A distributed edge intelligent diagnosis method based on bee colony self-organizing network

Abstract

The present invention discloses a distributed edge intelligent diagnosis method based on a bee colony self-organizing network, which belongs to the field of mechanical equipment fault diagnosis. The present invention includes: deploying edge nodes for a preset number of target objects, and each edge node collects the original vibration waveform data of the target object in an equal time interval manner; converting the original vibration waveform data of the target object into a two-dimensional time-frequency graph; training and verifying the distributed edge intelligent diagnosis model according to the collected historical time-frequency graph to obtain a trained distributed edge intelligent diagnosis model; deploying the trained distributed edge intelligent diagnosis model on each edge node; constructing a distributed edge intelligent diagnosis network based on a bee colony self-organizing network; converting real-time data into a two-dimensional time-frequency graph to obtain a diagnosis result through a trained distributed edge intelligent diagnosis model. The present invention solves the problems of difficulty in managing multi-node equipment fault diagnosis, difficulty in deploying deep learning large models on edge devices, and difficulty in data management.

Description

Technical Field

[0001] The invention relates to a distributed edge intelligent diagnosis method based on a bee colony self-organizing network, and belongs to the field of mechanical equipment fault diagnosis. Background Art

[0002] At present, the development of the manufacturing industry is receiving more and more attention, and the vigorous development of the manufacturing industry is inseparable from various mechanical equipment. Many mechanical equipment are prone to degradation or even failure due to long-term, high-load operation and lack of professional maintenance, which may affect the equipment life and production efficiency at the least, or cause casualties at the worst.

[0003] However, facing a large number of mechanical equipment, how to achieve effective real-time management has become a major problem. In order to detect equipment failures in time and reduce the impact on production, many companies have invested too many professionals in supervision. However, facing a large number of mechanical equipment, labor efficiency is low, untimely and costly, and this requires managers to have high professional knowledge, otherwise it is difficult to correctly judge the problems of the equipment in time. Many companies have deployed a complete equipment status monitoring system, that is, installing a large number of sensors on the equipment, transmitting a large amount of collected data to a central processor for analysis and processing to monitor the equipment status, but this method has problems such as large data transmission volume, high processing pressure on the central processor and difficult wiring. Summary of the invention

[0004] The present invention provides a distributed edge intelligent diagnosis method based on a bee colony self-organizing network, so as to overcome at least one technical problem existing in the prior art.

[0005] The technical solution of the present invention is:

[0006] A distributed edge intelligent diagnosis method based on a bee colony self-organizing network includes the following steps:

[0007] Data edge collection and conversion steps: deploy edge nodes for a preset number of target objects, and each edge node collects the original vibration waveform data of the target object at equal time intervals; after each collection is completed, the original vibration waveform data of the target object is converted into a two-dimensional time-frequency graph;

[0008] Model training and deployment steps: Based on the collected historical time-frequency graphs, the distributed edge intelligent diagnosis model is trained and verified to obtain a trained distributed edge intelligent diagnosis model; the trained distributed edge intelligent diagnosis model is deployed on each edge node;

[0009] Steps for constructing a distributed edge intelligent diagnosis network based on a swarm self-organizing network: connect each edge node with its own swarm gateway to form a corresponding edge intelligent diagnosis node, and construct a distributed edge intelligent diagnosis network based on a swarm self-organizing network for each edge intelligent diagnosis node based on the swarm self-organizing network technology;

[0010] Target object fault diagnosis steps: After the edge intelligent diagnosis node collects real-time data, it converts the real-time data into a two-dimensional time-frequency graph; the real-time two-dimensional time-frequency graph is used as the input of the trained distributed edge intelligent diagnosis model deployed in the edge intelligent diagnosis node to obtain the diagnosis result;

[0011] Cloud platform data management steps: randomly select an edge intelligent diagnosis node as the cloud node; connect the cloud node to the cloud gateway, and the cloud node obtains the numbers and diagnosis results of the remaining edge intelligent diagnosis nodes in the network through the distributed edge intelligent diagnosis network based on the swarm self-organizing network; the number and diagnosis result of the cloud node itself, and the numbers and diagnosis results of the remaining edge intelligent diagnosis nodes can be sent to the cloud gateway through the swarm gateway of the cloud node, and the cloud gateway sends the results to the cloud platform.

[0012] Furthermore, the distributed edge intelligent diagnosis method based on the swarm self-organizing network also includes an edge node data storage step; the edge node data storage step is specifically: each edge node stores the original vibration waveform data, the characteristic values ​​calculated based on the original vibration waveform data, and the generated time-frequency diagram according to a preset storage method.

