A distributed transfer learning machine pump fault diagnosis interactive method, system and device

Abstract

The present invention relates to the field of fault diagnosis and intelligent manufacturing technology, and in particular to a distributed transfer learning machine pump fault diagnosis interactive method, system and device. Its technical solution The system includes a sensor array, a distributed slave station module, a communication gateway, and a parameter master station server; the sensor array includes a plurality of types of sensors installed according to the site environment and actual needs, and is equipped with a power supply module; the distributed slave station module integrates a multi-channel acquisition module, with the RK3588 development motherboard as the main body, and multiple NPU neural computing units are externally mounted, including a custom deep migration network model. The present invention has good generalization performance, matches the detection equipment, optimizes the data transmission flow problem, and improves the model training efficiency. At the same time, it avoids the external leakage of the original equipment operation data of the enterprise, and realizes data privacy security. At the same time, the parameters are encrypted to further ensure the possibility of reverse cracking of the parameter data.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis and intelligent manufacturing, and in particular to a distributed transfer learning machine pump fault diagnosis interactive method, system and device. Background Art

[0002] At present, rotating equipment, including various motors and pumps, mostly uses traditional modeling methods for fault diagnosis, which requires a lot of experience in fault diagnosis. In addition, the equipment is widely distributed, and a large amount of raw data cannot be uploaded to a unified server for synchronous training due to data privacy, resulting in poor data utilization. At the same time, with the advent of the big data era, the scale of data continues to increase, and the traditional centralized transfer learning method faces the problems of insufficient computing resources and high communication overhead. Therefore, we propose a distributed transfer learning interactive method, system and device for machine pump fault diagnosis. Summary of the invention

[0003] The purpose of the present invention is to propose a distributed transfer learning pump fault diagnosis interactive method, system and device to address the problems of insufficient computing resources and high communication overhead faced by traditional centralized transfer learning methods in the background technology.

[0004] In a first aspect, the present application provides an interactive system for pump fault diagnosis based on distributed transfer learning, the system comprising a sensor array, a distributed slave module, a communication gateway, and a parameter master server;

[0005] The sensor array includes various types of sensors installed according to the on-site environment and actual needs, and is equipped with a power supply module;

[0006] The distributed slave module integrates a multi-channel acquisition module, with the RK3588 development motherboard as the main body, and multiple NPU neural computing units are plugged in, including a custom deep migration network model, which extracts general and device differentiated data features through a two-stage neural network, and the gradient parameters are encrypted and uploaded to the communication gateway;

[0007] The communication gateway receives the encrypted gradient parameters uploaded by the distributed slave station module, packages and uploads them to the parameter master station server, and transmits the encrypted average gradient parameters issued by the parameter master station server to the distributed slave station module at the same time;

[0008] The parameter master server receives the encrypted gradient parameters uploaded by the communication gateway, decrypts and performs weighted calculation to obtain the average gradient, and then encrypts and sends it to each distributed slave module through the communication gateway.

[0009] Optionally, the multiple types of sensors include any one or more of a three-axis vibration accelerometer sensor, a three-axis vibration velocity sensor, a three-phase current transformer, a three-phase voltage transformer, an acoustic emission sensor, a thermal imaging sensor or a camera.

[0010] Optionally, the distributed slave station module also includes a model parameter storage unit and a communication module that supports multiple wireless communication modes, and the model parameter storage unit is used to store the gradient parameters trained by the distributed slave station module and the updated gradient parameters from the parameter master station server.

[0011] Optionally, encryption includes two parts: a public key pk and a private key sk, wherein the encryption process is to input a ciphertext for a plaintext polynomial m∈R:

[0012] c←r·pk+(m+e0,e1)mod Q

[0013] Among them, c is the ciphertext, r is the size of the plaintext space, pk is the public key, m is the plaintext, e0, e1 is the noise, and Q is the modulus;

[0014] The decryption process is to Calculate with private key sk:

[0015]

[0016] Among them, m is the plaintext, c0, c1 is the ciphertext, sk i is a single private key, e i is the error term;

[0017] And return the plaintext;

[0018] Gradient parameters δ uploaded by different distributed slave modules i and the amount of data k i Determine the proportion of weighted gradients, N is the total number of slaves, and the weighted formula is as follows:

[0019]

[0020] The obtained weighted average gradient is encrypted and then transmitted to the distributed slave module through the communication gateway for gradient update.

