了所提长序列双预测方法的传输量减少比例和适用频率。
针对现有的在线故障诊断方法无法在有限的带宽资源下，同时兼顾高频数据传输量与故障诊断准确性之间的平衡，本申请提出了一种高频数据传输与精度自适应故障诊断的协同设计方法，该方法协同设计了在线故障诊断应用和高频数据传输算法，实现了在有限带宽资源下准确、实时的故障诊断。
如图1所示，本申请实施例提出的一种高频数据传输与精度自适应故障诊断的协同设计方法，包括以下步骤：
步骤1：边缘网关采集高频数据，对高频数据进行预处理，将经过预处理后的高频数据写入数据集，并发送给云服务器。
在线故障诊断需要实时传输振动等高频信号，这些物理世界的模拟信号由传感器转换为特定范围内的电平，然后通过高速ADC模块转换成数字信号。这些数字信号通过连接边缘网关对应接口，如并口，以有线高速线缆传输到边缘网关，网关通过相应的驱动和对应的传输协议解码获得原始数据。获得原始数据后，对原始数据进行在线预处理。预处理可以包括时间序列转换和反正切函数归一化。具体包括：
1.时间序列转换：为每个采样值在线添加时间戳，以标记其采集时间。时间戳的格式为“yyyy-MM-dd HH:mm:ss.SSS”，即“年-月-日-时-分-秒-毫秒”。
2.反正切函数归一化：由于采样值变化很大，数值小的样本无法得到有效训练，这将影响时间序列预测的准确性，传统的归一化方法无法减少数据值之间的相对差异。因此，我们采用非线性的反正切函数归一化方法来平衡数值之间的相对差异，从而使数值相对较小的样本也能得到有效的训练。反正切函数归一化能将数值正则化到-1和1之间。同时，数值之间的相对差异也被有效缩小。
反正切函数归一化使用如下函数进行处理：
式中，xi为归一化后的值，x为原始值。
边缘网关在完成对高频数据与处理后，边缘网关将预处理后的高频数据写入数据集文件，当数据集文件达到足够数量后，边缘网关再将数据集发送到云服务器。
步骤2：云服务器接收边缘网关发送的数据集，并利用该数据集训练Informer长序列预测模型。
云服务器在接收到数据集后，将其划分为训练集，验证集和测试集。利用训练集对Informer长序列预测模型进行训练，然后通过验证集调整模型超参数，最后在测试集上验证效果，得到最优预测模型，并将该最优预测模型保存为*.pkl文件。
Informer长序列预测模型的作用是为长序列双预测方法(L-DPS)提供准确快速地长序列时序预测。历史时序数据被输入到Informer长序列预测模型中，通过Informer长序列预测模型的预测推理，输出对未来时序数据的预测值。
云服务器在训练Informer长序列预测模型时，包括如下子步骤：
步骤2.1：云服务器利用训练集对Informer长序列预测模型进行训练；
步骤2.2：云服务器利用验证集调整Informer长序列预测模型的超参数；
步骤2.3：云服务器在测试集上验证效果，得到最优预测模型；
步骤2.4：云服务器将训练得到的最优模型保存为文件。
步骤3：云服务器和边缘网关同步加载训练好的Informer长序列预测模型。
云服务器将最优预测模型的模型文件发送给边缘网关，并与边缘网关同步加载该最优预测模型，即可实现对高频数据的预测。
步骤4：边缘网关运行长序列双预测方法，对高频数据进行预测，计算预测值和实际值的偏差，将偏差超过阈值的实际值发送给云服务器。
在边缘网关和云服务器运行长序列双预测方法对高频数据进行预测之前，需要提前配置L-DPS中的DPS长度参数，该参数会影响传输量减少比例和还原数据精度要求，DPS长度参数需要人工配置。DPS长度参数的确定，可以结合高频数据的频率来确定。DPS长度参数为一个具体的数值，比如10或50，DPS长度不同，还原后的数据精度也存在一定的差异。一般来说，DPS长度越小，数据还原精度越高。当边缘网关和云服务器成功加载相同的预测模型后，边缘网关实时采集和预处理原始数据得到初始值，并先向云服务器发送模型输入长度n个初始值，作为预测模型的初始输入。之后，边缘网关便可以将采集和预处理得到的初始值，使用L-DPS实时压缩并传输给云服务器。
