Reciprocating compressor fault diagnosis method and diagnosis device based on global knowledge enhancement and electronic equipment

Abstract

The invention discloses a reciprocating compressor fault diagnosis method, which comprises the following steps of: respectively extracting diagnosis characteristics in working data of a plurality of corresponding reciprocating compressors by utilizing a plurality of edge diagnosis models so as to form a plurality of diagnosis characteristic sets; the global diagnosis model obtains a plurality of global diagnosis predicted values based on each diagnosis feature set in the plurality of diagnosis feature sets; respectively training a plurality of edge diagnosis models by using the plurality of global diagnosis predicted values, and optimizing the plurality of edge diagnosis models and the global diagnosis model through iterative learning; and inputting the real-time working data of the plurality of reciprocating compressors into the corresponding plurality of edge diagnosis models to obtain a plurality of real-time fault prediction values so as to accurately detect heterogeneous faults of the plurality of reciprocating compressors.

Description

Technical Field

[0001] The present application belongs to the technical field of intelligent fault diagnosis of reciprocating compressors, and specifically, relates to a reciprocating compressor fault diagnosis method based on global knowledge enhancement, a diagnostic device and an electronic device. Background Art

[0002] Reciprocating compressors are the most widely used type of machinery and equipment in petrochemical production process equipment, and are mainly used in oil refining, oil production, gas transportation, aerodynamics, etc. Reciprocating compressors are often in a high temperature, high pressure, flammable, explosive, corrosive and multi-excitation source working environment, and there are many problems of frequent failures. Once a component fails, it may trigger a chain reaction, causing the entire system to shut down. In severe cases, it may cause major accidents such as explosions, causing huge economic losses and adverse social impacts.

[0003] With the further development of new-generation information technologies such as industrial Internet and big data, the petrochemical industry has entered the era of digital transformation. At present, oilfield stations are generally equipped with industrial control systems and status monitoring systems such as SCADA / DCS / PLC, which can monitor the operating status of reciprocating compressors and their key components, forming a massive amount of monitoring big data, which is an important resource for revealing the evolution process and nature of equipment failures. The reciprocating compressor fault diagnosis method based on deep learning needs to centralize these monitoring data, which conflicts with the business principle that actual equipment manufacturers and users do not want to share original equipment data. How to make full use of multi-party data for multi-user collaborative modeling and diagnosis is an important problem that needs to be solved in order to realize intelligent fault diagnosis of reciprocating compressors.

[0004] Federated learning, which uses data features rather than raw data for distributed fault prediction, provides a possible solution to the above problems. However, the heterogeneity of fault modes between devices in traditional federated systems will have a great impact on the effectiveness of the global diagnostic model. Traditional federated learning methods based on average aggregation of client model parameters are difficult to cope with heterogeneous challenges. In addition, the entire process of model training in traditional federated learning methods is completed on the client, and it is usually difficult to configure abundant computing resources at the edge where reciprocating compressors are located, which causes a mismatch between demand and resources. The relatively more abundant computing resources on the cloud server side are not fully utilized, and the transmission of model parameters increases the resource communication cost. Summary of the invention

[0005] In order to solve the problems existing in the existing federated learning methods, the purpose of this application is to provide a reciprocating compressor fault diagnosis method based on global knowledge enhancement.

[0006] The reciprocating compressor fault diagnosis method based on global knowledge enhancement proposed in the present application utilizes massive working data of reciprocating compressors in multiple working scenarios to extract diagnostic feature sets, trains a global diagnostic model through a collection of multiple diagnostic feature sets and obtains a global diagnostic prediction value based on heterogeneous reciprocating compressor data, and uses the global diagnostic prediction value to train small-estimation edge diagnostic models for fault diagnosis of each reciprocating compressor, so as to improve the ability of multiple edge diagnostic models in different application scenarios to detect heterogeneous fault modes and reduce the computing power required by the edge diagnostic model, thereby improving the fault diagnosis performance of each local reciprocating compressor in a collaborative diagnosis system based on federated learning.

[0007] Specifically, this application involves the following aspects:

[0008] 1. Reciprocating compressor fault diagnosis method, including:

[0009] Extracting diagnostic features from historical working data of a plurality of corresponding reciprocating compressors using a plurality of edge diagnostic models respectively to form a plurality of diagnostic feature sets;

[0010] The global diagnosis model obtains multiple global diagnosis prediction values ​​based on each of the multiple diagnostic feature sets;

[0011] Using multiple global diagnosis prediction values ​​to train multiple edge diagnosis models respectively, and optimizing the global diagnosis model and the multiple edge diagnosis models through iterative learning; and

[0012] The real-time working data of multiple reciprocating compressors are respectively input into the corresponding optimized multiple edge diagnosis models to obtain multiple real-time fault prediction values.

[0013] 2. A method for diagnosing a reciprocating compressor fault according to item 1, wherein:

[0014] The diagnostic features extracted from the historical working data of the corresponding reciprocating compressors using the multiple edge diagnostic models include: thermal performance features, including inlet and exhaust temperatures, inlet and exhaust pressures, exhaust volume, lubricating oil circuit pressure, and cooling water circuit temperature;

[0015] Vibration monitoring and piston rod axis trajectory characteristics, including crankcase vibration parameters, crosshead vibration parameters and cylinder vibration parameters;

[0016] Among them, historical working data include thermal performance parameters, vibration data, piston rod displacement data, cylinder pressure data, piston outer dead center / key phase data, noise monitoring data and oil monitoring data.