[0021] Optionally, the training process of the custom deep migration network model includes:

[0022] Slice preprocessing of multiple source domain fault data built into the distributed slave module and the collected target domain data of the detection device to obtain the original waveform of the data;

[0023] After variational mode decomposition VMD, the time-frequency feature map of the data is obtained;

[0024] The time-frequency feature map is input into the backbone network ResNet50 for training to obtain the universal feature expression of the data;

[0025] The common features are input into multiple sub-networks and trained independently with the target domain data to obtain the differential feature expression of the data. The sub-network adopts a multi-level CNN network and adds a DANet dual attention mechanism internally to adaptively integrate local features and global dependencies. The training effect of the custom deep migration network model is evaluated by calculating the overall loss.

[0026] In each round of iteration, the gradient parameters are fully encrypted and uploaded to the parameter master server for weighted processing. Then, the parameter master server sends the updated encrypted gradient, decrypts it, and then updates the entire network in reverse.

[0027] Optionally, the overall loss includes: classification loss, migration loss and distance loss, and the function calculation formula of the overall loss is:

[0028] L total =L CLS +αL TL +βL DISC ;

[0029] in,

[0030] The function calculation formula for classification loss is:

[0031]

[0032] in, is the source domain data, is the source domain data label, X sj is the source domain dataset, N is the total number of datasets, C is the Softmax classification prediction operation, F is the feature extraction operation of the secondary sub-network, and G is the general feature extraction operation of the primary main network;

[0033] The function calculation formula of migration loss is:

[0034]

[0035] in, is the Frobenius norm, F j is the feature extraction operation of the secondary sub-network j, and G is the general feature extraction operation of the primary main network. is the source domain j sample, and its number is N sj , is the target domain sample, and its number is N t , the total number of source domain data sets is N;

[0036] The function calculation formula for distance loss is:

[0037]

[0038] Among them, C j The operation is the Softmax classification prediction operation, F i is the feature extraction operation of the secondary sub-network j, and G is the general feature extraction operation of the primary main network. is the target domain sample, and its number is N t , the total number of source domain datasets is N.

[0039] Optionally, the function contribution to the overall loss is matched in a weighted manner, and a hyperparameter search is performed on the parameters in the formula to obtain the optimal parameter set.

[0040] In a second aspect, the present application provides an interactive method for pump fault diagnosis based on distributed transfer learning, comprising the following steps:

[0041] Install sensor arrays on motors, pumps, fans, and reducer equipment in multiple plants;

[0042] The distributed slave module collects data from the sensor array and performs model training. The training process is synchronized regularly. After each round of training, the gradient parameters are encrypted by the encryption and decryption device, and the encrypted gradients are sent to the parameter master server regularly.

[0043] After the parameter master server obtains a round of encrypted gradients from all slave stations, it performs weighted calculation on the decrypted gradients to obtain the average gradient of this round of calculation, and then encrypts the average gradient and sends it to each distributed slave station module to update the gradient.

[0044] In a third aspect, the present application provides a machine pump fault diagnosis interactive device based on distributed transfer learning, including the machine pump fault diagnosis interactive system based on distributed transfer learning described above.

[0045] Compared with the prior art, this application has the following beneficial technical effects:

[0046] The present invention performs model training based on a transfer learning network, so that the model has good generalization performance and matches the detection equipment.

[0047] Based on the idea of ​​distributed training, the data transmission traffic problem is optimized and the model training efficiency is improved. At the same time, the leakage of the original equipment operation data of the enterprise is avoided, and data privacy security is achieved.

[0048] By encrypting the parameters, the possibility of reverse cracking of the parameter data is ensured. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 Provide a principle block diagram of the distributed slave station module of the present invention;

[0050] Figure 2This is a schematic diagram of the distributed migration fault diagnosis system framework;

[0051] Figure 3 Diagram of the improved transfer learning network structure. DETAILED DESCRIPTION

[0052] The technical scheme of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. The components of the embodiments of the present invention generally described and shown in the accompanying drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention.

[0053] Example 1

[0054] See also Figure 1 ,An interactive system for fault diagnosis of pump based on distributed transfer learning,the system mainly consists of four parts: sensor array, distributed slave module, communication gateway and parameter master server.