在高频数据实时压缩传输算法L-DPS中，L-DPS压缩算法包括双预测，确认和还原三个机制，其中，双预测机制利用Informer长序列预测模型来进行预测，边缘网关根据预测结果来组织上传数据；确认机制筛选判断预测值偏差大于阈值的部分真实值，如果预测结果不准确，每个采样值中的数据点预测偏差都大于阈值，则需要传输所有真实值，也就是高频数据的采样值，L-DPS没有压缩任何高频数据；如果预测非常准确，所有采样值中的数据点预测偏差全少于阈值，则不需要上传任何采样值，L-DPS压缩效率最高。具体的L-DPS算法实现细节，可以参考CN114640695A公开的技术方案，本申请对L-DPS算法没有进行优化。
本步骤包括如下子步骤：
步骤4.1：边缘网关采集高频数据，并对高频数据进行预处理，得到高频数据的采样值；
边缘网关对高频数据的预处理，和步骤1中的预处理方法一致，主要包括时间序列转换和反正切函数归一化处理，经过预处理后的初始高频数据，就得到高频数据的采样值。
步骤4.2：边缘网关利用Informer长序列预测模型对高频数据进行预测，得到高频数据的第一预测值；
步骤4.3：边缘网关计算第一预测值和采样值之间的偏差，并判断偏差是否超过阈值；
步骤4.4：如果偏差超过阈值，则边缘网关发送高频数据的实际值；
步骤4.5：边缘网关利用实际值和预测值，还原高频数据。
边缘网关对高频数据进行预测得到的预测值，可能会和采样值之间存在一定的偏差，如果偏差没有超过阈值，边缘网关不需要发送实际的采样值，否则，边缘网关需要发送实际的采样值，云服务器根据这个实际的采样值来还原高频数据。
步骤5：云服务器运行长序列双预测方法，对高频数据进行预测，并根据接收到的数据对高频数据进行还原。
由于云服务器和边缘网关加载相同的Informer长序列预测模型，云服务器对高频数据的预测结果，应该和边缘网关的预测结果基本一致，因此，云服务器可以使用L-DPS接收边缘网关发送的数据，并对压缩数据进行还原。当边缘网关预测结果和实际采样值一致时，边缘网关并不需要上报实际值，因此，在此种情况下，云服务器在还原数据时，直接使用云服务器的预测值来还原。
本步骤包括如下子步骤：
步骤5.1：云服务器利用Informer长序列预测模型对高频数据进行预测，得到高频数据的第二预测值；
步骤5.2：云服务器接收边缘网关发送的高频数据的实际值；
步骤5.3：云服务器利用实际值和第二预测值，还原高频数据。
步骤6：云服务器将还原后高频数据和对应的DPS长度输入给精度自适应故障诊断模型，精度自适应故障诊断模型在线输出故障诊断结果。
精度自适应故障诊断模型基于一维卷积神经网络架构，包括块堆叠、全局平均池化层和全连接层，完成振动信号的局部特征和全局特征的提取。块堆叠由多个区块相互堆叠而成，一个区块包括两个卷积层和一个最大池化层。
本步骤包括如下子步骤：
步骤6.1：云服务器将还原后高频数据和对应的DPS长度输入给精度自适应故障诊断模型；
步骤6.2：块堆叠提取振动信号的局部特征和高层次特征；
步骤6.3：全局平均池化层根据高层次特征提取振动信号的全局特征；
步骤6.4：全连接层根据全局特征得到振动信号的故障标签，输出故障诊断结果。
本实施例提供的协同高频数据实时压缩传输算法和其传输还原数据所支持的在线故障诊断应用，以同时实现高传输减少率和故障诊断准确率。本实施例提出了一种传输与诊断的协同设计架构，以平衡数据传输量减少与故障诊断准确率保证，从而实现有限带宽资源下准确、实时的故障诊断。针对传输还原数据的精度差异，提出了精度自适应故障诊断模型，它可以利用高频数据实时压缩传输算法L-DPS中DPS长度的先验信息作为精度参考，对不同精度的传输还原数据进行自适应，从而有效提高故障诊断的准确率。针对高频振动信号采样值剧烈变化导致L-DPS中时序预测模型无法有效训练的问题，提出使用反正切函数对数据进行非线性归一化预处理。通过非线性的反正切函数归一化平衡高频振动信号数值间的相对差异，从而使数值相对较小的样本也能得到有效的训练，有效提高L-DPS中时序预测的准确性，进一步提高传输减少率。本申请解决了有限带宽资源下准确、实时故障诊断的实际，重要和关键问题，具有实用性，先进性和有效性。
