Method for processing fault diagnosis data of electromechanical equipment

Abstract

The invention relates to the technical field of information communication, and provides a method for processing fault diagnosis data of electromechanical equipment, which comprises the following steps: acquiring a local data set, calculating a stochastic gradient, and obtaining a momentum buffer function required by gradient tracking; calculating an adaptive learning rate according to the stochastic gradient; repeating the above steps, collecting the momentum buffer function and the adaptive learning rate of the client, and uploading the momentum buffer function and the adaptive learning rate to the server after a preset period is reached; and aggregating the momentum buffer function and the adaptive learning rate to obtain a global adaptive matrix, generating and outputting a model, and providing the model to a client. According to the method, optimization is carried out from client data, model optimization is carried out by integrating information of all clients, and the accuracy of equipment health assessment and fault prediction is improved while data privacy is protected.

Description

Technical Field

[0001] The present invention relates to the field of information and communication technologies, and more particularly, to a method for processing fault diagnosis data of electromechanical equipment. Background Art

[0002] All kinds of electromechanical equipment have been highly informatized and networked, and can generate and collect a large amount of equipment status data. After being processed by machine learning algorithms, these massive data can achieve intelligent diagnosis and predictive maintenance of electromechanical equipment. However, equipment status data often contains the core business information of enterprises and involves data privacy protection. In addition, the data distributions collected by different enterprises may vary, increasing the difficulty of data utilization. Therefore, how to achieve collaborative analysis of massive heterogeneous equipment status data, accurately evaluate the equipment health status, and achieve fault prediction under the premise of protecting data privacy is a difficult problem that needs to be solved urgently at present. Currently, the centralized data analysis method for equipment maintenance requires enterprises to upload local data to a third-party cloud service, which has the risk of data leakage. Uploading data after processing by methods such as encryption or differential privacy will reduce the accuracy of the model.

[0003] Federated learning is a machine learning mode that has emerged in recent years and can collaborate data of multiple enterprises without transmitting the original data. Currently, there are two main algorithms in the field of federated learning, the FedAvg algorithm and the FedAdam algorithm: 1) The FedAvg (Federated Averaging) algorithm is a basic federated learning algorithm. It allows each client to perform multiple local model trainings, and then periodically sends the parameters of the local model to the server side for model averaging. Specifically, the client uses local data to run SGD to train the model, the server collects the model parameters from all clients, averages them, and broadcasts them to all clients as the global model; 2) The FedAdam (Federated Adaptive Moment Estimation Optimization) algorithm is an improved algorithm of FedAvg. On the basis of FedAvg, the gradients calculated locally by the device are also aggregated in the central server, and the model parameters are adaptively adjusted on the server side. However, the existing algorithms: 1) require a large number of training iterations and have a slow convergence speed; 2) lack self-adaptability and rely on manual adjustment of the learning rate; 3) rely on data homogeneity, do not match the growth of the model and data scale, and have poor processing effects on heterogeneous data, making it difficult to be applied to actual predictive maintenance.

[0004] Therefore, there is an urgent need to develop a more rapid and efficient federated learning algorithm to improve the accuracy of equipment health assessment and fault prediction while protecting data privacy. Summary of the Invention

[0005] The present invention aims to provide a method for processing fault diagnosis data of electromechanical equipment to solve the technical problems of low accuracy of equipment fault prediction and poor protection of data privacy.

[0006] The technical solution adopted by the present invention to solve its technical problems is: A method for processing fault diagnosis data of electromechanical equipment, including:

[0007] S1 Client information collection

[0008] S11 Collect the local dataset D i , calculate the stochastic gradient and obtain the momentum buffer function m required for gradient tracking t,i ;

[0009] S12 Calculate the adaptive learning rate v according to the stochastic gradient in step S11 t,i ;

[0010] Repeat step S11 and step S12, collect the momentum buffer function m of t clients t,i and the adaptive learning rate, and upload them to the server side when t reaches the preset period;

[0011] S2 Aggregate the m t,i and v t,i collected from t clients to obtain the global and Obtain the global adaptive matrix, generate the output model, and provide it to the client.