[0017] 3. A reciprocating compressor fault diagnosis method according to item 1, wherein:

[0018] Using multiple global diagnosis prediction values ​​to train multiple edge diagnosis models respectively, optimizing the global diagnosis model and multiple edge diagnosis models through iterative learning includes:

[0019] Outputting each of the multiple global diagnosis prediction values ​​to the corresponding multiple edge diagnosis models;

[0020] Each edge diagnostic model in the plurality of edge diagnostic models determines a diagnostic loss function value using each corresponding diagnostic feature set and each global diagnostic prediction value, updates each edge diagnostic model using the diagnostic loss function value, extracts diagnostic features from the corresponding historical working data of each reciprocating compressor, and updates the plurality of diagnostic feature sets;

[0021] The global diagnosis model uses multiple diagnostic feature sets to determine a global loss function value, uses the global loss function value to update the global diagnosis model, and updates multiple global diagnosis prediction values ​​respectively.

[0022] 4. A method for diagnosing a reciprocating compressor fault according to item 3, wherein:

[0023] Each edge diagnosis model in the plurality of edge diagnosis models determines a diagnosis loss function value using each corresponding diagnosis feature set and each global diagnosis prediction value, including:

[0024] Each edge diagnosis model in the plurality of edge diagnosis models obtains an edge diagnosis prediction value of the reciprocating compressor using each corresponding diagnosis feature set, and obtains an edge diagnosis loss function value based on the edge fault prediction value;

[0025] Each edge diagnosis model in the plurality of edge diagnosis models obtains an enhanced diagnosis loss function value using each corresponding edge diagnosis prediction value and each global diagnosis prediction value; and

[0026] The diagnosis loss function value is determined based on a weighted sum of the edge diagnosis loss function value and the enhancement diagnosis loss function value.

[0027] 5. The reciprocating compressor fault diagnosis method according to item 3 or 4, further comprising:

[0028] Repeating the steps of determining a diagnostic loss function value, updating a plurality of diagnostic feature sets, determining a global loss function value, and updating a plurality of global diagnostic prediction values, to iterate a plurality of edge diagnostic models and a global diagnostic model.

[0029] 6. The reciprocating compressor fault diagnosis method according to item 5, further comprising:

[0030] Determine whether each edge diagnosis model and / or global diagnosis model after iteration converges;

[0031] In response to each edge diagnosis model and / or the global diagnosis model converging after iteration, the iteration of each edge diagnosis model and / or the global diagnosis model is stopped.

[0032] 7. The reciprocating compressor fault diagnosis method according to item 5, further comprising:

[0033] determining a maximum number of iterations for each edge diagnostic model and / or a global diagnostic model; and

[0034] In response to the number of iterations of each edge diagnostic model and / or the global diagnostic model reaching a maximum number of iteration updates, the iterations of each edge diagnostic model and / or the global diagnostic model are stopped.

[0035] 8. Reciprocating compressor fault diagnosis device, comprising:

[0036] A diagnostic feature extraction unit, which respectively uses a plurality of edge diagnostic models to extract diagnostic features from the working data of a plurality of corresponding reciprocating compressors to form a plurality of diagnostic feature sets;

[0037] A global diagnosis prediction unit, wherein the global diagnosis model obtains a plurality of global diagnosis prediction values ​​based on each of the plurality of diagnosis feature sets;

[0038] an edge model updating unit, which uses a plurality of global diagnosis prediction values ​​to respectively train a plurality of edge diagnosis models, and optimizes the plurality of edge diagnosis models through iterative learning; and

[0039] The fault diagnosis unit inputs the real-time working data of the multiple reciprocating compressors into the corresponding multiple edge diagnosis models after optimization to obtain multiple real-time fault prediction values.

[0040] 9. The reciprocating compressor fault diagnosis device according to item 8, which executes the reciprocating compressor fault diagnosis method as described in any one of items 1-7.

[0041] 10. An electronic device comprising:

[0042] Processor; and

[0043] A memory, in which computer program instructions are stored, and when the computer program instructions are executed by a processor, the processor executes a reciprocating compressor fault diagnosis method as described in any one of items 1-7.

[0044] Beneficial Effects

[0045] 1. A federated learning architecture is used to address the issue of privacy protection of user original data in intelligent fault diagnosis of reciprocating compressors across factories. At the same time, a federated collaborative diagnosis framework enhanced with global diagnostic knowledge is proposed to address the problem of serious heterogeneity among the data failure modes of various participants in the actual operation of reciprocating compressors. This framework makes full use of the diagnostic knowledge learned from different clients to build a global diagnostic knowledge set covering a wide range of fault modes on the server, and uses this knowledge set to train a global diagnostic model that can cope with more complex fault types. The client uses the knowledge learned from the global model to enhance its own diagnostic capability in a distilled manner, which can not only improve the generalization performance of each client's diagnostic model, but also effectively alleviate the impact of the heterogeneity of client fault modes on the diagnostic performance of the federated system.