如图2所示，本申请实施例提供的一种数据传输和故障诊断协同设计架构，该架构协同数据传输算法和故障诊断应用，以实现减少高频数据传输量与保证在线故障诊断准确率之间的平衡。在硬件层面，采用云边协同的方式，在云服务器上进行预测模型的训练，有效利用云上资源进行预测模型的训练，并在预测模型训练完成后，和边缘网关同步加载训练后的模型，边缘网关实时采集设备产生的高频数据，并在边缘侧实时进行采集的高频数据进行预处理，并通过加载的预测模型进行预测。其中，DPS长度连接了传输算法和故障诊断应用。利用其中的精度参考，精度自适应故障诊断模型可以适应不同精度的还原数据，有效提高故障诊断的准确率。通过传输与诊断之间的协同作用，该架构可同时实现较高的传输减少率和故障诊断准确率。
在本实施例中，边缘网关负责对设备侧高频数据的采集，并对采集后的高频数据进行预处理，该预处理包括时间序列转换和反正切函数归一化处理，并将预处理后的数据发送给云服务器，作为训练Informer长序列预测模型的数据集。云服务器负责Informer长序列预测模型的训练，通过边缘网关发送的测试用数据集，云服务器完成Informer长序列预测模型的训练，并将训练好的Informer长序列预测模型发送给边缘网关。在完成Informer长序列预测模型训练后，云服务器和边缘网关同时加载训练好的Informer长序列预测模型，对高频数据进行预测。
在对高频数据进行预测时，云服务器和边缘网关同时运行高频数据实时压缩传输算法L-DPS，各自对高频数据进行预测，根据预测结果进行数据还原。
云服务器在还原出高频数据后，云服务器将还原后高频数据和对应的DPS长度输入给精度自适应故障诊断模型，精度自适应故障诊断模型在线输出故障诊断结果。
现有的在线故障诊断方法需要振动等高频信号的实时可靠传输，在有限的带宽资源难以满足高频率和实时性要求，然而，通过减少数据传输量来缓解带宽压力会带来传输数据的精度偏差，降低故障诊断的准确率。本申请提出的数据传输与故障诊断的协同设计架构，通过对高频数据实时压缩传输算法L-DPS和其支持的在线故障诊断应用的协同设计，以同时实现高传输减少率和故障诊断准确率。通过对高频数据实时压缩传输算法和其支持的在线故障诊断应用的协同设计，能够实现减少数据传输量和保证故障诊断准确率之间的平衡。在实时压缩传输高频数据以满足有限带宽限制的同时，协同保证数据传输的精度和故障诊断的准确率。
如图3所示，本申请实施例提出的一种精度自适应故障诊断模型，用于解决L-DPS带来的传输还原数据的精度差异问题，该模型基于一维卷积神经网络(Convolutional Neural Network，CNN)架构。与传统的CNN不同，本申请考虑了数据传输时模型输入的精度差异。利用L-DPS中的DPS长度连接数据传输和诊断，为模型输入数据的精度提供了强有力的参考，该模型可以适应不同精度的还原数据，从而在减少数据传输，同时有效提高故障诊断的准确率。该精度自适应故障诊断模型包括卷积层、最大池化层、全局平均池化层和全连接层。该模型的主要部分是块堆叠，一个区块由两个卷积层和一个最大池化层组成，多个区块相互堆叠。该模型挖掘DPS长度中包含的有关振动信号精度的先验信息，在识别输入振动值精度状态的同时，提取振动信号的局部和全局特征。其中，DPS长度是L-DPS算法中一个由用户自己设置的参数，这个参数会影响L-DPS算法还原出来的振动信号的精度。精度自适应故障诊断模型能够通过训练，挖掘和学习DPS长度与L-DPS还原数据精度间的这种关系，并在推理时利用DPS长度所提供的还原数据精度的先验信息进行更准确的故障诊断。该模型有效降低了数据的维度和噪声，增强了信号的表示性。同时，它还能通过多层非线性变换学习复杂的故障模式，从而提高诊断的准确性和鲁棒性。如图4和图5所示，不同DPS长度，还原出来的数据精度是不同的，可以看到两条曲线是不同的。图4中的DPS长度为10时，还原后的数据更加精确；而图5中的DPS长度为50时，曲线比较平缓和不精确，很多特征平滑掉了。因此，在本申请中，考虑到不同DPS长度还原出来的数据精度有明显差异，所以本申请才需要进行精度适应，这也是本申请考虑精度自适应的动机。
