A Visual Method for Fault Diagnosis of Reciprocating Compressors

Abstract

The present invention provides a visualization method for reciprocating compressor fault diagnosis, including S1 generation and accumulation of training samples; S2 batch acquisition of training data and labels; S3 AI model design and initialization; S4 training to obtain the AI model; S5 selecting a data acquisition device that meets the requirements, synchronously collecting dynamic pressure data and rotational speed pulse data during operation, preprocessing the collected data to obtain test samples, inputting the obtained test samples into the AI model, forward calculating to obtain the network output, and taking the most likely category corresponding to the output as the recognition result; S6 using Guided Grad CAM to obtain the contribution degree of each value of the test sample to the recognition result of the AI model; S7 displaying an indicator diagram that can explain the diagnosis result; S8 manually confirming the accuracy of the model recognition result and processing it; S9 storing the result in the training sample library. The present invention can give the basis for the model to make a judgment without affecting the accuracy of the intelligent diagnosis model, improve the credibility of intelligent diagnosis, and facilitate manual verification of the diagnosis result.

Description

Technical Field

[0001] The present invention belongs to the technical field of mechanical fault diagnosis, and particularly relates to a visualization method for reciprocating compressor fault diagnosis. Background Art

[0002] The body of a reciprocating compressor consists of a cylinder block and a crankcase. The important working component inside the cylinder block is the piston, and the important working component inside the crankcase is the crankshaft. The piston and the crankshaft are connected by a connecting rod and a crosshead. An intake valve and an exhaust valve are installed at the top of the cylinder. The intake valve is connected to the intake pipe through an intake chamber, and the exhaust valve is connected to the exhaust pipe through an exhaust chamber. The main shaft of the crankshaft is connected to a prime mover and is driven by the prime mover to rotate. Through the transmission of the connecting rod, the piston is driven to reciprocate inside the cylinder block, and with the cooperation of the intake and exhaust valves, the intake, compression, and discharge of the medium are completed.

[0003] A reciprocating compressor has four important working processes: compression, exhaust, expansion, and intake. As shown in Figure 1 , when the piston moves from right to left, the volume of the gas inside the cylinder decreases and the pressure increases, which is the "compression" process of the compressor; when the gas pressure increases to a certain extent, the exhaust valve opens, and the piston continues to move leftward, and the gas continues to be discharged until the piston moves to the leftmost position, which is the "exhaust" process of the compressor; then, the piston moves from the leftmost position to the right. Since there is a clearance volume in the cylinder at the leftmost position, the gas in the clearance volume still has a relatively high pressure. When the piston just leaves the leftmost position, the intake valve cannot be immediately opened due to the insufficient pressure difference. The volume of the gas in the clearance volume increases continuously as the piston moves rightward, and the pressure decreases continuously. This is the "expansion" process of the compressor; the piston continues to move rightward, and when the intake valve reaches the required pressure difference, it starts to intake air until the piston moves to the rightmost position of the cylinder and the intake is completed. This is the "intake" process of the compressor. The working process of a reciprocating compressor is a cycle of these four processes.

[0004] During the above four processes, the pressure and volume of the gas inside the cylinder are constantly changing. An indicator diagram refers to a closed curve in which the dynamic pressure of the gas inside the cylinder changes with the piston displacement (or the volume inside the cylinder) during one cycle of a reciprocating compressor. Figure 2The figure is the indicator diagram of a reciprocating compressor under ideal conditions. The indicator diagram is an important tool for the fault diagnosis of reciprocating compressors. Through the indicator diagram, serious faults that may cause explosion accidents, such as air valve leakage and air valve blockage, can be warned more sensitively and earlier. At present, there are two commonly used diagnostic methods: one is manual diagnosis, where the indicator diagram within the normal time range is mapped as a benchmark, and the real-time mapped indicator diagram is obtained manually at regular intervals and compared with the benchmark for diagnosis based on experience; the other is to use the method of machine learning, where a large amount of dynamic pressure data of reciprocating compressors with known fault types (dynamic pressure data is the necessary data for drawing the indicator diagram) is used as the original data, and the features that can represent the indicator diagram are extracted as training samples to train an intelligent recognition model based on the indicator diagram for diagnosis.

[0005] However, diagnosing the indicator diagram of a reciprocating compressor manually based on the fault mechanism is not intelligent enough, relying too much on human experience, and it is difficult to guarantee real-time performance. And using general machine learning means for fault identification cannot give the basis for diagnosis, lacks credibility, and requires a large amount of manual review work. Summary of the Invention

[0006] To solve the deficiencies of the prior art, the present invention provides a method for visualizing the faults of a reciprocating compressor, which can give the basis for the model's judgment and display it on the indicator diagram without affecting the accuracy of the intelligent diagnosis model.

