Reciprocating compressor fault prediction method based on multi-source data analysis and PCAn-TSLG-LSTM model

Abstract

The application belongs to the field of compressors. The purpose is to propose a PCAn-TSLG-LSTM model for reciprocating compressor fault prediction. Including the following steps: (1) preprocessing the sensor collected data, and designing PCAn method to extract compressor health curve. (2) the overall model is subdivided into spatial domain and time domain, and the spatial domain is subdivided into local and global feature analysis, and an adaptive feature extraction mechanism is designed combined with LSTM network. (3) the input sequence is divided into four blocks, the internal effective information is extracted by applying the feature extraction mechanism, the local feature extraction results are combined, and then the global feature information is extracted by using the feature extraction mechanism, and the hidden layer state value of the LSTM network is output. (4) the global output values at multiple time steps are combined, the feature information at different time steps is extracted by applying the feature extraction mechanism, and the feature information is output. (5) input the information of time domain into the LSTM network, and predict the health curve value at next time. If it is not within the preset threshold range, it is determined to be a fault.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of reciprocating compressors, and particularly relates to a reciprocating compressor fault prediction method for offshore oil platforms. BACKGROUND

[0002] The petroleum and chemical industry is an important foundation and pillar of the national economy and plays a decisive role in the development of the macro economy. Reciprocating compressors are widely used in offshore oil drilling platforms due to their mature production process, simple structure, strong adaptability and other characteristics. However, due to the influence of complex wind and waves, tides and ocean currents on offshore platforms, reciprocating compressors are subjected to strong external environmental loads. Moreover, due to the space limitations of offshore oil platforms, compressor units are often arranged adjacent to each other, which can cause vibration interference between the compressor units. At the same time, there are many internal excitation sources in reciprocating compressors, which makes the offshore platform compressors more prone to failure. Therefore, ensuring the normal operation of reciprocating compressors and predicting faults in advance can greatly protect the interests of enterprises and ensure the safety of platform workers' lives and property.

[0003] The compressor fault prediction method can be divided into two categories: vibration signal analysis method and multivariate data analysis method. However, due to the vibration interference between the compressors on the offshore platform and the internal excitation sources of the compressors, the vibration signal has the characteristics of nonlinearity, non-stationarity and modal aliasing. It is difficult to extract effective information from the vibration signal alone, and it is not suitable for practical application on offshore oil platforms. SUMMARY

[0004] In order to solve the problem that the prior art cannot accurately predict the fault of the reciprocating compressor on the oil platform, the application proposes a reciprocating compressor fault prediction method based on multi-source data information and a PCAn-TSLG-LSTM model, which maximizes the use of compressor sensor data, uses the proposed PCAn method to construct a compressor health curve, and then accurately extracts compressor health information from the time and space, local and global angles, greatly improving the accuracy of fault prediction.

[0005] The specific steps of the reciprocating compressor fault prediction method are as follows:

[0006] 1. Sampling the state information of different structures inside the reciprocating compressor, such as the gas valve, main motor, bearing, scrubber, etc.

[0007] 2. Preprocessing the collected multi-dimensional data, including removing abnormal points and filling missing values, and obtaining 32-dimensional input sequence parameters of the model after processing.

[0008] 3. Use the proposed PCAn method to extract the curve with the best correlation as the health curve of the compressor during operation. The PCAn method uses principal component analysis (PCA) to extract features from the input sequence, and then calculates the health curve based on the feature extraction sequence and its correlation coefficient with the original data. The specific calculation formula is as follows:

[0009] (a1, a2, ..., a n )=softmax(β1β2,...,β n )

[0010]

[0011] Among them, β n is the correlation coefficient of the nth output sequence after principal component analysis. The first n correlation coefficients are calculated by the softmax function and normalized to the interval [0, 1] to obtain the corresponding weight coefficient a n . The first n weight coefficients and their corresponding correlation sequence l i After multiplication and addition, we get the final health curve L.

[0012] 4. Based on the information of different structures of the input sequence in the compressor, the 32 sequences are subdivided into four sub-modules: motor module, valve module, scrubber module and bearing module.