[0055] Among them, the sensor array mainly includes various types of sensor equipment, including three-axis vibration accelerometer sensor, three-axis vibration velocity sensor, three-phase current transformer, three-phase voltage transformer, acoustic emission sensor, thermal imaging sensor and camera. According to the site environment and actual needs, the corresponding sensors are installed. At the same time, the corresponding power supply module is required to power the sensor.

[0056] In addition, the distributed slave module integrates a multi-channel acquisition module, which can collect data from the sensor array and perform signal conditioning through the built-in signal conditioning board. The main body of the distributed slave module is the RK3588 development motherboard, and multiple NPU neural computing units are plugged in, which can realize the instruction operation and model training tasks at the operating system level. The custom deep migration network model is an improved migration network model. By designing a two-stage neural network, it relies on the first-level main network to extract common features, and relies on multiple second-level sub-networks to extract differentiated data features of each device, so as to achieve data feature distribution matching of different devices, and finally realize the improvement of model generalization ability. The model parameter storage unit is mainly used to store the gradient parameters of the distributed slave module training and the updated gradient parameters from the parameter master server. The communication module supports a variety of communication modes, such as Lora, Zigbee, NB-IOT, 4G, and WIFI wireless communication modes. After the gradient parameters are encrypted by the full-state encryptor, they are uploaded to the communication gateway through the built-in communication module, and then uploaded to the parameter master server after packaging. The full-state decryptor decrypts the encrypted gradient parameters transmitted by the communication gateway and stores them in the model parameter storage unit.

[0057] See also Figure 2 , multiple factories contain multiple motors, pumps, fans and reducers. Sensor arrays such as three-axis accelerometers, vibration velocity sensors, microphone arrays or cameras are installed according to needs. The sensor array data is collected through the distributed slave module and the model is trained. The training process needs to be synchronized regularly. After each round of training is completed, the gradient parameters will be encrypted by the encryption and decryption device, and the encrypted gradient will be sent to the parameter master server regularly. After the parameter master obtains a round of encrypted gradients from all slaves, it will perform weighted calculations on the decrypted gradients to obtain the average gradient of the round calculation, and encrypt the average gradient and send it to each distributed slave module to update the gradient.

[0058] The encryption consists of two parts: the public key pk and the private key sk. The encryption process is to take a plaintext polynomial m∈R, e0, e1 as noise, and input a ciphertext

[0059] c←r·pk+(m+e0,e1)modQ

[0060] Among them, m is the plaintext, c0, c1 is the ciphertext, sk i is a single private key, e i is the error term.

[0061] The decryption process is to and the private key sk, calculate

[0062] m=c0+∑(c1·sk i +e i )

[0063] Among them, m is the plaintext, c0, c1 is the ciphertext, sk i is a single private key, e i is the error term.

[0064] And returns the plaintext.

[0065] Gradient parameters δ uploaded by different distributed slave modules i and the amount of data k i Determines the proportion of weighted gradients, N is the total number of slaves, and the specific weighted formula is as follows:

[0066]

[0067] The obtained weighted average gradient is encrypted and then transmitted to the distributed slave module through the communication gateway for gradient update.

[0068] See also Figure 3, as the core of the distributed transfer learning network, the model structure determines the quality of the generalization performance. This application designs a machine pump fault diagnosis migration network based on a two-level neural network. The original waveform of the data is obtained by slicing and preprocessing the multiple source domain fault data built into the distributed slave module and the collected target domain data of the detection equipment. After variational mode decomposition VMD, the time-frequency feature map of the data is obtained. The time-frequency feature map is input into the backbone network ResNet50 network for training to obtain the common feature expression of the data; the common features are input into multiple sub-networks and independently trained with the target domain data to obtain the differential feature expression of the data. Among them, the sub-network adopts a multi-level CNN network, and the DANet dual attention mechanism is added internally to adaptively integrate local features and global dependencies. In each round of iteration, the gradient parameters are fully encrypted, uploaded to the parameter master server for weighted processing, and then wait for the parameter master server to send the updated encrypted gradient. After decryption, the entire network is updated in reverse. Its specific mathematical expression is as follows:

[0069] Assume that the multi-source domain dataset is The total number of source domain datasets is N;

[0070] Source domain training dataset: Source domain training dataset label:

[0071] Target domain training dataset:

[0072] Randomly select m source domain samples and m target domain samples, and

[0073] After performing variational mode decomposition, the corresponding time-frequency diagram is obtained.