[0007] The object of the present invention is achieved through the following technical solutions:

[0008] A method for visualizing the fault diagnosis of a reciprocating compressor, characterized by including the following steps:

[0009] S1. Generation and accumulation of training samples: The data collected by the dynamic pressure sensor installed in the cylinder and the pulse signal data synchronously collected by the key phase sensor installed on the edge of the flywheel are collected as training samples;

[0010] S2. Batch acquisition of the training data and labels generated in S1;

[0011] S3. Design and initialization of the AI model: A neural network model is adopted, the input of the network is a one-dimensional array with N points in a single channel, and the output is the label encoding corresponding to the health status, including normal, fault type 1, fault type 2, etc.; the neural network model includes but is not limited to ANN, ResNet, VGG, LSTM;

[0012] S4. Train and obtain the AI model;

[0013] S5. Select a data acquisition device that meets the requirements, synchronously collect dynamic pressure data and speed pulse data during operation, preprocess the collected data to obtain test samples, and input the obtained test samples into the AI model obtained in S4. Forward calculation is performed to obtain the network output, and the maximum possible category corresponding to the output is used as the recognition result;

[0014] S6. Use Guided Grad CAM to obtain the contribution of each value of the test sample in S5 to the AI model recognition result;

[0015] S7. Display the dynamometer diagram that can be interpreted for the diagnostic structure;

[0016] S8. Manually confirm the accuracy of the model recognition results and perform corresponding processing based on actual conditions;

[0017] S9. Storing the processing result of S8 in a training sample library to improve the accuracy of subsequent retraining.

[0018] Preferably, the S1 includes the following steps:

[0019] S11. Before data acquisition, the data acquisition device is set so that the rising edge of the speed pulse corresponds to the starting point of the expansion process, the same and fixed sampling frequency Fs is set for the sensor generating the training sample, and the dynamic pressure data and the speed pulse data are synchronously acquired;

[0020] S12. Take the synchronously collected dynamic pressure data corresponding to the data between the rising edges of two adjacent speed pulses as the dynamic pressure data within one cycle, and record the number of data points as N. If the data collected once contains complete dynamic pressure data for m consecutive cycles, calculate the average of the pressure values with the same index in each cycle as the effective dynamic pressure value of the index, and then obtain the effective dynamic pressure data with N points after averaging the data of each cycle;

[0021] S13, assuming that the crankshaft moves at a uniform speed, perform approximate processing on the N valid dynamic pressure data in S12, calculate the crank phase angle corresponding to each dynamic pressure value, and draw an indicator diagram based on information such as piston radius, crank length, and connecting rod length;

[0022] S14. Manually confirm the health status of the dynamometer diagram, including the fault type, create labels for the dynamic pressure data with N points in S12, and store the N dynamic pressure data together with the labels in a training sample library.

[0023] Preferably, the preprocessing step in S5 is the same as S12.

[0024] Preferably, the S6 comprises the following steps:

[0025] S61. During the forward calculation of the network in S5, the output result of the last convolutional layer in the AI model is used as the feature-map. The partial derivative of the score of the most likely class output by the network with respect to each value in the feature-map layer is calculated, and then the global average pooling (GAP) operation is used to obtain the contribution degree of each channel to the classification result, which is denoted as the channel weight. The number of the channel weights depends on the design of the AI model in S3.

[0026] S62. Multiply the above channel weights by the corresponding feature-map and then input it into the activation function. Finally, the weight array represented by Grad CAM is obtained. This weight array can be regarded as a one-dimensional heatmap "bar".

[0027] S63. Use the Guided Backpropagation algorithm to obtain the guided weight array, which has the same dimension as the weight array obtained in S62.

[0028] S64. Multiply the weights at the corresponding positions in S62 and S63 as the importance degree weights of each data point at the corresponding position in the test sample obtained by preprocessing in S5.

[0029] Preferably, S7 includes the following steps. Combine the method of drawing the indicator diagram mentioned in S13 to draw the indicator diagram of the test sample, and highlight the connection line where the dynamic pressure value points with larger importance degree weights in S6 are located as the "reason explanation" for the recognition result obtained by the model. The highlighting is used to distinguish the importance degree, and it can be in the form of the color of the heatmap, the thickness of the line, or the transparency of the line.

[0030] Preferably, during the manual confirmation in S8, the Guided Grad CAM will prompt the parts of the indicator diagram that need to be focused on for confirmation on the indicator diagram. If it is confirmed as a fault, the fault will be processed.

[0031] Preferably, S9 is to use the final result verified manually as a label and store it together with the samples of this diagnosis in the training sample library to improve the accuracy through subsequent retraining.