在本实施例中，精度自适应故障诊断模型基于一维卷积神经网络架构，包括块堆叠、全局平均池化层和全连接层，完成振动信号的局部特征和全局特征的提取。
其中，块堆叠由多个区块相互堆叠而成，一个区块包括两个卷积层和一个最大池化层。卷积层用来提取振动信号的局部特征，卷积层定义为：
式中，是张量，Wi是卷积滤波器，bi是第一偏置项，ReLU(·)是第一激活函数，i、i+1是层序号，t是时间步，Conv1d是卷积层。
最大池化层用来提取振动信号的高层次特征，最大池化层定义如下：
式中，MaxPool是最大池化层，max为取最大值，i、i+1是层序号，t是时间步，是张量。
全局平均池化层接收块堆叠输出的高层次特征，提取振动信号的全局特征，全局平均池化层定义如下：
式中，GlobalAvgPool是全局平均池化层，j是序号，n是模型输入数据的长度，i、i+1是层序号，t是时间步，是张量。
全连接层接收全局平均池化层输出的全局特征，全连接层输出故障标签，全连接层定义如下：
式中，FC是全连接层，Wi是权重矩阵，bi是第二偏置项，Softmax(·)是第二激活函数，i、i+1是层序号，t是时间步，是张量。
在本申请另一个实施例中，考虑到精度自适应故障诊断模型的核心在于输入还原数据的同时输入其对应的DPS长度，使得模型能够通过挖掘其中的精度参考自适应不同精度的还原数据。因此，虽然CNN架构是目前最有效的故障诊断架构，但该精度自适应故障诊断模型不局限于CNN架构，也可以基于变换器(Transformer)，循环神经网络(Recurrent Neural Networks，RNN)，多层感知机(Multilayer Perceptron，MLP)等架构。
本申请实施例提供的精度自适应故障诊断模型，能够与传输算法相匹配，以实现对不同精度的传输还原数据的高故障诊断准确率。
和现有技术相比，本申请提供的高频数据传输与精度自适应故障诊断的协同设计方法，具有如下有益效果：
1、针对现有的在线故障诊断方法需要振动等高频信号的实时可靠传输，有限的带宽资源难以满足高频率和实时性要求，通过减少数据传输量来缓解带宽压力会带来传输数据的精度偏差，降低故障诊断的准确率。为了在有限的带宽资源下实现准确、实时的故障诊断，本申请提出的一种数据传输与故障诊断的协同设计架构，通过对高频数据实时压缩传输算法L-DPS和其支持的在线故障诊断应用的协同设计，以同时实现高传输减少率和故障诊断准确率。通过对高频数据实时压缩传输算法和其支持的在线故障诊断应用的协同设计，能够实现减少数据传输量和保证故障诊断准确率之间的平衡。在实时压缩传输高频数据以满足有限带宽限制的同时，协同保证数据传输的精度和故障诊断的准确率，实现了减少数据传输量与保证故障诊断准确性之间的平衡，从而实现有限带宽资源下准确、实时的故障诊断。
2、现有的故障诊断模型没有考虑模型输入数据的精度差异，对于数据压缩传输后带来的还原数据精度差异和损失没有适应性，无法实现准确的故障诊断。本申请针对传输恢复数据的精度差异，提出了精度自适应故障诊断模型。通过利用DPS长度的先验信息作为传输还原数据的精度参考，提出的模型可以对不同精度的传输还原数据进行自适应，从而提高故障诊断的准确率。
3、现有方案中普遍采用线性归一化的方法对数据进行预处理，无法减少数据值之间的相对差异。对于高频振动信号，由于采样值变化很大，数值小的样本无法得到有效训练，难以实现准确的时间序列预测。针对高频振动信号采样值剧烈变化导致时序预测模型无法有效训练的问题，本申请提出使用反正切函数对数据进行非线性归一化预处理。通过非线性的反正切函数归一化平衡高频振动信号数值间的相对差异，从而使数值相对较小的样本也能得到有效的训练，有效提高高频数据实时压缩传输算法中时序预测的准确性。
以上详细描述了本申请的较佳具体实施例。应当理解，本领域的普通技术人员无需创造性劳动就可以根据本申请的构思作出诸多修改和变化。因此，凡本技术领域中技术人员依本申请的构思在现有技术的基础上通过逻辑分析、推理或者有限的实验可以得到的技术方案，皆应在由权利要求书所确定的保护范围内。