[0046] 2. Considering the problem of insufficient computing resources on edge clients, a targeted model training mechanism is designed. In this training mechanism, part of the diagnostic model training task is assigned to the server side with more abundant computing resources, while the client can flexibly choose the client model structure that best suits its own conditions. This distributes the computing pressure of model training more reasonably, which not only reduces the computing burden on edge clients, but also makes full use of the more powerful computing power of the server. At the same time, this training mechanism does not transmit all the parameters of the model between the client and the server, but replaces them with the intermediate feature information of the client model and the global diagnostic knowledge inferred from the global model, which further reduces the demand for communication resources of the federated diagnostic system and alleviates the burden of communication bandwidth, thereby speeding up the response speed of the overall federated collaborative diagnosis system. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 The figure illustrates a flow chart of a reciprocating compressor fault diagnosis method according to an embodiment of the present application.

[0048] Figure 2 A flowchart of obtaining multiple global diagnosis prediction values ​​according to a reciprocating compressor fault diagnosis method according to an embodiment of the present application is illustrated.

[0049] Figure 3 A schematic diagram of edge diagnosis model iteration of a reciprocating compressor fault diagnosis method according to an embodiment of the present application is illustrated.

[0050] Figure 4 A block diagram of a reciprocating compressor fault diagnosis device according to an embodiment of the present application is illustrated.

[0051] Figure 5 A block diagram of an electronic device according to an embodiment of the present application is illustrated. DETAILED DESCRIPTION

[0052] The present application is further described below in conjunction with embodiments. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0053] Unless otherwise defined, the technical and scientific terms in this specification have the same meaning as those generally understood by those skilled in the art. Although similar or identical methods and materials as described herein may be used in experiments or practical applications, materials and methods are described herein below. In the event of a conflict, the present specification including the definitions therein shall prevail. In addition, materials, methods and examples are provided for illustration only and are not restrictive. The present application is further described below in conjunction with specific embodiments, but is not intended to limit the scope of the present application.

[0054] Application Overview

[0055] As mentioned above, the problem with the traditional federated learning diagnosis method is that, on the one hand, in the traditional federated learning framework, the heterogeneity of the failure modes between devices will have a great impact on the effect of the global diagnosis model. The reciprocating compressor device client equipped with a small-scale diagnosis model uploads the parameters to the global model and trains the global model through parameter average aggregation. In this way, the global model still cannot learn the complex feature representations in the data of multiple devices, and the prediction results obtained based on the heterogeneous multi-device failure mode features are still a shallow integration result. Therefore, the performance improvement of the small-scale diagnosis model is limited. This federated learning method is difficult to cope with the challenges of fault prediction accuracy brought about by the heterogeneity of failure modes.

[0056] On the other hand, the entire process of training the model in the traditional federated learning method is completed on the reciprocating compressor device client equipped with a small-scale diagnostic model. The small-scale diagnostic model of the reciprocating compressor device client needs to learn more global knowledge of the global model by matching its feature representation with the global model. However, it is usually difficult to configure abundant computing resources at the edge where the reciprocating compressor is located, which causes a mismatch between demand and resources, and the relatively more abundant computing resources on the cloud server side are not fully utilized. Moreover, in traditional federated learning, there are multiple transmissions of model parameters between the edge segment and the server side, which is also a large demand in terms of communication resources, and better communication infrastructure needs to be configured for multiple edge clients.

[0057] Based on the above problems, the basic idea of ​​the present application is to construct a federated collaborative diagnosis method, taking the distributed reciprocating compressors across factories as an edge client, and building a cloud server on it. Information communication can be carried out between each client and the server, but the original private data of the client will not be transmitted and shared. On this basis, a federated collaborative diagnosis system for reciprocating compressors is formed; a client diagnosis model is constructed and global knowledge is distilled locally. The client model is mainly divided into a feature extractor and a classifier in structure. Each client participating in the federated diagnosis method builds the most suitable local diagnosis model for itself according to its own local actual situation. Under this condition, the structure of each local model, especially the feature extractor structure, can be different from each other and each client can make targeted adjustments to it; in the federated diagnosis method, after each client completes the construction of the local model, it uses local private data to train the model. During the training process, the global diagnosis knowledge downloaded from the server is used to enhance the effect of each client's local model. The global diagnosis knowledge enhancement is completed through the knowledge distillation method; after the client completes the local model training, it uploads its corresponding diagnostic features and their corresponding labels to the server;

[0058] Furthermore, the global diagnostic model on the server side is trained. After the cloud server receives the local diagnostic knowledge of each edge client, it first forms a global diagnostic knowledge set that comprehensively covers the failure modes of different clients. The server then uses this knowledge set to train the global diagnostic model. Since the cloud server has better computing resources and needs to face a larger amount of data and more complex failure modes than the edge client, the global diagnostic model on the server can be constructed according to the actual situation. The structure is more complex than the client model to cope with the challenges of such large data volumes and complex failure modes; after the server completes the global model training, it performs a forward reasoning and then distributes the reasoning result labels to the corresponding edge clients. The federated collaborative diagnosis method iteratively trains the model. During this process, each client receives the diagnostic label from the cloud server, and then repeats the iterative training according to the above steps until the entire training process reaches the threshold set in advance and the training of the entire model is fit.

[0059] After introducing the basic principles of the present application, various non-limiting embodiments of the present application will be described in detail with reference to the accompanying drawings.

[0060] Exemplary Methods

[0061] Figure 1 The figure illustrates a flow chart of a reciprocating compressor fault diagnosis method according to an embodiment of the present application.

[0062] like Figure 1 As shown, the reciprocating compressor fault diagnosis method according to an embodiment of the present application includes the following steps.