[0074] Input the time-frequency graph into the primary backbone network for training to obtain the universal feature expression and

[0075] Input the common feature expression into different sub-networks for training to obtain the differential feature expression and Calculate multiple loss functions to get the overall loss function.

[0076] Among them, this method designs a special loss function evaluation method and defines

[0077] Classification loss function

[0078]

[0079] in, is the source domain data, is the source domain data label, X sj is the source domain dataset, N is the total number of datasets, C is the Softmax classification prediction operation, F is the feature extraction operation of the secondary sub-network, and G is the general feature extraction operation of the primary main network;

[0080] Transfer loss function

[0081]

[0082] in, is the Frobenius norm. j Operation j is the feature extraction operation of the secondary sub-network, and operation G is the general feature extraction operation of the primary main network. is the source domain j sample, and its number is N sj . is the target domain sample, and its number is N t , the total number of source domain data sets is N;

[0083] Distance loss function

[0084]

[0085] Among them, C j The operation is the Softmax classification prediction operation, F i Operation j is the feature extraction operation of the secondary sub-network, and operation G is the general feature extraction operation of the primary main network. is the target domain sample, and its number is N t , the total number of source domain datasets is N.

[0086] And the loss function contribution is matched by weighted method to obtain the overall loss function formula

[0087] L total =L CLS +αL TL +βL DISC

[0088] The parameters in the formula are subjected to hyperparameter search to obtain the optimal parameter group.

[0089] The present invention performs model training based on the transfer learning network, so that it has good generalization performance and matches the detection equipment. Based on the distributed training idea, the data transmission flow problem is optimized and the model training efficiency is improved. At the same time, the external leakage of the original equipment operation data of the enterprise is avoided, and data privacy security is achieved. At the same time, the parameters are encrypted to further ensure the possibility of reverse cracking of the parameter data.

[0090] The present application provides an interactive method for pump fault diagnosis based on distributed transfer learning, comprising the following steps:

[0091] Install sensor arrays such as triaxial accelerometers, vibration velocity sensors, microphone arrays or cameras on motors, pumps, fans and reducer equipment in multiple factories as needed;

[0092] The sensor array data is collected through the distributed slave module, and the model training is carried out. The training process needs to be synchronized regularly. After each round of training is completed, the gradient parameters will be encrypted through the encryption and decryption device, and the encrypted gradient will be sent to the parameter master server regularly.

[0093] After the parameter master obtains a round of encrypted gradients from all slave stations, it performs weighted calculation on the decrypted gradients to obtain the average gradient of this round of calculation, and then encrypts the average gradient and sends it to each distributed slave station module to update the gradient.