[0032] Preferably, in S11, by adjusting the position of the key phase block on the flywheel, when the piston moves to the limit position, the normal vector of the plane where the key phase block is located is on the same axis as the axis of the key phase sensor probe so that the rising edge of the rotational speed pulse corresponds to the starting point of the expansion process.

[0033] Preferably, the limit position in S11 refers to the end of the compression process and the start of the expansion process.

[0034] Preferably, if the number of data points within a period in S12 is not N, an interpolation fitting method is used to process it to be N. Here, m is a positive integer not less than 1, and the index range is 0 to N - 1.

[0035] The beneficial effects of the present invention are as follows: Without affecting the accuracy of the intelligent diagnosis model, this method can provide the basis for the model to make a judgment, improve the credibility of intelligent diagnosis, and facilitate manual verification of the diagnosis results. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 : The local structure of a reciprocating compressor in the prior art.

[0037] Figure 2 : The indicator diagram of a reciprocating compressor under ideal conditions

[0038] Figure 3 : The schematic diagram of the fault visualization process of the present invention.

[0039] Figure 4 : The indicator diagram obtained by combining the Guided Grad CAM method in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0040] The present invention discloses a method for visualizing the fault diagnosis of a reciprocating compressor, which combines Figure 3 As shown below, it includes the following steps:

[0041] S1. Generation and accumulation of training samples: The data collected by the dynamic pressure sensor installed in the cylinder and the pulse signal data synchronously collected by the key-phase sensor installed on the edge of the flywheel are collected as training samples.

[0042] S2. Batch acquisition of the training data and labels generated in S1;

[0043] S3. Design and initialization of the AI model: A neural network model is adopted. The input of the network is a one-dimensional array with N points in a single channel, and the output is the label encoding corresponding to the health status, including normal, fault type 1, fault type 2, etc. The neural network model includes but is not limited to ANN, ResNet, VGG, and LSTM.

[0044] S4. Train and obtain the AI model;

[0045] S5. Select a data acquisition device that meets the requirements, synchronously collect the dynamic pressure data and rotational speed pulse data during operation, preprocess the collected data to obtain test samples, and input the obtained test samples into the AI model obtained in S4. Through forward calculation, the output of the network is obtained, and the category corresponding to the maximum possibility of the output is used as the recognition result.

[0046] S6. Use Guided Grad CAM to obtain the contribution degree of each value of the test sample in S5 to the recognition result of the AI model;

[0047] S7. Display the indicator diagram that can explain the diagnosis structure;

[0048] S8. Manually confirm the accuracy of the model recognition result and perform corresponding processing in combination with the actual situation;

[0049] S9. Store the processing result of S8 into the training sample library to improve the accuracy of subsequent retraining.

[0050] The following elaborates on the present invention in detail with specific embodiments for easy understanding.

[0051] In this embodiment, for a number of first-stage reciprocating compressors of a certain existing model, a data acquisition device has been installed according to the requirements of the present invention, and the training sample library has accumulated a sufficient amount of samples in the normal state and several fault states, and has been labeled according to the requirements.

[0052] Each training sample averages the dynamic pressure data of multiple cycles of the reciprocating compressor operating at rated power, and the number of data points of the dynamic pressure data per cycle is 456.

[0053] Batch obtain multiple training samples and labels from the training sample library;

[0054] Design the network model as ResNet18. The input of the network is a one-dimensional array with a single channel and 456 data points, and the output is the one-hot label corresponding to the health status.

[0055] Train the ResNet18 network model until the network loss, training accuracy, and validation accuracy meet the expectations, complete the training, and obtain the trained network model.

[0056] Collect the dynamic pressure data and rotational speed pulse data of the compressor to be detected at rated power, preprocess the data according to the method proposed in this patent, and obtain the average dynamic pressure data of multiple cycles of the equipment operation as the sample to be detected.

[0057] Input the sample to be detected into the trained network model to obtain the recognition result.

[0058] Adopt the Guided Grad CAM algorithm to obtain the influence of each data point of the dynamic pressure data on the model recognition result, that is, the importance degree of each data point.

[0059] Combine the processing result of the previous step with the sample to be detected, and use the method proposed in the present invention to draw the indicator diagram of the reciprocating compressor, as shown in Figure 4As shown, the data points with higher importance from the previous step are highlighted. After manual verification of the diagnosis results, it is determined that the exhaust valve is not tight. The data from this diagnosis and the confirmed label are added to the sample library.