[0013] Furthermore, the storage method is ring storage, and the storage space of the ring storage logically forms a ring. Data is written from one end and moves along the ring space until it reaches the other end; when the buffer is full, the new data will overwrite the oldest data.

[0014] Furthermore, the converting of the original vibration waveform data of the target object into a two-dimensional time-frequency diagram is specifically as follows: firstly filtering and denoising the original vibration waveform data of the target object, and then converting the filtered and denoised data into a two-dimensional time-frequency diagram using wavelet transform.

[0015] Furthermore, the model training and deployment steps are specifically as follows:

[0016] Construct teacher model and student model;

[0017] Create category labels for the collected historical time-frequency graphs; divide the labeled historical time-frequency graphs into training sets and validation sets according to a preset ratio; use the training set and validation set as inputs of the teacher model to obtain a trained teacher model;

[0018] The training set and validation set are used as the input of the student model. During each round of student model training, the outputs of the student model and the trained teacher model are softened respectively by the set temperature parameters to obtain their respective soft labels, and the distillation loss is calculated by the soft labels. Then, the student loss is calculated using the hard labels generated by the student model without softening and the true labels. Finally, the total loss is obtained by taking the weighted sum of the distillation loss and the student loss. Finally, minimizing the total loss is used as the training goal to obtain the trained student model, and the trained student model is deployed on each edge node as the trained distributed edge intelligent diagnosis model.

[0019] Furthermore, the teacher model uses resnet50 as the framework, uses three 3×3 convolution kernels to replace the original 7×7 convolution kernel of resnet50; introduces a self-attention module in the residual block; and the student model uses depthwise separable convolution.

[0020] The beneficial effects of the present invention are: on the one hand, the present invention applies the model compression technology of knowledge distillation to transfer the knowledge of a large and complex model (teacher model) to a small model (student model), which greatly reduces the number of parameters, computational complexity and memory usage, thereby obtaining a high-performance lightweight model for deployment to edge nodes, which greatly reduces the diagnosis time while ensuring the accuracy of fault diagnosis; on the other hand, based on the edge nodes constructed in the first aspect, the present invention uses a swarm self-organizing network to build a one-to-many distributed fault diagnosis network, that is, multiple edge intelligent diagnosis nodes close to the target object are responsible for collecting data for diagnosis and broadcasting the diagnosis results, and a receiving end is responsible for monitoring the results sent by the node and sending it to the cloud for monitoring, thereby realizing a complete process for mechanical equipment from data collection to fault diagnosis and then to the cloud; based on the distributed fault diagnosis network, a maximum of 1024 nodes can be networked. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a flowchart of the present invention;

[0022] Figure 2 A schematic diagram of the process of edge data collection, conversion, storage, and diagnosis provided according to an embodiment of the present invention;

[0023] Figure 3 A schematic diagram of data storage provided according to an embodiment of the present invention;

[0024] Figure 4 A schematic diagram of a model training process provided according to an embodiment of the present invention;

[0025] Figure 5 A schematic diagram of a residual block structure introducing a self-attention module according to an embodiment of the present invention;

[0026] Figure 6 A schematic diagram of a network topology structure provided by an embodiment of the present invention;

[0027] Figure 7 A schematic diagram of the cloud migration process of a swarm ad hoc network data provided in an embodiment of the present invention;

[0028] Figure 8 A schematic diagram of data transmission in a swarm self-organizing network provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical scheme and advantages of the embodiments of the present invention clearer, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. It should be noted that the embodiments in this application and the features in the embodiments can be combined with each other arbitrarily without conflict.

[0030] When managing a large number of devices, it is difficult for traditional manual work to correctly judge the status of the devices, and the efficiency is low and the cost is high; while the method of deploying a large number of sensors to collect data to the central processor has problems such as large data transmission volume, high processing pressure on the central processor and difficult wiring. The present invention provides a distributed edge intelligent diagnosis method based on a bee colony self-organizing network. The distributed edge intelligent diagnosis method based on a bee colony self-organizing network provided by the present invention is described below:

[0031] like Figure 1-Figure 8 As shown, according to a first aspect of an embodiment of the present invention, a distributed edge intelligent diagnosis method based on a swarm ad hoc network is provided, comprising:

