Small hermetic compressor hidden defect detection method based on graph machine learning

Abstract

The application discloses a kind of small fully-enclosed compressor implicit defect detection methods based on graph machine learning, focus on solving small fully-enclosed compressor when factory whole machine inspection, since compressor model is numerous and frequently replaced, traditional machine learning model generalization is insufficient, and lead to the problem of diagnostic accuracy decline after model replacement.This application makes full use of the vibration signal of the shell during the operation of the compressor, which is converted into a graph structure according to a special graphing method based on logical relationship, and then characterized by graph convolutional neural network, to mine more common features in the vibration signal, thereby improving the generalization of the model.The application effectively improves the problem that the existing small fully-enclosed compressor implicit defect diagnosis method cannot be applied to the diagnosis of multiple models of compressors, and improves the possibility of practical application of implicit defect diagnosis technology.

Description

Technical Field

[0001] This invention belongs to the technical field of rotating machinery fault diagnosis methods, specifically relating to a method for detecting latent defects in a small, fully enclosed compressor based on graph machine learning. Background Technology

[0002] As a core component of household refrigeration equipment such as refrigerators, the small, fully enclosed compressor's inherent defects can significantly reduce the performance and lifespan of the refrigeration system. Among these defects, latent defects are the most difficult to identify. These refer to flaws that arise during the compressor's production process due to variations in component manufacturing techniques and assembly precision. Compressors with latent defects may function normally during the initial factory inspection, failing to reach a failure level, but the malfunction gradually becomes apparent after the compressor has been in operation for a period of time. Therefore, accurately identifying latent defects in compressors during factory inspection and preventing malfunctions before they actually occur is of great practical significance.

[0003] Currently, with the development and popularization of artificial intelligence technology, deep learning technology has been widely used in the field of rotating machinery fault diagnosis. This technology greatly saves manpower and resources, and can extract more complex and reliable features when there are sufficient samples. However, traditional deep learning methods extract features from data as a whole, considering only the numerical attributes of the data itself, ignoring the influence of physical factors such as time and space between data on feature extraction. The single source of feature indicators leads to a lack of generalization in the extracted features.

[0004] In recent years, graph machine learning models, which combine graph data and deep neural networks, have shown great potential in the field of rotating machinery fault diagnosis. Graph data consists of a series of nodes and edges. Nodes represent the attributes and features of the data itself, while edges establish the relationships between nodes and are used to transmit and update node vectors. The graph structure contains richer data information, making it easier to extract high-dimensional features during representation learning. Graph data can not only mine the numerical features of the nodes themselves but also integrate features from neighboring nodes through edge connections, resulting in more powerful reasoning capabilities compared to ordinary neural networks. At the same time, the nodes in graph data are connected by constraints of basic physical laws, and the features mined by graph machine learning have better generalization ability compared to general deep neural networks. In addition, this data structure implicitly contains potential mandatory logical relationships, reducing the impact of some irrelevant interferences on the model, thus making the model more robust. Summary of the Invention

[0005] This invention addresses three problems in the final inspection stage of small fully enclosed compressor production: (1) Currently, most small fully enclosed compressor manufacturers still use manual inspection for quality control, resulting in serious misjudgments and omissions of compressors with hidden defects; (2) The vibration signal of the small fully enclosed compressor casing has great randomness and nonlinearity, making manual feature extraction extremely difficult; (3) There are many different models of compressors on the production line, and they are frequently changed, making it too labor-intensive and resource-intensive to model different models of compressors separately. This invention proposes a method for detecting hidden defects in small fully enclosed compressors based on graph machine learning. Addressing the definition of graph data nodes and the connection of edges, and considering the real-world situation of small fully enclosed compressor final inspection, a matching graph data construction scheme is proposed. This scheme converts the time-domain vibration signal into a graph structure to obtain richer information representation. The graph machine learning mechanism is then used to obtain more common feature indicators from the vibration signal, thereby improving the generalization of the diagnostic model.

[0006] The specific plan is as follows:

[0007] S1. Build a casing vibration signal acquisition system, use an accelerometer to generate vibration, and collect casing vibration data of multiple compressor models during operation in an actual production environment.

[0008] S2. Preprocess the collected raw vibration data to obtain vibration data for the complete start-stop process and establish the raw dataset.

[0009] Furthermore, considering that different types of latent defects exhibit different degrees of fault manifestation during the stable and non-stable phases of compressor operation, individual samples are divided into three phases: startup, stable, and termination, for feature extraction, thereby better uncovering fault characteristics.