[0012] The present invention optimizes the adaptive learning rate by using the client historical gradient information, uploads the optimized adaptive learning rate information and gradient information to the server, and the server then optimizes the model with the adaptive learning rate information and gradient information of different clients, and finally provides the optimized model information to the client. The present invention optimizes from the client data and comprehensively optimizes the model with the information of all clients, improving the accuracy of equipment health assessment and fault prediction while protecting data privacy.

[0013] In the present invention, it can be understood that: in each iteration cycle, the server terminal is connected to t clients, each client contains b motor devices, and the number of each motor device is i. Based on this, the following calculation description is carried out.

[0014] In the present invention, the initial information of the local dataset described in step S11 includes at least one of the current, voltage, and vibration of the motor device; the initial information is converted into a time-frequency domain image through continuous wavelet transform.

[0015] Preferably, the calculation of the stochastic gradient described in step S11 includes:

[0016] Local dataset D i includes collecting b local motor data, and the b local motor data are expressed as where i represents the electromechanical equipment number; D i represents the local dataset of the i-th electromechanical equipment;

[0017] Using the current model parameters x t,i , and using the collected local motor data B t,i to calculate the loss function f i (x t,i ; B y,i ):

[0018]

[0019] Calculate the stochastic gradient corresponding to the local motor data B in the client t,i

[0020]

[0021] The stochastic gradient of the t-th iteration is denoted as Calculate the stochastic gradient of each client i on the current model parameters x t,i

[0022]

[0023] Calculating the local stochastic gradient can generate immediate feedback on the global model parameters from the local dataset, which helps to improve the adaptability and accuracy of the model on heterogeneous data. At the same time, local processing reduces the dependence on the central server and speeds up the overall learning process.

[0024] Preferably, the calculation of the momentum buffer function m required for gradient tracking in step S11 t,i includes:

[0025] Calculate the stochastic gradient of the current step represents the stochastic gradient calculated by the i-th client point in the t-th round:

[0026]

[0027] Calculate the model parameters x of the previous step t-1,i on the current dataset B t,i stochastic gradient represents the stochastic gradient calculated by the i-th client point in the (t - 1)-th round:

[0028] ​​

[0029] Momentum buffer function m t,i :

[0030]

[0031] where α t ∈(0,1) represents the momentum parameter.

[0032] m t,i tracks gradient information by accumulating a linear combination of the current gradient and the content of the previous buffer step.

[0033] Smoothing the learning process and accelerating convergence based on historical gradient information are the main functions of tracking gradients. Momentum helps the model descend stably during training, adjusting the update direction of the current step by considering previous gradients, thereby overcoming possible training fluctuations and quickly approaching the optimal solution.

[0034] Preferably, the calculation of the adaptive learning rate in step S12 includes:

[0035] Calculating the stochastic gradient of the current step from the previous step Further obtaining the square of the current stochastic gradient:

[0036]

[0037] Calculating the adaptive learning rate v based on the square of the current stochastic gradient t,i :

[0038]

[0039] where β ∈(0,1) represents the momentum parameter.

[0040] v t,i estimates the adaptive learning rate by accumulating the exponential moving average of the historical gradient squares. By considering the gradient change amount of each client, the adaptive learning rate helps to maintain stability when facing different data distributions, avoiding slow or unstable learning caused by too large or too small learning rates.

[0041] Preferably, the global and calculation in step S2 includes:

[0042] Collecting m t,i and v t,i uploaded by each client i;

[0043] Taking the average of m t,i for all clients to obtain the global

[0044]

[0045] Take the average of v for all clients to obtain the global t,i average, obtaining the global

[0046]

[0047] Repeat the above collection and averaging operations every q steps to obtain the global and

[0048] Through the aggregation of information from all clients, the global and are obtained, preparing for generating the adaptive matrix. Synchronize and aggregate the gradient and learning rate information from all clients globally to achieve coordinated model updates. While reducing the model bias caused by uneven data distribution, it can also improve the overall model's robustness and performance through collective wisdom.