[0032] S1. Data edge acquisition and conversion: Edge nodes are deployed for a preset number of target objects. Each edge node acquires the original vibration waveform data of the target object in an equal time interval manner. After each acquisition is completed, the original vibration waveform data of the target object is converted into a two-dimensional time-frequency graph. For example, the equal time interval acquisition method can be specifically: each acquisition lasts 1 second, and the time interval between two acquisitions is 60 seconds. Each edge node is established using a Raspberry Pi and a vibration acquisition card (refer to Figure 2 ), the vibration acquisition card adopts MCC172, and taking the target object as a rolling bearing as an example, the original vibration waveform data is collected according to MCC172 and transmitted to the Raspberry Pi, and the Raspberry Pi is used to convert the original vibration waveform data of the target object into a two-dimensional time-frequency diagram, which can be further used to calculate the eigenvalue.

[0033] Furthermore, the converting of the original vibration waveform data of the target object into a two-dimensional time-frequency diagram is specifically as follows: firstly filtering and denoising the original vibration waveform data of the target object, and then converting the filtered and denoised data into a two-dimensional time-frequency diagram using wavelet transform.

[0034] Furthermore, the wavelet transform is a CMOR (Complex Morlet) continuous wavelet transform, which decomposes the original vibration waveform data into different frequency scales and time positions, and generates a time-frequency graph as a model input for training;

[0035] The continuous wavelet transform The formula is: ;

[0036] Among them, a is the scale, b is the translation, is the conjugate of the wavelet function, Represents raw vibration waveform data.

[0037] S2. Edge node data storage: Each edge node stores the original vibration waveform data, the characteristic values ​​calculated based on the original vibration waveform data, and the generated time-frequency diagram according to a preset storage method.

[0038] Furthermore, the storage method is a ring storage, and the storage space of the ring storage logically forms a ring, and data is written from one end and moves along the ring space until it reaches the other end; when the buffer is full, the new data will overwrite the oldest data, thereby avoiding the problem of buffer overflow. Figure 3 , each edge node stores the original vibration waveform data, the eigenvalues ​​calculated based on the original vibration waveform data, and the generated time-frequency diagram in the corresponding folder according to the circular storage method.

[0039] By applying the above technical solution, it can be seen that in order to solve the problem of limited storage capacity of edge devices, the present invention adopts a ring storage method to store data. All data is stored in a specific directory and overwritten at a certain time interval. This storage method can not only reduce the occupancy of node storage, but also retain data of sufficient time period for data tracing.

[0040] It should be noted that the characteristic values ​​include but are not limited to: peak-to-peak value, mean, variance, standard deviation, mean square value, root mean square value, mean square error, waveform factor, kurtosis factor, peak factor, skewness, frequency domain peak, six-segment spectrum energy, and twelve-segment harmonic energy; the main function of the above-mentioned characteristic values ​​is to be used for data traceability (the data traceability specifically lies in: for the diagnosis results obtained by the distributed edge intelligent diagnosis method based on the bee swarm self-organizing network, the monitoring personnel can use the characteristic values ​​stored on the edge node to verify the diagnosis results obtained by the distributed edge intelligent diagnosis method based on the bee swarm self-organizing network to achieve traceability).

[0041] S3, model training and deployment: Based on the collected historical time-frequency graphs, the distributed edge intelligent diagnosis model is trained and verified to obtain a trained distributed edge intelligent diagnosis model; the trained distributed edge intelligent diagnosis model is deployed on each edge node.

[0042] refer to Figure 4 , the model training and deployment are specifically as follows:

[0043] Construct teacher model and student model;

[0044] Create category labels for the collected historical time-frequency graphs; divide the labeled historical time-frequency graphs into training sets and validation sets according to a preset ratio; use the training set and validation set as inputs of the teacher model to obtain a trained teacher model;

[0045] The training set and validation set are used as the input of the student model. During each round of student model training, the outputs of the student model and the trained teacher model are softened respectively by the set temperature parameter T to obtain their respective soft labels, and the distillation loss is calculated by the soft labels. Then, the student loss is calculated using the hard labels generated by the student model without softening and the true labels. Finally, the total loss is obtained by taking the weighted sum of the distillation loss and the student loss. Finally, minimizing the total loss is used as the training goal to obtain the trained student model, and the trained student model is deployed on each edge node as the trained distributed edge intelligent diagnosis model.