[0010] S3. Time Domain Nodes: Taking the vibration data in the X-axis direction as an example, the vibration data in the three operating stages are divided to generate an even number of time domain nodes.

[0011] Furthermore, during the compressor's start-up and operation phase, the vibration signal exhibits a specific pattern over time. Connecting the nodes sequentially according to their sampling times reveals a physical relationship between the edges, implying a temporal progression.

[0012] Furthermore, during the stable operation phase of the compressor, the vibration data exhibits a periodic variation pattern as the internal mechanisms of the compressor operate smoothly. A stochastic strategy is used for edge connections to reduce the influence of irrelevant interference.

[0013] Furthermore, during the compressor's end-of-run phase, nodes are connected via edge pairs with implicit time attributes, and then connected to the time-domain nodes during the start-up phase, thus fusing the time-domain information from both phases.

[0014] S4. Frequency Domain Nodes: Taking vibration data in the X-axis direction as an example, the frequency domain features are calculated by concatenating the data from two time domain nodes to obtain frequency domain nodes. Furthermore, the frequency domain nodes of the three stages are connected to each other and connected to their corresponding time domain nodes. The frequency domain nodes of the start and end stages are also fused with the frequency domain information of the two stages through edge connections.

[0015] S5. Complete the construction of the X-axis vibration data subplot. Repeat steps 3-4 above to obtain the Y-axis and Z-axis vibration data subplots. Connect the nodes of the three subplots in pairs according to the structural order to complete the construction of a sample graph.

[0016] S6. Repeat steps 3-5 above for the samples in the original dataset to complete the creation of the vibration signal image sample dataset, and divide the training set and test set in an 8:2 ratio.

[0017] S7. Construct a Graph Convolutional Neural Network (GCN) model to perform feature extraction and pattern recognition on graph-structured data. The model consists of two graph convolutional layers and two dropout layers, followed by a fully connected layer, and finally a softmax classification layer to output diagnostic results.

[0018] S8. Feed the graph data training set into the graph convolutional neural network model described in S7 to train the optimal diagnostic model. The test set is used to evaluate the performance of the diagnostic model. The obtained model can detect and classify multiple models of small, fully enclosed compressors.

[0019] S9. Using the OpenVINO tool, the model is converted to IR format, and the IR model is deployed and called on the LabVIEW platform to improve the system inference speed and stability. In conjunction with the data acquisition device, a small fully enclosed compressor latent defect online diagnosis system is built.

[0020] The beneficial effects of this invention are as follows:

[0021] 1) Calculate the time-frequency domain characteristics of the vibration signal to generate data nodes, and use the forced basic physical laws to constrain the edges to construct the original vibration signal into graph data with rich logical relationships;

[0022] 2) The graph structure dataset generated by the graph construction method described in this invention can be efficiently extracted from vibration signals by training with GCN, thereby improving the generalization of the model and solving the problem of insufficient generalization of single diagnostic models in actual industrial scenarios due to the large number of compressor models.

[0023] 3) By using the OpenVINO tool to obtain the IR model and deploying and calling it using software, the system stability is effectively guaranteed, the model inference speed is improved, and the possibility of the diagnostic system being applied in practice is increased. Attached Figure Description

[0024] Figure 1 This is a flowchart of the data acquisition system of the present invention;

[0025] Figure 2 This is a waveform diagram of vibration data along the X-axis for a sample in the dataset.

[0026] Figure 3 This is a flowchart illustrating the graph data construction process proposed in this invention.

[0027] Figure 4 This invention presents a graph data composition scheme.

[0028] Figure 5 This is a diagram of the graph convolutional neural network model structure of the present invention;

[0029] Figure 6 This is the deployment scheme for the diagnostic system of the present invention. Detailed Implementation

[0030] To more clearly illustrate the purpose, technical solution, and advantages of the embodiments of the present invention, the method of the present invention will be described in detail below with reference to the relevant accompanying drawings and embodiments.

[0031] Example

[0032] S10. The deployment of the data acquisition system and the data acquisition work are as follows:

[0033] S101. The data acquisition system includes: a compressor, a compressor start-up platform, a triaxial accelerometer, a data acquisition board, and a host computer;

[0034] Hardware-wise: the compressors include VTB1114YL, VTH111Y, and VTB11133Y inverter compressors; the three-axis accelerometer is from KISTLER Corporation, model SN 5750330; the data acquisition board is from National Instruments, model NI-9234; the host computer is a laptop computer equipped with an i7-11800H CPU.