[0013] 5. Design an adaptive feature extraction mechanism based on TSLG-LSTM (Time-Space Learning Grouped-LSTM) spatiotemporal feature combination learning in conjunction with the LSTM network, and apply it to the local, global, and spatial domain feature analysis of the model. The adaptive feature extraction method is as follows:

[0014] 1) For the kth input sequence The hidden layer output h of the LSTM network at the previous moment can be t-1 , and cell state s t-1 Calculate the weight of the sequence k at this moment together:

[0015]

[0016] Among them, W and U are the weight matrix and bias matrix that need to be trained and learned, tanh is the tanh nonlinear activation function, n is the number of all motor sequences, [h t-1 ;s t-1 ;d kis the concatenation operation of hidden layer state, cell unit and input sequence. The network state and the correlation of different input sequences are calculated by combining the LSTM unit state and output at the last time and each input sequence at this time through the above formula. Since the correlation of different sequences and target sequences is different, the corresponding weight of each input sequence is also different through the softmax layer, which realizes the extraction of effective information of multiple input sequences.

[0017] After weighting, the new output sequence is as follows:

[0018]

[0019] 2) According to the weighted sequence, the hidden layer output of the LSTM network can be updated as:

[0020]

[0021] Where f is the LSTM network. The network parameters of the LSTM are updated through the hidden layer output at the last time and the input at this time.

[0022] 6. For the data of motor module, air valve module, washer module and bearing module, the above adaptive feature extraction method is used to assign weights to the internal sequences of each module at the same time step, realizing the extraction of internal features of each module.

[0023] 7. After the feature extraction of each sub-module, the total sequence of each weighted sequence is obtained according to the following formula

[0024]

[0025] Where, represent the parameters of the air valve sequence, washer sequence and bearing sequence at time t after weighting, respectively, and [;] represents the concatenation operation.

[0026] 8. At time t, the total sequence is subjected to the above adaptive feature extraction, and different feature weights are assigned to the internal sequences of the total sequence, realizing the purpose of global feature extraction and extracting the most relevant data information from the total sequence.

[0027] 9. At multiple time steps (1,..., t,..., T-1), the proposed adaptive feature extraction method is used again to assign different weights to the data at different times, and the weight assigned to the time step that has a greater impact on the health curve is larger, realizing the extraction of compressor health information on the time scale.

[0028] 10. Add different weighted information at different times to obtain the final prediction input item c tThe time domain LSTM network hidden layer output d' at T-1 time T-1 The health curve data at the last T time is calculated by an LSTM network in combination:

[0029]

[0030] Where [d' T ; c T ] is the concatenation operation of the time domain LSTM hidden layer output and the prediction input item c t . f is the LSTM network used to calculate the final output

[0031] 11. The proposed network is trained and tested using multiple sets of fault data and health data, and the threshold of the health curve is determined through experiments. If the health curve changes above or below the threshold, it is determined that a fault occurs. Through the above model design, the purpose of reciprocating compressor fault prediction and diagnosis is achieved. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 Reciprocating compressor motor related sequence

[0033] Figure 2 PCAn method extracted compressor health curve

[0034] Figure 3 Model overall architecture

[0035] Figure 4 TSLG-LSTM prediction model framework

[0036] Figure 5 Model batch error

[0037] Figure 6 Model iteration error

[0038] Figure 7 Normal working model training test results

[0039] Figure 8 Fault occurrence model training test results

[0040] Figure 9 Model fault prediction results, threshold =-0.75 DETAILED DESCRIPTION

[0041] The application will be further described below in combination with specific examples and drawings.

[0042] The reciprocating compressor fault prediction method based on multi-source data analysis and PCAn-TSLG-LSTM model. It includes the following steps:

[0043] First step: data processing

[0044] 1.1 Collect data sets by field sensors, and determine 4000 groups of 32-dimensional data after screening. The sampling time is 1 min. The data set includes different measurement point vibration signal frequency domain characteristics, different position temperature, pressure, flow and other measurement point variables, and the data is processed for abnormal value elimination and missing value processing.

[0045] 1.2 Classify the pre-processed data into motor, gas valve, bearing and scrubber module data, and divide the time step into 10, with an actual time span of 10 min. Among them, the motor related sequence is as shown in Figure 1 .

[0046] Second step: health curve

[0047] For the pre-processed data, the PCAn method is used to extract the health curve of the reciprocating compressor. The health curve is used as the overall target sequence of the model. The health curve is as shown in Figure 2 .

[0048] Third step: establishment of prediction model

[0049] 3.1 Design an adaptive feature extraction mechanism based on LSTM network, and apply it to the feature information extraction in the following three stages.