Claims (20)

  1. 一种高频数据传输与精度自适应故障诊断的协同设计方法，所述方法包括以下步骤：
    S101：边缘网关采集高频数据，对所述高频数据进行预处理，将经过预处理后的所述高频数据写入数据集，并发送给云服务器；
    S102：所述云服务器接收所述边缘网关发送的所述数据集，并利用所述数据集训练Informer长序列预测模型；
    S103：所述云服务器和所述边缘网关同步加载训练好的所述Informer长序列预测模型；
    S104：所述边缘网关运行长序列双预测方法，对所述高频数据进行预测，计算预测值和实际值的偏差，将所述偏差超过阈值的所述实际值发送给所述云服务器；
    S105：所述云服务器运行所述长序列双预测方法，对所述高频数据进行预测，并根据接收到的数据对所述高频数据进行还原；
    S106：所述云服务器将还原后的所述高频数据和所述高频数据对应的DPS长度输入至精度自适应故障诊断模型，所述精度自适应故障诊断模型在线输出故障诊断结果。
  2. 如权利要求1所述的方法，其中，在所述S101中，所述边缘网关对所述高频数据进行预处理包括：时间序列转换和归一化。
  3. 如权利要求2所述的方法，其中，所述时间序列转换是为所述高频数据的每个采样值在线添加时间戳，以标记所述高频数据的采集时间。
  4. 如权利要求3所述的方法，其中，所述时间戳包括年、月、日、时、分、秒和毫秒信息，所述时间戳的格式为“yyyy-MM-dd HH:mm:ss.SSS”。
  5. 如权利要求4所述的方法，其中，所述归一化是将所述高频数据的数值正则化到-1和1之间的数值。
  6. 如权利要求5所述的方法，其中，所述归一化利用反正切函数归一化方法来完成，所述反正切函数归一化使用如下函数进行处理：
    式中，xi为归一化后的值，x为原始值。
  7. 如权利要求6所述的方法，其中，在所述S102中，所述云服务器在接收到所述边缘网关发送的所述数据集后，将所述数据集划分为训练集、验证集和测试集。
  8. 如权利要求7所述的方法，其中，在所述S102中，所述云服务器利用所述数据集训练所述Informer长序列预测模型包括如下子步骤：
    S1021：所述云服务器利用所述训练集对所述Informer长序列预测模型进行训练；
    S1022：所述云服务器利用所述验证集调整所述Informer长序列预测模型的超参数；
    S1023：所述云服务器在所述测试集上验证效果，得到最优预测模型；
    S1024：所述云服务器将训练得到的所述最优模型保存为文件。
  9. 如权利要求8所述的方法，其中，在所述S103中，所述云服务器将所述Informer长序列预测模型的所述最优预测模型发送给所述边缘网关，并与所述边缘网关同步加载所述最优预测模型。
  10. 如权利要求9所述的方法，其中，所述S104包括如下子步骤：
    S1041：所述边缘网关采集所述高频数据，并对所述高频数据进行所述预处理，得到所述高频数据的所述实际值；
    S1042：所述边缘网关利用所述Informer长序列预测模型对所述高频数据进行预测，得到所述高频数据的第一预测值；
    S1043：所述边缘网关计算所述第一预测值和所述实际值之间的所述偏差，并判断所述偏差是否超过所述阈值；
    S1044：如果所述偏差超过所述阈值，则所述边缘网关发送所述高频数据的所述实际值；
    S1045：所述边缘网关利用所述实际值和所述第一预测值，还原所述高频数据。
  11. 如权利要求10所述的方法，其中，所述S105包括如下子步骤：