[0035] In terms of software: The operation panel of the data acquisition device was developed using LabVIEW 2018 64-bit software. This software can control the acquisition in real time, convert vibration data into visual images and save them as .csv files;

[0036] S102. Deploy the data acquisition system in the production environment and place the triaxial accelerometer at the center of the compressor housing, wherein the X-axis of the sensor is parallel to the axis of the compressor cylinder, the Y-axis is perpendicular to the X-axis and parallel to the placement plane of the compressor, and the Z-axis is perpendicular to the placement plane of the compressor.

[0037] S103. Data Acquisition System Workflow as follows: Figure 1 As shown, when the "Start Acquisition" button is pressed, the data acquisition board begins to receive vibration data from the accelerometer. Each acquisition lasts for 10 seconds and includes data from the entire operating cycle of the fully enclosed compressor, from power-on startup to stable operation and then to power-off shutdown.

[0038] S20. Data preprocessing, taking vibration data in the X-axis direction as an example. A program written in Python 3 is used to capture 3 seconds of vibration data during the complete start-stop process, such as... Figure 2 As shown, a total of 76,800 data points were collected. Considering that some latent defects, such as casing collisions, are more obvious during the non-steady operation phase of the compressor, while the faults are less likely to appear during steady operation, the first 25,600 data points were defined as the startup phase, the middle 25,600 data points as the steady phase, and the last 25,600 data points as the termination phase. To verify the generalization ability of the model, three sets of experimental datasets were designed, as detailed in Table 1 below. Dataset 1 was used to train the diagnostic model, and datasets 2 and 3 were used to verify the generalization ability of the diagnostic model.

[0039] Table 1 Experimental Dataset Information

[0040]

[0041]

[0042] S30. The graph data construction process is as follows: Figure 3 As shown, taking the vibration data in the X-axis direction as an example, the vibration data of the three running stages are divided to generate an even number of time-domain nodes. Considering that there are 25,600 data points in each of the three running stages, and that the node feature dimension of the graph data is too large, it will not only consume a lot of training time, but also be detrimental to feature mining. Therefore, the original data is divided into 12 groups of vectors as node attributes, with the dimension uniformly set to (6400,).

[0043] During the compressor's start-up and operation phase, the vibration signal exhibits a specific pattern over time. Connecting the nodes sequentially according to their sampling times reveals a physical relationship inherent in the time frame.

[0044] During the stable operation phase of the compressor, the vibration data exhibits a periodic variation pattern as the internal mechanisms of the compressor operate smoothly. A stochastic strategy is used for edge connections to reduce the influence of irrelevant interference.

[0045] Furthermore, during the compressor's end-of-run phase, nodes are connected via edge pairs with implicit time attributes, and then connected to the time-domain nodes during the start-up phase, thus fusing the time-domain information from both phases.

[0046] The specific partitioning steps and edge connection scheme are as follows:

[0047] S301. During the startup phase, nodes 1, 2, 3, and 4 are generated, as follows: Figure 2 As shown in the first second, the vibration data exhibits a clear temporal characteristic. It shows a pattern of large amplitude fluctuations at the initial moment, followed by irregular shaking in both positive and negative directions, eventually stabilizing. Therefore, sorting nodes according to their sampling time sequence is more effective in extracting useful features. Connecting nodes 1 and 2, 2 and 3, and 3 and 4 with edges according to their temporal relationship is shown below. Figure 4 As shown.

[0048] S302. During the stable phase, nodes 5, 6, 7, and 8 are generated, as follows: Figure 2 As shown in the second second, with the smooth operation of the compressor's internal mechanism, the vibration data exhibits stable periodic changes. To reduce the impact of irrelevant interference on the quality of extracted features, the four nodes are randomly shuffled and numbered 5, 6, 7, and 8. Then, nodes 5 and 6, 6 and 7, and 7 and 8 are connected by edges to further reduce the impact of irrelevant interference on the quality of extracted features. Figure 4 As shown.

[0049] S303. In the final stage, nodes 9, 10, 11, and 12 are generated, and the nodes are connected by edges in chronological order. Nodes 1 and 12, 2 and 11, 3 and 10, 4 and 9, 13 and 18, and 14 and 17 are also connected by edges. The node information from the start and end stages is integrated to perform characterization learning of the vibration signal in the non-stationary operation stage.