[0060] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A visualization method for reciprocating compressor fault diagnosis, characterized in that It includes the following steps: S1. Generation and accumulation of training samples: Collect the data collected by the dynamic pressure sensor installed in the cylinder and the pulse signal data synchronously collected by the key-phase sensor installed on the edge of the flywheel as training samples; S2. Batch acquisition of the training data and labels generated in S1; S3. Design and initialization of the AI model: Adopt a neural network model. The input of the network is a one-dimensional array with N points in a single channel, and the output is the label encoding corresponding to the health status. The neural network model includes ANN, ResNet, VGG, and LSTM; S4. Train and obtain the AI model; S5. Select a data acquisition device that meets the requirements, synchronously collect the dynamic pressure data and rotational speed pulse data during operation, preprocess the collected data to obtain test samples, and input the obtained test samples into the AI model obtained in S4. Perform forward calculation to obtain the output of the network, and take the most likely category corresponding to the output as the recognition result; S6. Use Guided Grad CAM to obtain the contribution degree of each value of the test samples in S5 to the recognition result of the AI model; S7. Display the indicator diagram that can explain the diagnosis result; S8. Manually confirm the accuracy of the model recognition result and perform corresponding processing in combination with the actual situation; S9. Store the processing result of S8 into the training sample library to improve the accuracy of subsequent retraining; For the above S1, it includes the following steps: S11. Set the data acquisition device before data collection so that the rising edge of the rotational speed pulse corresponds to the starting point of the expansion process. Set the same and fixed sampling frequency Fs for the sensors generating the training samples, and synchronously collect the dynamic pressure data and rotational speed pulse data; S12. Take the synchronously collected dynamic pressure data corresponding to the data between two adjacent rising edges of the rotational speed pulse as the dynamic pressure data within one cycle. Record the number of data points as N. If the data collected at one time contains the complete continuous m cycles of dynamic pressure data, calculate the average value of the pressure values with the same index in each cycle as the effective dynamic pressure value of this index, and then obtain the effective dynamic pressure data with N points after averaging the data of each cycle; S13. Assume that the crankshaft rotates at a constant speed, approximately process the N effective dynamic pressure data in S12, calculate the crank phase angle corresponding to each dynamic pressure value, and draw an indicator diagram in combination with the piston radius, crank length, and connecting rod length information; S14. Manually confirm the health status of the indicator diagram, including the fault type, establish labels for the dynamic pressure data with N points in S12, and store the N dynamic pressure data together with the labels into the training sample library; The preprocessing steps in S5 are the same as those in S12; The above S6 includes the following steps: S61. During the forward calculation of the network in S5, the output result of the last convolutional layer in the AI model is used as the feature-map. The partial derivative of the score of the most likely category output by the network with respect to each value in the feature-map layer is calculated, and then the global average pooling operation is used to obtain the contribution degree of each channel to the classification result, which is denoted as the channel weight. The number of the channel weights depends on the design of the AI model in S3. S62. Multiply the above channel weights by the corresponding feature-map and then input them into the activation function. Finally, a weight array represented by GradCAM is obtained. This weight array can be regarded as a one-dimensional heatmap "bar". S63. Use the guided backpropagation algorithm to obtain a guided weight array, which has the same dimension as the weight array obtained in S62. S64. Multiply the weights at the corresponding positions in S62 and S63 as the importance degree weights of each data point at the corresponding position in the test sample obtained by preprocessing in S5.

2. The visualization method for reciprocating compressor fault diagnosis according to claim 1, characterized in that, S7 includes the following steps: Combine the method of drawing the indicator diagram mentioned in S13 to draw the indicator diagram of the test sample, and highlight the connection line where the dynamic pressure value points with larger importance degree weights in S6 are located as the "reason explanation" for the recognition result obtained by the model. The highlighting is used to distinguish the importance degree, which is in the form of the color of the heatmap, the thickness of the line, or the transparency of the line.

3. The visualization method for reciprocating compressor fault diagnosis according to claim 1, characterized in that, When performing manual confirmation in S8, the interpretable display of Guided Grad CAM on the indicator diagram will prompt the parts of the indicator diagram that need to be focused on for confirmation. If it is confirmed as a fault, the fault will be processed.

4. The visualization method for reciprocating compressor fault diagnosis according to claim 1, wherein, S9 is to use the final result verified manually as a label and store it together with the sample of this diagnosis in the training sample library for subsequent retraining to improve the accuracy.

5. The visualization method for reciprocating compressor fault diagnosis according to claim 1, wherein, In S11, by adjusting the position of the key phase block on the flywheel, when the piston moves to the extreme position, the normal vector of the plane where the key phase block is located is on the same axis as the axis of the key phase sensor probe so that the rising edge of the rotational speed pulse corresponds to the starting point of the expansion process.

6. The visualization method for reciprocating compressor fault diagnosis according to claim 5, wherein The extreme position in S11 refers to the end of the compression process and the start of the expansion process.

7. A reciprocating compressor fault diagnosis visualization method according to claim 6, characterized in that, In S12, if the number of data points within a period is not N, the interpolation fitting method is used to process it to make it N. Here, m is a positive integer not less than 1, and the index range is 0 to N - 1.

Images