    S1051：所述云服务器利用所述Informer长序列预测模型对所述高频数据进行预测，得到所述高频数据的第二预测值；
    S1052：所述云服务器接收所述边缘网关发送的所述高频数据的所述实际值；
    S1053：所述云服务器利用所述实际值和所述第二预测值，还原所述高频数据。
  12. 如权利要求11所述的方法，其中，在所述S106中，所述精度自适应故障诊断模型是基于一维卷积神经网络、变换器、循环神经网络和多层感知机中的一种架构。
  13. 如权利要求12所述的方法，其中，所述精度自适应故障诊断模型是基于所述一维卷积神经网络架构，所述一维卷积神经网络架构包括块堆叠、全局平均池化层和全连接层，并且被配置为能够完成振动信号的局部特征和全局特征的提取。
  14. 如权利要求13所述的方法，其中，所述块堆叠是由多个区块相互堆叠而成，所述多个区块中的每一个区块包括两个卷积层和一个最大池化层。
  15. 如权利要求14所述的方法，其中，所述两个卷积层用来提取所述振动信号的所述局部特征，所述卷积层定义为：
    式中，是张量，Wi是卷积滤波器，bi是第一偏置项，ReLU(·)是第一激活函数，i、i+1是层序号，t是时间步，Conv1d是卷积层。
  16. 如权利要求15所述的方法，其中，所述最大池化层用来提取所述振动信号的高层次特征，所述最大池化层定义如下：
    式中，MaxPool是最大池化层，max为取最大值，i、i+1是层序号，t是时间步，是张量。
  17. 如权利要求16所述的方法，其中，所述全局平均池化层用来接收所述块堆叠输出的所述高层次特征，提取所述振动信号的所述全局特征，所述全局平均池化层定义如下：
    式中，GlobalAvgPool是全局平均池化层，j是序号，n是模型输入数据的长度，i、i+1是层序号，t是时间步，是张量。
  18. 如权利要求17所述的方法，其中，所述全连接层用来接收所述全局平均池化层输出的所述全局特征，所述全连接层被配置为能够输出故障标签，所述全连接层定义如下：
    式中，FC是全连接层，Wi是权重矩阵，bi是第二偏置项，Softmax(·)是第二激活函数，i、i+1是层序号，t是时间步，是张量。
  19. 如权利要求18所述的方法，其中，所述S106包括如下子步骤：
    S1061：所述云服务器将所述还原后的所述高频数据和所述高频数据对应的所述DPS长度，输入至所述精度自适应故障诊断模型；
    S1062：所述块堆叠提取所述振动信号的所述局部特征和所述高层次特征；
    S1063：所述全局平均池化层根据所述高层次特征提取所述振动信号的所述全局特征；
    S1064：所述全连接层根据所述全局特征得到所述振动信号的所述故障标签，输出故障诊断结果。
  20. 如权利要求19所述的方法，其中，所述高频数据对应的所述DPS长度，根据所述高频数据的频度进行配置。
PCT/CN2024/139533 2024-07-04 2024-12-16 一种高频数据传输与精度自适应故障诊断的协同设计方法 Pending WO2026007341A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US19/220,787 US20260009673A1 (en) 2024-07-04 2025-05-28 Method of design for synergy between high-frequency data transmission and precision-adaptive fault diagnosis