[0063] Step S110, using multiple edge diagnostic models to extract diagnostic features from the working data of multiple corresponding reciprocating compressors to form multiple diagnostic feature sets. In actual production work, reciprocating compressors are usually distributed in different industrial sites and operate independently. These distributed reciprocating compressors across factories can be used as multiple independent edge clients, and an edge diagnostic model can be deployed on each edge client.

[0064] According to the reciprocating compressor fault diagnosis method of the embodiment of the present application, the edge diagnosis model of each edge client can fully control all the working data of the reciprocating compressor in which it is located, that is, these data are privately used by each edge diagnosis model and are not shared and transmitted between edge clients, thus fully ensuring the privacy security of the local original data of each reciprocating compressor. Each edge diagnosis model can be a pre-trained low-complexity model that extracts a set of diagnostic features from the historical working data generated by the reciprocating compressor in which it is located. The diagnostic features can only preliminarily characterize the fault category of the reciprocating compressor and need to be further explored.

[0065] Next, the diagnostic feature sets extracted by each edge diagnostic model are integrated to form a high-dimensional collection database of diagnostic feature sets. For example, a cloud server can be built, and the edge diagnostic model of each edge client outputs the obtained diagnostic feature set and uploads it to the cloud server to form a database on the cloud server. In this process, the historical working data of the reciprocating compressor processed by the edge diagnostic model on each edge client remains invisible to the cloud server, which further ensures the privacy of the historical working data.

[0066] In particular, the reciprocating compressor fault diagnosis method according to the embodiment of the present application utilizes multiple edge diagnosis models to extract historical working data of multiple corresponding reciprocating compressors, such as thermal performance parameters, vibration data, piston rod displacement data, cylinder pressure data, piston outer dead center / key phase data, noise monitoring data and oil monitoring data, etc., which may have diagnostic features in historical working data with large environmental differences. These features may include thermal performance features, such as intake and exhaust temperature, intake and exhaust pressure, exhaust volume, lubrication oil circuit pressure and cooling water circuit temperature, as well as vibration monitoring and piston rod axis trajectory features, such as crankcase vibration parameters, crosshead vibration parameters and cylinder vibration parameters.

[0067] In one example, after each edge diagnosis model is deployed, the corresponding edge client inputs the locally private historical working data of the reciprocating compressor into the edge diagnosis model. The process can be expressed as:

[0068]

[0069] in, represents the kth edge diagnosis model in the tth global update, represents the training parameters of the kth edge diagnosis model, represents the sample variable of the i-th training sample in the local historical work data, so these private local historical work data can be further represented as a historical work dataset, that is, Where N k represents the total number of samples in the historical working dataset of the kth edge diagnosis model, Representing the sample variables and their corresponding fault labels in the historical working data set, the edge diagnosis model can extract simple diagnostic features. It should be noted that each edge diagnosis model will update its training parameters and update the historical working data set locally in each global update, but it can do so without receiving data from other edge diagnosis models. In this way, each edge diagnosis model can receive and learn the diagnostic labels of the global features of all edge clients, and the process does not generate data exchange between edge clients.

[0070] Edge diagnosis models can be divided into feature extractors according to their functions. and classifier In one example, a feature extractor The output is It is a collection of simple primary fault mode classification features; classifier The output of the edge diagnosis model is It is the logical value output by the fully connected layer of the last layer of the classifier of the edge diagnosis model. The logical value can be converted into the predicted probability of each fault mode by the edge diagnosis model through the following formula:

[0071]

[0072] in, is the predicted probability value of the fault mode of the edge diagnosis model based on multiple sample variables in the input historical working data, and τ is the temperature normalization factor, which is used to smooth the probability value to improve the diagnosis effect. Feature Extractor Different structures and / or complexities can be set according to actual conditions. For example, in an application environment with poor hardware conditions, a low-performance edge diagnosis model can be configured for each edge client of a reciprocating compressor, and its feature extractor can only extract simple low-level features from the original historical data, thereby ensuring that it can run smoothly on the edge client of the reciprocating compressor. In this way, sufficient configuration flexibility can be provided for edge clients under different conditions, reducing the configuration requirements of different reciprocating compressor fault diagnosis terminals to improve the practicality of fault detection.

[0073] That is, each edge diagnostic model can be a classification model with lower complexity, such as a linear binary classification model such as a decision tree or logistic regression, and the historical working data set used to train each edge diagnostic model can also be a small sample data set. Reciprocating compressors usually operate in high temperature, high pressure, flammable and explosive environments, and their working environment may not allow the long-term use of high-configuration computing equipment; the edge diagnostic model provided in this application can be deployed and trained in low-configuration computers, servers or workstations, so it can adapt to the harsh working environment of reciprocating compressors, that is, the edge client configures the device or electronic device with the edge diagnostic model at the working end or data storage end of the reciprocating compressor, and first considers improving the safety performance of the device or electronic device rather than improving the computing performance, so as to meet the long-term and safe operation requirements of the edge diagnostic model.

[0074] In step S120, the global diagnostic model obtains multiple global diagnostic prediction values ​​based on each diagnostic feature set in the multiple diagnostic feature sets. The global diagnostic model is different from the edge diagnostic model. It can have a higher model complexity and can use multiple diagnostic feature sets for global training. Therefore, in step S120, it can further include the following steps: Figure 2 The following steps are shown.