[0046] The teacher model in knowledge distillation needs to have high performance, which is very important for improving the accuracy of the student model. Therefore, the present invention uses resnet50 as the framework to build the teacher model. Resnet50 effectively alleviates the gradient vanishing problem, has a more complex hierarchical structure, can more effectively extract features, and obtain better accuracy. In order to obtain better performance, the present invention optimizes resnet50: first, three 3×3 convolution kernels are used to replace the original 7×7 convolution kernel, which deepens the network hierarchy and the model can obtain better performance; second, a self-attention module is introduced in the residual block (an example of a residual block introducing a self-attention module is shown in FIG. Figure 5As shown). In the above, the self-attention module is as follows: first, a 1×1 convolution layer is used to reduce the channel dimension of the input feature map to a single channel, and the single-channel feature map is multiplied by an adjustment factor α; then the adjusted single-channel map is normalized using the sigmoid activation function to obtain an attention weight distribution between 0 and 1; finally, the weight is used to weight the input feature map of the self-attention module element by element, so as to better emphasize or suppress the feature responses at different positions, thereby enhancing the model's ability to capture key information and improving the model's feature extraction and representation performance. The teacher model parameters are shown in Table 1 below:

[0047] Table 1

[0048]

[0049] Considering the real-time requirements in edge diagnosis and the limited computing power of edge devices, the model should be as small as possible. The student model of the present invention adopts depthwise separable convolution, which decomposes the standard convolution operation into depthwise convolution and pointwise convolution; depthwise convolution is a convolution operation performed on each input channel separately, which is only responsible for filtering and does not change the number of channels; pointwise convolution uses a 1×1 convolution kernel to fuse the channel information after depth convolution and change the number of channels. This structure greatly reduces the number of parameters and the amount of calculation of the model, making the model more lightweight and suitable for resource-constrained environments such as mobile devices. Both the student model and the teacher model belong to convolutional neural networks, and they both use convolution layers to extract features of data such as images. In both architectures, convolution operations are the core parts. Therefore, it is more helpful for the student model to receive the rich knowledge in the soft labels of the teacher model, so that the student model can obtain a higher accuracy. The parameters of the student model are shown in the following table. The parameter size of the student model is only 0.8MB, which is less than one percent of the parameter size of the teacher model of 86.4MB. The parameters of the student model are shown in Table 2 below:

[0050] Table 2

[0051]

[0052] The outputs of the student model and the trained teacher model are softened by setting the temperature parameter T to obtain their respective soft labels. The calculation formula is:

[0053] ;

[0054] Where T is the temperature parameter, , are the logits vectors of the i-th and j-th categories respectively. Applying the above technical solution, we can see that when T=1, it is the standard Softmax formula; when T>1, it is the softened Softmax formula.

[0055] The total loss function L is calculated as follows:

[0056] ;

[0057] Among them, L S is the distillation loss function, L H is the student loss function, is the coefficient.

[0058] By applying the above technical solution, it can be seen that the present invention first learns a large amount of historical data through the teacher model to obtain sufficient information and diagnostic knowledge; then trains the student model, and its training goal is to minimize two loss functions; one is the student loss related to the real label, which is similar to the traditional model training, to ensure that the student model can be correctly classified; the other is the distillation loss related to the soft label of the teacher model, and the distillation loss makes the output probability distribution of the student model as close as possible to the soft label of the teacher model, thereby ensuring that knowledge can be learned from the teacher model and an efficient and lightweight student model is obtained; then the trained student model is deployed to the edge node to realize diagnostic analysis at the edge.

[0059] Exemplarily, the diagnostic model is used for the diagnosis of rolling bearings, and the diagnostic categories include: normal, roller fault, outer ring fault, inner ring fault, outer ring + roller fault, inner ring + roller fault, outer ring + inner ring fault, outer ring + inner ring + roller fault, and the corresponding labels are Arabic numerals 0, 1, 2, 3, 4, 5, 6, 7.

[0060] S4. Construct a distributed edge intelligent diagnosis network based on bee colony self-organizing network: connect each edge node with its own bee colony gateway to form a corresponding edge intelligent diagnosis node, and construct a distributed edge intelligent diagnosis network based on bee colony self-organizing network for each edge intelligent diagnosis node based on bee colony self-organizing network technology to realize the communication and data transmission of edge intelligent diagnosis nodes in the network. After adding or deleting nodes, the network can automatically identify and update the network topology structure; Figure 6 The figure shows an example of the network topology structure provided for this embodiment; the distributed edge intelligent diagnosis network based on the swarm self-organizing network is a network composed of devices that can move autonomously, and can form a dynamic, self-organizing network through mutual communication and collaboration without a central controller. It has the characteristics of self-organization, self-repair, high flexibility, and distribution; supports 16-level physical layer relay, 1 second fast networking, and communication with 1024 devices; can perform conversion between serial port / network port and swarm wireless network, and can support the function of wireless switch.