[0049] Generate the adaptive matrix in the server. The main purpose of generating the adaptive matrix is to create a global update strategy that will reflect the common contributions of different clients to the model update. The adaptive matrix aggregates the adaptive learning rate information of different clients, enabling each global update to take into account the learning characteristics of all clients, thereby optimizing the model performance and ensuring the consistency and effectiveness of the updates.

[0050] Preferably, the generation of the global adaptive matrix described in step S2 includes:

[0051]

[0052] where ρ is a constant greater than 0, used to smooth the adaptive information and prevent overfitting.

[0053] It should be noted that the adaptive matrix A t is a diagonal matrix, obtained through the global stochastic gradient to provide adaptive information for the local model updates of the clients.

[0054] Finally, perform personalized model updates on the output model according to the data of the clients to improve the accuracy and adaptability of the model on specific client data. Ensure the effective execution of the global update strategy in the local environment and enhance the model's response ability to local data characteristics.

[0055] Preferably, the generation of the output model described in step S2 includes:

[0056] Each client i performs the following operations:

[0057] Obtain the global adaptive matrix A sent by the servert , calculate its inverse matrix A t -1 ;

[0058] Use m t,i to update the global adaptive matrix as follows:

[0059]

[0060] where η t represents the learning rate.

[0061] The beneficial effects of the present invention are as follows:

[0062] Compared with the prior art, the present application provides a method for processing electromechanical equipment fault diagnosis data. By using the client historical gradient information to optimize the adaptive learning rate, and uploading the optimized adaptive learning rate information and gradient information to the server, the server then optimizes the model with the adaptive learning rate information and gradient information of different clients, and finally provides the optimized model information to the client. Optimize from the client data and comprehensively optimize the model with the information of all clients, while protecting data privacy, improving the accuracy of equipment health assessment and fault prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 is a schematic flowchart of a method for processing electromechanical equipment fault diagnosis data according to the present invention.

[0064] The drawings are only for illustrative purposes and should not be construed as a limitation of the present invention; for better illustration of the embodiments, some components in the drawings may be omitted, enlarged or reduced, and do not represent the dimensions of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0065] The following will further describe the present invention in detail with reference to the drawings and specific embodiments.

[0066] As Figure 1 shown, N electromechanical devices are regarded as N clients, and each device is assigned a number i. The time-series data collected on the electromechanical device i, such as the time-series signals of the current, voltage, vibration, etc. of the motor, are converted into time-frequency domain images through continuous wavelet transform, and this image is used as the dataset D i input into the algorithm. At the same time, the server will distribute the initial model parameters x 0 to each client for local learning of the client.

[0067] A method for processing electromechanical equipment fault diagnosis data according to the present invention includes the following steps:

[0068] 1. Calculate the local random gradient

[0069] Randomly sample a mini - batch of data from the local dataset D i where b is the size of the mini - batch. The mini - batch contains motor operation data such as current, voltage, and vibration that have been transformed into time - frequency domain images through continuous wavelet transform;

[0070]

[0070] Use the current model parameters x t,i , and use the collected local motor data B t,i to calculate the loss function f i (x t,i ; B t,i ):

[0071]

[0072] Calculate the random gradient corresponding to the local motor data B in the client t,i The random gradient at the t - th iteration is denoted as

[0073]

[0074] Calculate the random gradient of each client i on the current model parameters x t,i t,i t,i

[0075]

[0076] 2. Gradient tracking

[0077] Calculate the random gradient at the current step represents the random gradient calculated at the i - th client point in the t - th round:

[0078]

[0079] Calculate the random gradient of the model parameters x at the previous step t-1,i on the current dataset B t,i The random gradient is represents the random gradient calculated at the i - th client point in the (t - 1) - th round:

[0080]

[0081] The momentum buffer function m t,i :

[0082]

[0083] where αt ∈(0,1) represents the momentum parameter.