[0075] Step S1201, uploading a local diagnostic feature set. After the local multiple edge diagnostic models are trained, the diagnostic feature set obtained locally and the labels of multiple diagnostic features in the diagnostic feature set can be uploaded to a cloud server. In an example, the diagnostic feature set uploaded by the kth edge diagnostic model to an available cloud server can be represented as The cloud server receives local diagnostic feature sets from at least some edge diagnostic models, preferably all edge diagnostic models, to form a global diagnostic knowledge feature set. In this way, the diagnostic knowledge feature set obtained by the reciprocating compressor fault diagnosis method according to the embodiment of the present application can cover different fault mode information on the reciprocating compressors where different edge diagnostic models are located, and is a reciprocating compressor working data set with a larger data volume, more complex data distribution and more fault types, as shown in the following formula:

[0076]

[0077] Among them, N clientRepresents the number of edge diagnostic models deployed on edge clients based on reciprocating compressors. Those skilled in the art will appreciate that, although the edge diagnostic model has low complexity and the computing power required for feature calculation and fault mode output can be very low, in actual production work, not all edge clients based on reciprocating compressors can deploy the model. Therefore, the edge diagnostic model can also be deployed on the console work end or server of one or more reciprocating compressors in certain scenarios to effectively obtain at least part of the historical working data of the reciprocating compressor in the scenario to ensure the information richness of the diagnostic knowledge feature set.

[0078] Step S1202, train the global diagnostic model using the set of diagnostic knowledge feature sets to obtain a global diagnostic model with better global diagnostic capabilities. Specifically, the global diagnostic model can be deployed on a server that stores the diagnostic knowledge feature set or can access the cloud server where the diagnostic knowledge feature set is located, so that the global diagnostic model can be flexibly trained according to the feature dimension of the knowledge feature set, the fault mode category and the complexity of the data distribution; the global diagnostic model obtained in this way can further obtain the fault mode prediction value of the reciprocating compressor where each edge diagnostic model is located based on the diagnostic feature set from each edge diagnostic model. Compared with the fault mode prediction value obtained by the edge diagnostic model based on the local example working data, the global diagnostic model focuses on the feature representation of more extensive and complex reciprocating compressor working faults. Therefore, although they are all based on the same diagnostic feature set input, the global diagnostic model can judge the fault heterogeneity that the edge diagnostic model cannot obtain, that is, the fault mode that has not yet appeared on the reciprocating compressor where the edge diagnostic model is located, the fault mode that occurs less frequently, or the complex fault mode that is difficult to learn due to the complexity limitation of the edge diagnostic model.

[0079] In one example, the global diagnostic model can learn multiple diagnostic feature sets that cover the generalized fault characteristics of all reciprocating compressors. After training, the global diagnostic model uses the existing diagnostic knowledge feature set to perform a forward inference on the server side, inputs each diagnostic feature set into the model separately, and obtains the corresponding global diagnostic prediction value. The server side then distributes the global diagnostic prediction value as a fault diagnosis result label to each edge diagnostic model for enhanced training of the edge diagnostic model.

[0080] Specifically, according to the diagnostic knowledge feature set obtained in step S1201 Train a global diagnostic model F with better performance GSince the diagnostic knowledge feature set includes both the fault mode feature knowledge output by multiple edge diagnostic models, i.e., the fault mode probability prediction value, and the label information of these feature knowledge, the training of the global diagnostic model can use the classification model training method known to those skilled in the art to achieve the global diagnostic model in the diagnostic knowledge feature set. For example, the global diagnostic model can be a linear model, a neural network, or a variety of pre-trained models, etc., and the model is fitted by setting a loss function and an appropriate learning rate. Training data and parameter adjustment can be performed based on selective fault mode prediction, which provides the feasibility and practicality of system construction and deployment application in different application scenarios for the fault diagnosis method of the reciprocating compressor according to the embodiment of the present application.

[0081] Step S1203, after the global diagnosis model training is completed, a forward reasoning is performed based on each diagnosis feature set in the existing diagnosis knowledge feature set to obtain the global diagnosis prediction values ​​of multiple reciprocating compressors. The reasoning process can be expressed as follows:

[0082]

[0083] in, That is, the global diagnostic prediction value learned by the global diagnostic model based on the diagnostic feature set provided by the kth edge diagnostic model; the global diagnostic model sends the global diagnostic prediction value to the edge client where the corresponding reciprocating compressor is located on the server side. The global diagnostic prediction value contains rich global fault diagnosis knowledge, which can improve the global diagnostic performance of the edge diagnostic model in each edge client during the enhanced training process, so as to effectively avoid the situation where the existing federated learning method suffers from the degradation of detection performance due to fault heterogeneity when systematically detecting multiple reciprocating compressors.

[0084] Furthermore, the fault diagnosis method for a reciprocating compressor according to an embodiment of the present application includes step S130, i.e., using multiple global diagnostic prediction values ​​to respectively train multiple edge diagnostic models, and optimizing multiple edge diagnostic models and the global diagnostic model through iterative learning. The purpose of this is that the edge diagnostic model can locally distill global knowledge using the global diagnostic prediction values ​​from the global diagnostic model, and the global diagnostic model can also be updated together with the edge diagnostic model during each round of iteration. As can be seen from step S110, an edge diagnostic model that best suits the actual situation of the equipment itself can be deployed on each reciprocating compressor, and preliminary training can be performed using the historical working data of the equipment; after receiving the global diagnostic prediction value, the edge diagnostic model can be collaboratively trained in combination with the local historical working data to achieve enhanced learning and improve the performance of the edge diagnostic model.