[0050] S40. Taking vibration data along the X-axis as an example, the frequency domain features are calculated by concatenating the data from two time-domain nodes to obtain frequency domain nodes. The frequency domain features of the vibration data are calculated using Fast Fourier Transform (FFT). Considering the symmetry of the Fourier transform results and the uniformity of the graph data node dimensions, frequency domain nodes 13 to 18 are obtained, with a unified dimension of (6400,). The specific calculation steps and edge connection scheme are as follows:

[0051] S401. During the startup phase, the data from nodes 1 and 2, and nodes 3 and 4 are concatenated separately, and the frequency domain characteristics are calculated using FFT. The calculated ΔF is 2Hz according to the following formula.

[0052]

[0053] Where ΔF is the spectral resolution, i.e., the spectral interval, f s N is the sampling frequency, and N is the number of sampling points.

[0054] After transformation, 12,800 data points are obtained. Half of the results are taken as node attributes, resulting in two frequency domain nodes, 13 and 14.

[0055] S402. During the steady phase, the data from nodes 5 and 6, and 7 and 8 are concatenated, and the frequency domain features are calculated using FFT to obtain two frequency domain nodes, 15 and 16.

[0056] S403. In the final stage, the data of nodes 9 and 10, and 11 and 12 are concatenated respectively, and the frequency domain features are calculated by FFT to obtain two frequency domain nodes, 17 and 18.

[0057] S404. Connect nodes 13 and 14, 15 and 16, and 17 and 18 with edges to facilitate the fusion of frequency domain features of the three-stage vibration data. Connect nodes 13 and 1, 13 and 2, 14 and 3, 14 and 4, 15 and 5, 15 and 6, 16 and 7, 16 and 8, 17 and 9, 17 and 10, 18 and 11, and 18 and 12 with edges to facilitate the fusion of time-frequency domain features of the three-stage vibration data.

[0058] S50. Complete the subplot construction for the X-axis vibration data. Repeat steps S30-S40 above to obtain subplots for the Y-axis and Z-axis vibration data.

[0059] Connect the three subgraphs by edges, linking 1 to 1' and 1' to 1" sequentially. Repeat this process to connect the remaining nodes, fusing the vibration data from the three axes. This completes the construction of the vibration data graph sample, as shown below. Figure 4 As shown.

[0060] S60. Repeat steps S30-S50 for the three datasets described in Table 1 to create the graph sample datasets, and divide them into training and test sets in an 8:2 ratio. Use dataset one for model training, and datasets two and three for model generalization validation.

[0061] S70. Construct a graph convolutional neural network model containing two graph convolutional layers, two dropout layers, one fully connected layer (Dense layer), and one output layer (Softmax layer). The graph convolutional layers extract features, the dropout layers prevent overfitting, the extracted features are fused in the Dense layer, and the Softmax layer serves as the diagnostic output layer. Figure 5 The diagram shows the topology of a graph convolutional neural network.

[0062] The GraphConvolution layer utilizes a message-passing mechanism to fuse features from neighboring nodes, enrich its own node information, and extract latent features from the data. This process can be described by the following mathematical model:

[0063]

[0064] Where W represents the weight matrix; A represents the adjacency matrix of the graph; Represents an adjacency matrix with self-connections I represents the identity matrix; H represents the degree matrix of the adjacency matrix with self-connections; H represents the node features of the network; and σ() is the activation function.

[0065] S80. 4000 samples from Dataset 1 are used as the training set and fed into the graph convolutional neural network model described in S70. 1000 samples are used as the test set to evaluate the performance of the diagnostic model. The training epochs are 20, the loss function is categorical_crossentropy, the batch size is 16, and the optimizer is Adam.

[0066] The generalization ability of the model was verified using datasets two and three. Training data was created using the graphing method described in this invention. Based on the fundamental physical laws enforced between the data, representation learning was performed. GCN can extract more generalizable feature indicators from vibration signals, making it applicable to different models of compressors with similar internal structures.

[0067] S90. Utilizing the diagnostic model of S80, the data acquisition system of S10, and LabVIEW software, an online diagnostic system for latent defects in a fully enclosed compressor is built. The specific construction scheme and steps are as follows:

[0068] The S901 utilizes the OpenVINO toolkit developed by Intel, and its specific deployment scheme is as follows: Figure 6 As shown, the optimal model obtained in S80 is converted into IR format, which effectively improves the inference speed of the model. The IR model includes an .xml file describing the network topology and a .bin file containing network weights and biases.