[0084] m t,i Track the gradient information by accumulating a linear combination of the current gradient and the content of the buffer in the previous step.

[0085] 3. Estimate the adaptive learning rate

[0086] Perform the following operations for each client i:

[0087] Calculate the stochastic gradient of the current step from the previous step Further obtain the square of the current stochastic gradient:

[0088]

[0089] Calculate the adaptive learning rate v based on the square of the current stochastic gradient t,i :

[0090]

[0091] where β ∈(0,1) represents the momentum parameter.

[0092] v t,i Estimate the adaptive learning rate by accumulating the exponential moving average of the historical gradient squares.

[0093] 4. Server-side average the adaptive information

[0094] The server performs the following operations:

[0095] Collect m t,i and v t,i ;

[0096] Average m t,i for all clients to obtain the global

[0097]

[0098] Average v t,i for all clients to obtain the global

[0099]

[0100] Repeat the above collection and averaging operations every q steps to obtain the global and

[0101] Through the aggregation of all client information, the global and Prepare to generate an adaptive matrix.

[0102] 5. Generate an adaptive matrix

[0103] The server performs the following operations. The generation of the global adaptive matrix includes:

[0104]

[0105] Among them, ρ is a constant greater than 0 (for example, it can be 1, 2, 3...), which is used to smooth the adaptive information and prevent overfitting.

[0106] It should be noted that the adaptive matrix A t is a diagonal matrix and is obtained through the global random gradient to provide adaptive information for the local model update of the client.

[0107] 6. Local client adaptive update

[0108] Each client i performs the following operations:

[0109] Obtain the global adaptive matrix A sent by the server t , and calculate its inverse matrix A t -1 ;

[0110] Use m t,i to update the global adaptive matrix as follows:

[0111]

[0112] Among them, η t represents the learning rate.

[0113] Compared with the existing federated learning algorithms, this algorithm effectively accelerates the convergence speed of the model by introducing a momentum mechanism to track and utilize historical gradient information. In addition, by collaboratively generating an adaptive matrix among clients, the number of communication rounds required for training is reduced, and the sample complexity is reduced to O(ε - 3), and the communication complexity is reduced to O(ε - 2). Specifically, the introduction of the adaptive matrix significantly improves the processing ability of heterogeneous data, enabling each client to quickly adapt to the characteristics of its own dataset, thereby accelerating the training process of the entire model. Existing federated learning algorithms usually require more communication and have a slower convergence speed. Through the above innovations, the present invention has significantly improved the communication and computing efficiency technically while ensuring the security and privacy of data, meeting the high standards of intelligent manufacturing and equipment maintenance technologies in the Industry 4.0 era. The present invention is easy to promote and apply and can be continuously upgraded. This provides an effective way to improve fault diagnosis by using massive distributed electromechanical device data.

[0114] By learning through the above process, the updated model for motor fault diagnosis has the following three main advantages compared with traditional fault detection methods.

[0115] (1) Data privacy and utilization efficiency: Traditional centralized data processing methods require centralized processing of sensitive data, which poses a risk of leakage. This patent uses the federated learning method to update the fault diagnosis model, enabling data analysis to be carried out through local computing and only sharing model parameters without sharing the original data, greatly enhancing data privacy protection. This method allows all participants to jointly optimize and update the model for fault diagnosis while ensuring data security.

[0116] (2) Adaptability to heterogeneous data: The electromechanical equipment of different enterprises may have differences in data collection standards and formats. Traditional methods are less efficient in processing such heterogeneous data. This patent adjusts the fault diagnosis model through an adaptive learning rate and a dynamic model, allowing the model to better adapt to data from different sources, thereby improving the accuracy and reliability of the model in a multi-source data environment.