[0061] S5. Target object fault diagnosis: After the edge intelligent diagnosis node collects real-time data, it converts the real-time data into a two-dimensional time-frequency graph; the real-time two-dimensional time-frequency graph is used as the input of the trained distributed edge intelligent diagnosis model deployed in the edge intelligent diagnosis node to obtain the diagnosis result.

[0062] S6. Cloud platform data management: randomly select an edge intelligent diagnosis node as the cloud node; connect the cloud node to the cloud gateway, and the cloud node obtains the numbers and diagnosis results of the other edge intelligent diagnosis nodes in the network through the distributed edge intelligent diagnosis network based on the swarm self-organizing network; the number and diagnosis result of the cloud node itself, and the numbers and diagnosis results of the other edge intelligent diagnosis nodes can be sent to the cloud gateway through the swarm gateway of the cloud node, and the cloud gateway sends the results to the cloud platform through the DTU network for statistics and display.

[0063] refer to Figure 7 , 8 According to the distributed edge intelligent diagnosis network based on swarm self-organizing network established by S4, an edge intelligent diagnosis node is randomly selected as the cloud node. The cloud node is responsible for continuously monitoring the numbers and diagnosis results from the remaining edge intelligent diagnosis nodes. After completing the diagnosis, the remaining edge intelligent diagnosis nodes will broadcast the numbers and diagnosis results to the swarm gateway of the cloud node through their own swarm gateway. The broadcasted data includes the diagnosis result and the node identification number. The cloud node as the receiving end identifies the data source and the corresponding diagnosis result after receiving a data, and then packages the received data and sends it to the cloud gateway, which then implements the data clouding.

[0064] The basic function of the cloud gateway is to realize two-way transparent transmission through RS232 / 485 to DTU, that is, it can convert the data received by the serial port into DTU and transmit it to the cloud.

[0065] By applying the above technical solution, it can be known that the present invention uses a swarm self-organizing network to build a multi-node distributed edge intelligent diagnosis network with self-organization. The network can realize the joint management of a large number of devices. When adding or deleting device nodes, it can also automatically identify and update the network topology. The network data is wirelessly transmitted without the need for wiring. At the same time, the edge nodes in the network deploy lightweight distributed edge intelligent diagnosis models to realize local processing and fault diagnosis of data, avoiding large amounts of data transmission; data clouding organizes the real-time status of all node devices to the cloud, and operators can easily monitor in real time, greatly improving the efficiency and cost of equipment management.

[0066] According to a second aspect of an embodiment of the present invention, a distributed edge intelligent diagnosis system based on a swarm ad hoc network is provided, comprising the following modules:

[0067] Data edge acquisition and conversion module: used to deploy edge nodes for a preset number of target objects, and each edge node collects the original vibration waveform data of the target object at equal time intervals; after each acquisition is completed, the original vibration waveform data of the target object is converted into a two-dimensional time-frequency graph;

[0068] Model training and deployment module: used to train and verify the distributed edge intelligent diagnosis model based on the collected historical time-frequency graphs, obtain the trained distributed edge intelligent diagnosis model; deploy the trained distributed edge intelligent diagnosis model on each edge node;

[0069] Construct a distributed edge intelligent diagnosis network module based on bee colony self-organizing network: used to connect each edge node with its own bee colony gateway to form a corresponding edge intelligent diagnosis node, and construct a distributed edge intelligent diagnosis network based on bee colony self-organizing network for each edge intelligent diagnosis node based on bee colony self-organizing network technology;

[0070] Target object fault diagnosis module: After the edge intelligent diagnosis node collects real-time data, it converts the real-time data into a two-dimensional time-frequency graph; the real-time two-dimensional time-frequency graph is used as the input of the trained distributed edge intelligent diagnosis model deployed in the edge intelligent diagnosis node to obtain the diagnosis result;

[0071] Cloud platform data management module: used to randomly select an edge intelligent diagnosis node as a cloud node; connect the cloud node to the cloud gateway, and the cloud node obtains the numbers and diagnosis results of the remaining edge intelligent diagnosis nodes in the network through the distributed edge intelligent diagnosis network based on the swarm self-organizing network; the number and diagnosis result of the cloud node itself, and the numbers and diagnosis results of the remaining edge intelligent diagnosis nodes can be sent to the cloud gateway through the swarm gateway of the cloud node, and the cloud gateway sends the results to the cloud platform.