[0069] S902. The IR model is deployed using the DNN module of LabVIEW software. In conjunction with the data acquisition system, the IR model is called to perform inference on the vibration data. Deploying the diagnostic system on the industrial production line enables online diagnostics.

Claims

1. A method for detecting latent defects in a small, fully enclosed compressor based on graph machine learning, characterized in that... Includes the following steps: 1) Build a casing vibration signal acquisition system, use an accelerometer to generate vibration, and collect casing vibration data of multiple compressor models during operation in an actual production environment; 2) Preprocess the collected raw vibration data to obtain vibration data for the complete start-stop process and establish the raw dataset; 3) Divide the vibration data of a single sample into three stages: start-up, steady-state, and end. Taking the vibration data in the X-axis direction as an example, divide the vibration data of the three operating stages and generate an even number of time domain nodes. 4) Taking the vibration data in the X-axis direction as an example, the frequency domain features are calculated by concatenating the data from two time domain nodes to obtain the frequency domain nodes; 5) Complete the construction of the X-axis vibration data subplot. Repeat steps 3) and 4) to obtain the Y-axis and Z-axis vibration data subplots respectively. The plot construction in steps 3) to 5) specifically includes: a) Time-domain node connection: During the start-up and end-up phases, time-domain nodes are connected sequentially according to the sampling time order; during the stationary phase, a random strategy is used to connect time-domain nodes by edges; at the same time, the corresponding time-domain nodes of the start-up and end-up phases are connected to fuse the time-domain information of the two phases. b) Frequency domain node connection: Connect the frequency domain nodes generated in the same stage, and connect each frequency domain node with the two time domain nodes that generated it; at the same time, connect the frequency domain nodes of the start stage and the end stage to fuse the frequency domain information of the two stages. c) Three-axis subgraph connection: Connect nodes with the same structural order in the X, Y, and Z direction subgraphs in pairs; 6) Repeat steps 3-5 above for the samples in the original dataset to complete the creation of the vibration signal image sample dataset, and divide the training set and test set in an 8:2 ratio; 7) Construct a graph convolutional neural network model to perform feature extraction and pattern recognition on graph-structured data; 8) Feed the graph data training set into the graph convolutional neural network model described in step 7) to train the optimal diagnostic model. The test set is used to evaluate the performance of the diagnostic model. The obtained model can detect and classify multiple models of small fully enclosed compressors. 9) Using the OpenVINO tool, the model is converted into IR format, and the IR model is deployed and called on the LabVIEW platform to improve the system inference speed and stability. With the help of data acquisition devices, the diagnostic system can be deployed on the industrial production line to realize online diagnosis.

2. The method for detecting latent defects in a small, fully enclosed compressor based on graph machine learning as described in claim 1, characterized in that... In step 2), considering that different types of latent defects exhibit different degrees of fault manifestation during the stable and non-stable phases of compressor operation, a single sample is divided into three phases: startup, stable, and termination, for feature extraction, thereby better uncovering fault features.

3. The method for detecting latent defects in a small, fully enclosed compressor based on graph machine learning as described in claim 2, characterized in that... Step 3) consists of three operating phases: startup phase, stable operation phase, and shutdown phase. During the startup phase of the compressor, the vibration signal exhibits a specific pattern over time. Nodes are connected sequentially according to the sampling time, and the connection of edges implies the physical relationship of time. During the stable operation phase of the compressor, the vibration data exhibits a periodic change pattern as the internal mechanism of the compressor operates smoothly. A random strategy is used to connect edges to reduce the influence of irrelevant interference. During the shutdown phase of the compressor, nodes are connected through edges with implicit time attributes and connected to the time domain nodes of the startup phase to fuse the time domain information of the two phases.

4. The method for detecting latent defects in a small, fully enclosed compressor based on graph machine learning as described in claim 3, characterized in that... The frequency domain nodes of the three stages are connected to each other and connected to their corresponding time domain nodes. The frequency domain nodes of the start and end stages are also connected by edges to fuse the frequency domain information of the two stages.

5. The method for detecting latent defects in a small, fully enclosed compressor based on graph machine learning as described in claim 4, characterized in that... The graph convolutional neural network model in step 7) consists of two graph convolutional layers and two Dropout layers, followed by a fully connected layer, and finally a Softmax classification layer to output the diagnostic results.

Images