[0050] 3.2 For the motor, gas valve, bearing and scrubber module, each module has an LSTM model, represented by “M1, M2, M3, M4” respectively. Note that the hidden layer nodes of each LSTM of the overall model are set to 64. The data of each module at time t and the LSTM hidden layer output and cell unit state at the previous time are used as input, and the adaptive feature extraction mechanism is used to extract the effective information of each input sequence at time t and output weighted, while updating the M1, M2, M3, M4 network information.

[0051] 3.3 For the global feature extraction stage, the weighted output of the four modules is concatenated as the input sequence of global feature extraction. This stage also has an LSTM model, represented by “M5”. The weighted input sequence at time t and the LSTM hidden layer output and cell unit state at the previous time are used as input, and the effective information of each input sequence at time t is extracted and output weighted, while updating the M5 network information.

[0052] 3.4 For the time domain feature extraction stage, the first 9 time steps of global feature output are used to assign feature weights. This stage is represented by "M6" LSTM model, which extracts the effective information of each input sequence at 9 time points and weights the output, and then accumulates it as the final prediction input item c. Finally, the health curve value at time T is predicted using the health curve value at time T-1 and the prediction input item c, as well as the hidden layer output of M6. The overall framework of the model is shown in Figure 3 .

[0053] Step 4: Model Effect

[0054] After practical training and testing, the error change of the model in each batch is shown in Figure 4 , and the overall error change is shown in Figure 5 . Figure 6 and Figure 7 are the prediction results of the model under normal working condition and fault occurrence condition respectively. Through the prediction results, it can be seen that the model can accurately predict the occurrence of faults, and the prediction accuracy is very high.

Claims

1. A reciprocating compressor fault prediction method based on multivariate information and PCAn-TSLG-LSTM model, characterized by: A PCAn method was designed to extract the compressor operating status health curve from multi-source data. Then, an adaptive feature extraction method combined with the LSTM network was designed. The model was divided into three parts: spatial local feature extraction, spatial global feature extraction, and time domain feature extraction. The proposed feature extraction method was applied to the above three parts, and the feature extraction results were used for health curve prediction, realizing fault warning of reciprocating compressors.

2. A reciprocating compressor fault prediction method based on multivariate information and PCAn-TSLG-LSTM model according to claim 1, characterized in that: The PCAn method was designed to extract the compressor health curve from the screened multidimensional input data. Using 32-dimensional data, PCA calculation was performed to obtain correlation sequences with different correlation coefficients. The top n most correlated sequences were selected, and their correlation coefficients were normalized to between 0 and 1 using the softmax function. The correlation coefficients were then multiplied and accumulated with the correlation sequences to obtain the final compressor health curve.

3. The reciprocating compressor fault prediction method based on multivariate information and PCAn-TSLG-LSTM model according to claim 1, characterized in that: Based on the design of the adaptive feature extraction method of the LSTM network, combined with the input multi-sequence information, the correlation between the input sequence and the LSTM network hidden layer output and cell unit state parameters is calculated, and this correlation coefficient is converted to between 0 and 1 through the softmax layer, thereby assigning different weights to different input sequences, realizing the selection of the most effective features of multiple sequences. At the same time, the weighted sequence and the LSTM network hidden layer output at the previous moment are used to update the LSTM unit information at this moment.

4. A reciprocating compressor fault prediction method based on multivariate information and PCAn-TSLG-LSTM model according to claim 1, characterized in that: By using the proposed adaptive feature extraction method, every part of the model is fully utilized; 1) For a 32-dimensional input sequence, the system is divided into four submodules according to the structure of the reciprocating compressor: the motor module, the valve module, the scrubber module, and the bearing module. For the internal sequence of each submodule, the proposed adaptive feature extraction method is used to assign different weights to the internal sequence at time t, thus achieving effective information extraction from the local submodule. 2) After extracting local effective information, the weighted outputs of the four submodules are connected in series, and global adaptive feature extraction is performed on the series data. Different weights are assigned to global variables, and higher weights are assigned to sequences most relevant to the health curve, thus achieving global feature information extraction in the spatial domain. 3) For the global information in the spatial domain at different time steps, since the data at different times have different effects on the health curve at the target time, they are integrated and adaptive feature extraction is performed on the time step. After feature extraction in the time domain, the final effective spatial and temporal information is obtained, which is then used as the input of the LSTM network of the reciprocating compressor health curve to predict the compressor health information at the target time.

Images