[0085] In this way, each edge diagnosis model uses the global fault diagnosis knowledge from the global diagnosis model in a knowledge distillation manner to enhance its own diagnostic performance and generalization ability. It can be understood by those skilled in the art that the diagnosis results of the complex model with better performance are efficiently transmitted to the small model with a simple structure for enhanced training, without the need to transmit a large amount of complex feature data or model parameters. In this way, the fault diagnosis method for a reciprocating compressor according to an embodiment of the present application avoids the need to use more data or parameters in the optimization process of the edge diagnosis model, thereby avoiding resource consumption and transmission costs during the optimization of the edge diagnosis model, making it possible to deploy it on multiple reciprocating compressors.

[0086] Specifically, the enhanced training of the edge diagnosis model can be divided into two parts, and the two parts use their own feature variable sets or diagnostic prediction values ​​to collaboratively optimize the edge diagnosis model parameters. In the first part, the edge diagnosis model continues to use its own diagnostic feature set extracted from historical working data or directly trains through historical working data. This process is consistent with the general classification model training; in one example, the cross entropy loss function is used to evaluate the model performance of each local update. In the other part, the edge diagnosis model receives the global diagnosis prediction value issued by the global diagnosis model as the global diagnosis label, which contains heterogeneous fault feature information that has not appeared in the current reciprocating compressor or is difficult to be extracted by the undistilled edge diagnosis model. The edge diagnosis model enhances its comprehensive diagnostic performance through knowledge distillation; in one example, the distillation loss function of distilled global knowledge is used to evaluate the model performance of each global update. Through the above two parts, the optimization process of the edge diagnosis model is constrained at the same time, so that it can not only fully learn its own fault mode, but also have its unseen faults to enhance its generalization performance, so as to better adapt to the heterogeneous challenges of systematic detection.

[0087] In the reciprocating compressor fault diagnosis method according to the embodiment of the present application, the loss function of the edge diagnosis model enhanced training is therefore composed of two parts, namely, the edge diagnosis loss function value of the edge diagnosis prediction value output by the edge diagnosis model for local update, exemplified by a cross entropy loss function, and the enhanced diagnosis loss function value for distilling the global diagnosis prediction value and enhancing the local model diagnosis effect through global update, exemplified by a KL divergence function. Therefore, the overall diagnostic loss function of the model can be expressed by the following formula:

[0088]

[0089] in, and are the cross entropy loss value and KL divergence value respectively, and the parameter λ is used to adjust the weight between the two loss functions. The cross entropy loss value and KL divergence value can be expressed as the following formula:

[0090]

[0091] Among them, C k represents the total number of fault modes that can be output by the kth edge diagnosis model, represents the true label of the sample variable of the i-th training sample in the local historical working data, is the predicted value of the global diagnostic model deployed on the server for the current sample variable, and KL(·) represents the KL divergence. By optimizing the edge diagnostic model deployed on the reciprocating compressor through the above loss function, the knowledge in the global diagnostic model deployed in the cloud is transferred to each locally deployed edge diagnostic model, and the performance of the local edge diagnostic model is enhanced by using the global knowledge. It should be noted that when the edge diagnostic model is trained for the first time using the diagnostic feature set, the global diagnostic model has not yet output the global fault diagnosis knowledge. At this time, the global fault diagnosis knowledge can be initialized to 0, and the KL divergence value is not included in the diagnostic loss function value.

[0092] In this way, Figure 3 As shown, the edge diagnosis model can be updated using the diagnosis loss function value, for example, the model parameters are updated by the gradient descent method, and the diagnosis loss function value can be obtained through the labels of the training data, which reduces the communication cost between the global diagnosis model deployed in the cloud and the edge diagnosis model deployed in the local edge client. In each global update communication round, only the label prediction value of each sample variable in the historical working data set can be transmitted between the models. The edge prediction model distills global knowledge and improves its own classification performance through the label prediction value. This can also reduce the communication frequency and time between the device or electronic device where the edge diagnosis model is located and the cloud, effectively improving its high deployment cost, data transmission delay and / or packet loss in the working environment of the reciprocating compressor, so as to improve the practical application feasibility of the diagnosis method of the present application.

[0093] In particular, the global diagnostic model can also select the cross entropy loss function value for each iteration update in the cloud server. In addition, different loss functions can also be selected according to the different classifiers of the global diagnostic model, such as logarithmic loss function, exponential loss function, hinge loss function, etc., which are not limited in this application. The present application may provide that it is determined whether each edge diagnostic model and / or each global diagnostic model has converged after each iteration. In response to each edge diagnostic model and / or global diagnostic model after the iteration having converged, the iteration of each edge diagnostic model and / or global diagnostic model can be stopped and the training can be completed; in response to each edge diagnostic model and / or global diagnostic model after the iteration not converging, the iteration of each edge diagnostic model and / or global diagnostic model can be continued. In addition, a preset maximum number of iteration updates can be set. When the number of iterations reaches the maximum number of iteration updates, the iteration is stopped directly to complete the training; higher model training efficiency can be provided when facing a high-dimensional diagnostic knowledge feature set.