[0117] (3) Real-time performance and accuracy: Using the federated learning algorithm, the model is allowed to receive and process updated data from each client in real time, quickly feedback the fault diagnosis results, speed up the fault detection speed, endow the model with the ability to integrate multi-party information, and improve the accuracy of fault judgment. In terms of predictive maintenance, the model can predict potential faults based on the real-time status data of the equipment, providing a scientific basis for maintenance decisions.

[0118] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, rather than limiting the implementation manners of the present invention. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to enumerate all implementation manners here. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. A method for processing electromechanical equipment fault diagnosis data, characterized in that: include: S1 Client Information Collection S11 collects local data set D i , calculate the stochastic gradient And obtain the momentum buffer function m required for gradient tracking t,i ; S12 calculates the adaptive learning rate v based on the stochastic gradient t,i ; Repeat steps S11 and S12 to collect the momentum buffer function m of t clients. t,i and adaptive learning rate v t,i , when t reaches the preset period, it is uploaded to the server; S2 collects the m of t clients t,i and v t,i Aggregate to get the global and Obtain the global adaptive matrix, generate and output the model, and provide it to the client.

2. A method for processing electromechanical equipment fault diagnosis data according to claim 1, characterized in that: The initial information of the local data set in step S11 includes at least one of the current, voltage, and vibration of the motor equipment; The initial information is converted into a time-frequency domain image through continuous wavelet transform.

3. A method for processing electromechanical equipment fault diagnosis data according to claim 1, characterized in that: The calculation of the stochastic gradient in step S11 includes: Local dataset D i It includes collecting b local motor data, and b local motor data is expressed as Where i represents the number of the electromechanical equipment; D i Represents the local data set of the i-th electromechanical equipment; Using the current model parameters x t,i , using the collected local motor data B t,i Calculate the loss function f i (x t,i ; B t,i ): Calculate the local motor data B in the client t,i The corresponding stochastic gradient The stochastic gradient of the tth iteration is recorded as Calculate the current model parameters x for each client i t,i The stochastic gradient on 4. A method for processing electromechanical equipment fault diagnosis data according to claim 1, characterized in that: The momentum buffer function required for gradient tracking in step S11 includes: Calculate the stochastic gradient for the current step Represents the stochastic gradient calculated by the i-th client node in the t-th round: Calculate the model parameters x of the previous step t-1,i On the current dataset B t,i The stochastic gradient Represents the stochastic gradient calculated by the i-th client node in the t-1th round: Momentum buffer function m t,i : Among them, α t ∈(0,1) represents the momentum parameter.

5. A method for processing electromechanical equipment fault diagnosis data according to claim 1, characterized in that: The step S12 of calculating the adaptive learning rate includes: Calculate the stochastic gradient of the current step from the previous step Further obtain the square of the current stochastic gradient: Calculate the adaptive learning rate v based on the square of the current stochastic gradient t,i : Among them, β∈(0,1) represents the momentum parameter.

6. A method for processing electromechanical equipment fault diagnosis data according to claim 1, characterized in that: The global and include: Collect m uploaded by each client i t,i and v t,i ; For all clients t,i Find the average and get the overall For all clients v t,i Find the average and get the overall Every q steps, repeat the above collection and averaging operations to obtain the global and 7. A method for processing electromechanical equipment fault diagnosis data according to claim 1, characterized in that: The step S2 of obtaining the global adaptive matrix includes: Here, ρ is a constant greater than 0.

8. A method for processing electromechanical equipment fault diagnosis data according to claim 1, characterized in that: The step S2 of generating and outputting the model includes: Each client i performs the following operations: Get the global adaptive matrix A sent by the server t , calculate its inverse matrix A t -1 ; Use m t,i The global adaptive matrix is ​​updated as follows: Among them, η t Represents the learning rate.

Images