[0072] For the parts of the modules not described in detail above, please refer to the relevant description of this embodiment.

[0073] The specific implementation modes of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above implementation modes, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

Claims

1. A distributed edge intelligent diagnosis method based on bee colony self-organizing network, characterized in that: The following steps are involved: Data edge collection and conversion steps: deploy edge nodes for a preset number of target objects, and each edge node collects the original vibration waveform data of the target object at equal time intervals; after each collection is completed, the original vibration waveform data of the target object is converted into a historical two-dimensional time-frequency graph; Model training and deployment steps: Based on the collected historical two-dimensional time-frequency graphs, the distributed edge intelligent diagnosis model is trained and verified to obtain a trained distributed edge intelligent diagnosis model; the trained distributed edge intelligent diagnosis model is deployed on each edge node; Steps for constructing a distributed edge intelligent diagnosis network based on a swarm self-organizing network: connect each edge node with its own swarm gateway to form a corresponding edge intelligent diagnosis node, and construct a distributed edge intelligent diagnosis network based on a swarm self-organizing network for each edge intelligent diagnosis node based on the swarm self-organizing network technology; Target object fault diagnosis steps: After the edge intelligent diagnosis node collects real-time data, the real-time data is converted into a two-dimensional time-frequency graph; the real-time two-dimensional time-frequency graph is used as the input of the trained distributed edge intelligent diagnosis model deployed in the edge node of the edge intelligent diagnosis node to obtain the diagnosis result; Cloud platform data management steps: randomly select an edge intelligent diagnosis node as the cloud node; connect the cloud node to the cloud gateway, and the cloud node obtains the numbers and diagnosis results of the other edge intelligent diagnosis nodes in the network through the distributed edge intelligent diagnosis network based on the swarm self-organizing network; the number and diagnosis result of the cloud node itself, and the numbers and diagnosis results of the other edge intelligent diagnosis nodes can be sent to the cloud gateway through the swarm gateway of the cloud node, and the cloud gateway sends the results to the cloud platform.

2. The distributed edge intelligent diagnosis method based on bee colony ad hoc network according to claim 1 is characterized in that: It also includes an edge node data storage step; the edge node data storage step is specifically: each edge node stores the original vibration waveform data, the characteristic values ​​calculated based on the original vibration waveform data, and the generated historical two-dimensional time-frequency diagram according to a preset storage method.

3. The distributed edge intelligent diagnosis method based on bee colony ad hoc network according to claim 2 is characterized in that: The storage method is ring storage, and the storage space of the ring storage logically forms a ring. Data is written from one end and moves along the ring space until it reaches the other end; when the buffer is full, the new data will overwrite the oldest data.

4. The distributed edge intelligent diagnosis method based on bee colony ad hoc network according to claim 1 is characterized in that: The method of converting the original vibration waveform data of the target object into a historical two-dimensional time-frequency diagram specifically comprises: firstly filtering and denoising the original vibration waveform data of the target object, and then converting the filtered and denoised data into a historical two-dimensional time-frequency diagram using wavelet transform.

5. The distributed edge intelligent diagnosis method based on bee colony ad hoc network according to claim 1 is characterized in that: The model training and deployment steps are specifically as follows: Construct teacher model and student model; Create category labels for the collected historical two-dimensional time-frequency graphs; divide the labeled historical two-dimensional time-frequency graphs into training sets and validation sets according to a preset ratio; use the training set and validation set as inputs of the teacher model to obtain a trained teacher model; The training set and validation set are used as the input of the student model. During each round of student model training, the outputs of the student model and the trained teacher model are softened respectively by the set temperature parameters to obtain their respective soft labels, and the distillation loss is calculated by the soft labels. Then, the student loss is calculated using the hard labels generated by the student model without softening and the true labels. Finally, the total loss is obtained by taking the weighted sum of the distillation loss and the student loss. Finally, minimizing the total loss is used as the training goal to obtain the trained student model, and the trained student model is deployed on each edge node as the trained distributed edge intelligent diagnosis model.

6. The distributed edge intelligent diagnosis method based on bee colony self-organizing network according to claim 5 is characterized in that: The teacher model uses resnet50 as the framework, and uses three 3×3 convolution kernels to replace the original 7×7 convolution kernel of resnet50; a self-attention module is introduced in the residual block; and the student model uses depthwise separable convolution.

Images