[0094] Step S140, the real-time working data of multiple reciprocating compressors are respectively input into the corresponding optimized multiple edge diagnosis models to obtain multiple real-time fault prediction values. It can be understood that the real-time fault prediction value can be the probability value of multiple types of fault modes of the corresponding reciprocating compressor output by each edge prediction model, or one or more types of fault modes. The classifier of the edge diagnosis model can be a simple binary classifier such as logistic regression, or a more complex multi-classifier, which can be determined according to the working environment of the edge client in which it is located. For example, the working environment of the reciprocating compressor is relatively harsh, and the edge client configured with it cannot carry more complex devices or electronic devices. The corresponding edge diagnosis model can be a simple classification model, and an output port is set separately for each fault mode or a specific type of fault mode to reduce its training cost, communication cost and data output complexity.

[0095] Exemplary Devices

[0096] Figure 4 A block diagram of a reciprocating compressor fault diagnosis device according to an embodiment of the present application is illustrated.

[0097] like Figure 4As shown, the reciprocating compressor fault diagnosis device 200 according to the embodiment of the present application includes: a diagnostic feature extraction unit 210, which uses multiple edge diagnostic models to extract diagnostic features from the historical working data of multiple corresponding reciprocating compressors to form multiple diagnostic feature sets; a global diagnostic prediction unit 220, the global diagnostic model obtains multiple global diagnostic prediction values ​​based on each diagnostic feature set in the multiple diagnostic feature sets; an edge model updating unit 230, which uses multiple global diagnostic prediction values ​​to train multiple edge diagnostic models respectively, and optimizes multiple edge diagnostic models and the global diagnostic model through iterative learning; and a fault diagnosis unit 240, which inputs the real-time working data of multiple reciprocating compressors into the corresponding multiple edge diagnostic models after optimization to obtain multiple real-time fault prediction values.

[0098] Here, those skilled in the art will appreciate that the specific functions and operations of the various units and modules in the reciprocating compressor fault diagnosis device 200 have been described in detail above. Figure 1 to Figure 2 The description of the reciprocating compressor fault diagnosis method has been described in detail, and therefore, its repeated description will be omitted.

[0099] As described above, the reciprocating compressor fault diagnosis device 200 according to the embodiment of the present application can be implemented in various terminal devices, such as a server for training a global diagnostic model and an electronic device for training an edge diagnostic model. In one example, the reciprocating compressor fault diagnosis device 200 according to the embodiment of the present application can be integrated into the terminal device as a software module and / or a hardware module. For example, the reciprocating compressor fault diagnosis device 200 can be a software module in the operating system of the terminal device, or can be an application developed for the terminal device; of course, the reciprocating compressor fault diagnosis device 200 can also be one of the many hardware modules of the terminal device.

[0100] Alternatively, in another example, the reciprocating compressor fault diagnosis apparatus 200 and the terminal device may also be separate devices, and the reciprocating compressor fault diagnosis apparatus 200 may be connected to the terminal device via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.

[0101] Exemplary Electronic Devices

[0102] Below, reference Figure 5 To describe an electronic device according to an embodiment of the present application.

[0103] Figure 5 A block diagram of an electronic device according to an embodiment of the present application is illustrated.

[0104] like Figure 5As shown, the electronic device 10 includes one or more processors 11 and a memory 12 .

[0105] The processor 13 may be a central processing unit (CPU) or other forms of processing units having data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device 10 to perform desired functions.

[0106] The memory 12 may include one or more computer program products, and the computer program product may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory (cache), etc. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 11 may run the program instructions to implement the reciprocating compressor fault diagnosis method of each embodiment of the present application described above and / or other desired functions. Various contents such as diagnostic feature sets, diagnostic loss function values, etc. may also be stored in the computer-readable storage medium.

[0107] In one example, the electronic device 10 may further include: an input device 13 and an output device 14, and these components are interconnected via a bus system and / or other forms of connection mechanisms (not shown).

[0108] The input device 13 may include, for example, a keyboard, a mouse, etc.

[0109] The output device 14 can output various information to the outside, including trained edge diagnosis models and / or global diagnosis models, etc. The output device 14 can include, for example, a display, a speaker, a printer, a communication network and a remote output device connected thereto, etc.

[0110] Of course, for simplicity, only some of the components in the electronic device 10 related to the present application are shown, omitting components such as a bus, an input / output interface, etc. In addition, the electronic device 10 may further include any other appropriate components according to specific application scenarios.

[0111] The basic principles of the present application are described above in conjunction with specific embodiments. However, it should be noted that the advantages, strengths, effects, etc. mentioned in the present application are only examples and not limitations, and it cannot be considered that these advantages, strengths, effects, etc. are required by each embodiment of the present application. In addition, the specific details disclosed above are only for the purpose of illustration and ease of understanding, not for limitation, and the above details do not limit the present application to being implemented by adopting the above specific details.

[0112] In this application, words such as "including", "comprising", "having", etc. are open words, meaning "including but not limited to", and can be used interchangeably therewith. The words "or" and "and" used herein mean the words "and / or" and can be used interchangeably therewith, unless the context clearly indicates otherwise. The word "such as" used herein means the phrase "such as but not limited to", and can be used interchangeably therewith.

[0113] It should also be noted that in the method, system and device of the present application, each step or each module can be decomposed and / or recombined. Such decomposition and / or recombination should be regarded as equivalent solutions of the present application.