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202410895486.3 2024-07-04
CN202410895486.3A CN118859715A (zh) 2024-07-04 2024-07-04 一种高频数据传输与精度自适应故障诊断的协同设计方法

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US19/220,787 Continuation-In-Part US20260009673A1 (en) 2024-07-04 2025-05-28 Method of design for synergy between high-frequency data transmission and precision-adaptive fault diagnosis

Publications (1)

Publication Number Publication Date
WO2026007341A1 true WO2026007341A1 (zh) 2026-01-08

Family

ID=93160167

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2024/139533 Pending WO2026007341A1 (zh) 2024-07-04 2024-12-16 一种高频数据传输与精度自适应故障诊断的协同设计方法

Country Status (3)

Country Link
US (1) US20260009673A1 (zh)
CN (1) CN118859715A (zh)
WO (1) WO2026007341A1 (zh)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118859715A (zh) * 2024-07-04 2024-10-29 上海交通大学 一种高频数据传输与精度自适应故障诊断的协同设计方法
CN121179477B (zh) * 2025-11-25 2026-02-24 广安职业技术学院 一种工业机器人实时故障检测方法及系统

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112163668A (zh) * 2020-09-29 2021-01-01 上海交通大学 一种基于预测和云边协同的减少时间序列数据传输量方法
CN114640695A (zh) * 2022-04-24 2022-06-17 上海交通大学 一种智能工厂基于长序列双预测和informer的高频时序数据有效传输方法
US20220365525A1 (en) * 2021-05-12 2022-11-17 Yokogawa Electric Corporation Systems, methods, and devices for equipment monitoring and fault prediction
CN116520814A (zh) * 2023-07-03 2023-08-01 清华大学 云边协同架构下基于联邦学习的设备故障预测方法及装置
CN117973502A (zh) * 2024-01-29 2024-05-03 重庆大学 一种基于云边协同的生产设备故障诊断系统及方法
CN118859715A (zh) * 2024-07-04 2024-10-29 上海交通大学 一种高频数据传输与精度自适应故障诊断的协同设计方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112163668A (zh) * 2020-09-29 2021-01-01 上海交通大学 一种基于预测和云边协同的减少时间序列数据传输量方法
US20220365525A1 (en) * 2021-05-12 2022-11-17 Yokogawa Electric Corporation Systems, methods, and devices for equipment monitoring and fault prediction
CN114640695A (zh) * 2022-04-24 2022-06-17 上海交通大学 一种智能工厂基于长序列双预测和informer的高频时序数据有效传输方法
CN116520814A (zh) * 2023-07-03 2023-08-01 清华大学 云边协同架构下基于联邦学习的设备故障预测方法及装置
CN117973502A (zh) * 2024-01-29 2024-05-03 重庆大学 一种基于云边协同的生产设备故障诊断系统及方法
CN118859715A (zh) * 2024-07-04 2024-10-29 上海交通大学 一种高频数据传输与精度自适应故障诊断的协同设计方法

Also Published As

Publication number Publication date
US20260009673A1 (en) 2026-01-08
CN118859715A (zh) 2024-10-29

Similar Documents

Publication Publication Date Title
WO2026007341A1 (zh) 一种高频数据传输与精度自适应故障诊断的协同设计方法
CN111913849B (zh) 一种用于运维数据的无监督异常检测和鲁棒趋势预测方法
CN108536123B (zh) 基于长短时记忆神经网络的列控车载设备故障诊断方法
CN114676782A (zh) 一种基于高频时序数据的车辆故障预警方法及系统
CN111562105B (zh) 一种基于小波包分解和卷积神经网络的风电机组齿轮箱故障诊断方法
CN113239720B (zh) 一种基于深度迁移学习的地铁车辆走行部故障诊断方法
CN116502162A (zh) 边缘算力网络中的异常算力联邦检测方法、系统及介质
CN111813084A (zh) 一种基于深度学习的机械装备故障诊断方法
CN108304927A (zh) 基于深度学习的轴承故障模式诊断方法及系统
CN118885731A (zh) 基于多源数据融合的采煤机智能故障诊断系统
CN113328755B (zh) 一种面向边缘计算的压缩数据传输方法
CN115051929A (zh) 基于自监督目标感知神经网络的网络故障预测方法及装置
CN113536671B (zh) 一种基于lstm的锂电池寿命预测方法
CN113642254A (zh) 一种基于多尺度注意力机制的风力发电功率预测方法
CN113705809A (zh) 一种数据预测模型训练方法、工业指标预测方法和装置
CN116011507A (zh) 融合元学习与图神经网络的稀有故障诊断方法
CN116961674A (zh) 一种分布式光伏数据压缩方法及系统
CN119202705A (zh) 基于知识蒸馏与协同增量学习的变压器故障诊断方法
CN115293150A (zh) 融合卷积神经网络与自注意力机制手术记录自动编码方法
CN119721196A (zh) 一种基于风电模型的协同与知识传递方法及系统
CN118862011A (zh) 一种基于DIformer的航空发动机剩余使用寿命预测方法
CN116070170B (zh) 一种基于深度学习的云边端数据融合处理方法及系统
CN114021628A (zh) 基于Transformer模型的制造业多源异构数据融合方法
CN118659367B (zh) 基于EFformer模型的电力负荷预测系统及方法
CN118378514A (zh) 一种基于多模特征融合的飞机发动机剩余寿命预测方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 24946207

Country of ref document: EP

Kind code of ref document: A1