[0094] The above specific embodiments are only several optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant inspirations of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A pump fault diagnosis interactive system based on distributed transfer learning, characterized in that: The system includes a sensor array, a distributed slave module, a communication gateway, and a parameter master server; The sensor array includes various types of sensors installed according to the on-site environment and actual needs, and is equipped with a power supply module; The distributed slave module integrates a multi-channel acquisition module, with the RK3588 development motherboard as the main body, and multiple NPU neural computing units are plugged in, including a custom deep migration network model, which extracts general and device differentiated data features through a two-stage neural network, and the gradient parameters are encrypted and uploaded to the communication gateway; The communication gateway receives the encrypted gradient parameters uploaded by the distributed slave station module, packages and uploads them to the parameter master station server, and transmits the encrypted average gradient parameters issued by the parameter master station server to the distributed slave station module at the same time; The parameter master server receives the encrypted gradient parameters uploaded by the communication gateway, decrypts and performs weighted calculation to obtain the average gradient, and then encrypts and sends it to each distributed slave module through the communication gateway; The training process of the custom deep migration network model includes: Slice preprocessing of multiple source domain fault data built into the distributed slave module and the collected target domain data of the detection device to obtain the original waveform of the data; After variational mode decomposition VMD, the time-frequency feature map of the data is obtained; The time-frequency feature map is input into the backbone network ResNet50 for training to obtain the universal feature expression of the data; The common features are input into multiple sub-networks and trained independently with the target domain data to obtain the differential feature expression of the data. The sub-network adopts a multi-level CNN network and adds a DANet dual attention mechanism internally to adaptively integrate local features and global dependencies. The training effect of the custom deep migration network model is evaluated by calculating the overall loss. In each iteration, the gradient parameters are fully encrypted and uploaded to the parameter master server for weighted processing. After that, the updated encrypted gradient is sent by the parameter master server, decrypted, and the entire network is updated in reverse. The overall loss includes classification loss, migration loss and distance loss, and the overall loss function calculation formula is: L total =L CLS +αL TL +βL DISC ; in, The function calculation formula for classification loss is: in, is the source domain data, is the source domain data label, X sj is the source domain dataset, N is the total number of datasets, C operation is the Softmax classification prediction operation, F operation is the secondary sub-network j feature extraction operation, and G operation is the primary main network general feature extraction operation; The function calculation formula of migration loss is: in, is the Frobenius norm, F j is the feature extraction operation of the secondary sub-network j, and G is the general feature extraction operation of the primary main network. is the source domain j sample, and its number is N sj , is the target domain sample, and its number is N t , the total number of source domain data sets is N; The function calculation formula for distance loss is: Among them, C j The operation is the Softmax classification prediction operation, F i is the feature extraction operation of the secondary sub-network j, and G is the general feature extraction operation of the primary main network. is the target domain sample, and its number is N t , the total number of source domain data sets is N; The function contribution of the overall loss is matched in a weighted manner, and a hyperparameter search is performed on the parameters in the overall loss function formula to obtain the optimal parameter group.

2. The pump fault diagnosis interactive system based on distributed transfer learning according to claim 1 is characterized in that: The multiple types of sensors include any one or more of a three-axis vibration accelerometer sensor, a three-axis vibration velocity sensor, a three-phase current transformer, a three-phase voltage transformer, an acoustic emission sensor, a thermal imaging sensor or a camera.

3. The pump fault diagnosis interactive system based on distributed transfer learning according to claim 1 is characterized in that: The distributed slave station module also includes a model parameter storage unit and a communication module supporting multiple wireless communication modes. The model parameter storage unit is used to store the gradient parameters trained by the distributed slave station module and the updated gradient parameters from the parameter master station server.

4. The machine pump fault diagnosis interactive system based on distributed transfer learning according to claim 3 is characterized in that: Encryption consists of two parts: public key pk and private key sk. The encryption process is to input a ciphertext for the plaintext polynomial m∈R: c←r·pk+(m+e0,e1)mod Q Among them, c is the ciphertext, r is the size of the plaintext space, pk is the public key, m is the plaintext, e0, e1 is the noise, and Q is the modulus; The decryption process is to Calculate with private key sk: m=c0+Σ(c1·sk i +e i ) Among them, m is the plaintext, c0, c1 is the ciphertext, sk i is a single private key, e i is the error term; And return the plaintext; Gradient parameters δ uploaded by different distributed slave modules i and the amount of data k i Determine the proportion of weighted gradients, N is the total number of slaves, and the weighted formula is as follows: The obtained weighted average gradient is encrypted and then transmitted to the distributed slave module through the communication gateway for gradient update.

5. An interactive method for pump fault diagnosis based on distributed transfer learning, characterized in that: The pump fault diagnosis interactive system based on distributed transfer learning as described in any one of claims 1 to 4 comprises the following steps: Install sensor arrays on motors, pumps, fans, and reducer equipment in multiple plants; The distributed slave module collects data from the sensor array and performs model training. The training process is synchronized regularly. After each round of training, the gradient parameters are encrypted by the encryption and decryption device, and the encrypted gradients are sent to the parameter master server regularly. After the parameter master server obtains a round of encrypted gradients from all slave stations, it performs weighted calculation on the decrypted gradients to obtain the average gradient of this round of calculation, and then encrypts the average gradient and sends it to each distributed slave station module to update the gradient.

6. A machine pump fault diagnosis interactive device based on distributed transfer learning, characterized in that: It includes an interactive system for machine pump fault diagnosis based on distributed transfer learning as described in any one of claims 1 to 4.

Images