[0114] The above description has been given for the purpose of illustration and description. In addition, this description is not intended to limit the embodiments of the present application to the forms disclosed herein. Although multiple example aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, changes, additions and sub-combinations thereof.

Claims

1. Reciprocating compressor fault diagnosis method, including: Extracting diagnostic features from historical working data of a plurality of corresponding reciprocating compressors using a plurality of edge diagnostic models respectively to form a plurality of diagnostic feature sets; The global diagnostic model obtains a plurality of global diagnostic prediction values ​​based on each diagnostic feature set in the plurality of diagnostic feature sets; Using the multiple global diagnosis prediction values ​​to train the multiple edge diagnosis models respectively, and optimizing the multiple edge diagnosis models and the global diagnosis model through iterative learning; as well as The real-time working data of a plurality of reciprocating compressors are respectively input into the correspondingly optimized plurality of edge diagnosis models to obtain a plurality of real-time fault prediction values.

2. The reciprocating compressor fault diagnosis method according to claim 1, wherein: The diagnostic features extracted from the historical working data of the corresponding reciprocating compressors using the plurality of edge diagnostic models respectively include: thermal performance features, including intake and exhaust temperatures, intake and exhaust pressures, exhaust volume, lubricating oil circuit pressure, and cooling water circuit temperature; Vibration monitoring and piston rod axis trajectory characteristics, including crankcase vibration parameters, crosshead vibration parameters and cylinder vibration parameters; The historical working data include thermal performance parameters, vibration data, piston rod displacement data, cylinder pressure data, piston outer dead center / key phase data, noise monitoring data and oil monitoring data.

3. The reciprocating compressor fault diagnosis method according to claim 1, wherein: The using the multiple global diagnosis prediction values ​​to respectively train the multiple edge diagnosis models, and optimizing the multiple edge diagnosis models and the global diagnosis model through iterative learning comprises: Outputting each of the plurality of global diagnosis prediction values ​​to the corresponding plurality of edge diagnosis models respectively; Each edge diagnostic model in the plurality of edge diagnostic models determines a diagnostic loss function value using each corresponding diagnostic feature set and each global diagnostic prediction value, updates each edge diagnostic model using the diagnostic loss function value, extracts diagnostic features from the corresponding historical operating data of each reciprocating compressor, and updates the plurality of diagnostic feature sets; The global diagnostic model determines a global loss function value using the multiple diagnostic feature sets, updates the global diagnostic model using the global loss function value, and updates the multiple global diagnostic prediction values ​​respectively.

4. The reciprocating compressor fault diagnosis method according to claim 3, wherein: Each edge diagnosis model in the plurality of edge diagnosis models determines the diagnosis loss function value by using each corresponding diagnosis feature set and each global diagnosis prediction value, including: Each edge diagnosis model in the plurality of edge diagnosis models obtains an edge diagnosis prediction value of the reciprocating compressor using each corresponding diagnostic feature set, and obtains an edge diagnosis loss function value based on the edge fault prediction value; Each edge diagnosis model of the plurality of edge diagnosis models obtains an enhanced diagnosis loss function value using each corresponding edge diagnosis prediction value and each global diagnosis prediction value; and The diagnosis loss function value is determined based on a weighted sum of the edge diagnosis loss function value and the enhancement diagnosis loss function value.

5. The reciprocating compressor fault diagnosis method according to claim 3 or 4, further comprising: The determining of the diagnostic loss function value, the updating of the plurality of diagnostic feature sets, the determining of the global loss function value, and the updating of the plurality of global diagnostic prediction values ​​are repeated to iterate the plurality of edge diagnostic models and the global diagnostic model.

6. The reciprocating compressor fault diagnosis method according to claim 5, further comprising: Determine whether each edge diagnosis model and / or the global diagnosis model after iteration converges; In response to each edge diagnosis model and / or the global diagnosis model converging after the iteration, the iteration of each edge diagnosis model and / or the global diagnosis model is stopped.

7. The reciprocating compressor fault diagnosis method according to claim 5, further comprising: Determining a maximum number of iterations of each edge diagnosis model and / or the global diagnosis model; as well as In response to the number of iterations of each edge diagnostic model and / or the global diagnostic model reaching the maximum number of iteration updates, the iterations of each edge diagnostic model and / or the global diagnostic model are stopped.

8. Reciprocating compressor fault diagnosis device, comprising: A diagnostic feature extraction unit, which respectively uses a plurality of edge diagnostic models to extract diagnostic features from the working data of a plurality of corresponding reciprocating compressors to form a plurality of diagnostic feature sets; A global diagnosis prediction unit, wherein the global diagnosis model obtains a plurality of global diagnosis prediction values ​​based on each of the plurality of diagnosis feature sets; An edge model updating unit, which uses the multiple global diagnosis prediction values ​​to respectively train the multiple edge diagnosis models, and optimizes the multiple edge diagnosis models through iterative learning; as well as The fault diagnosis unit inputs the real-time working data of the multiple reciprocating compressors into the corresponding multiple edge diagnosis models after optimization to obtain multiple real-time fault prediction values.

9. The reciprocating compressor fault diagnosis device according to claim 8, which executes the reciprocating compressor fault diagnosis method according to any one of claims 1-7.

10. An electronic device, comprising: processor; as well as A memory, in which computer program instructions are stored, and when the computer program instructions are executed by the processor, the processor executes the reciprocating compressor fault diagnosis method according to any one of claims 1 to 7.

Images