Emulsion pump flow monitoring method and system based on recurrent neural network

Abstract

The present application relates to the field of electric digital data processing, and particularly relates to an emulsion pump flow monitoring method and system based on a recurrent neural network, the method comprising: obtaining historical operation data and working condition state labels of an emulsion pump station, including flow, pressure and power signals, and obtaining a flow sample segment after preprocessing; setting a mode number and a bandwidth search range, taking parameter combinations as genetic individuals, constructing a population and optimizing to obtain optimal parameter combinations of each sample, and constructing a historical database and a real-time parameter candidate set; accordingly, performing mode decomposition, constructing a comprehensive synergy index according to the correlation between pressure, power signals and IMF components, calculating a reconstruction weight, and reconstructing a flow signal; training an LSTM model using the reconstructed data and working condition labels, outputting a working condition probability, reconstructing a real-time flow signal and inputting the model to realize real-time online monitoring of the emulsion pump flow; and the present application improves the signal-to-noise ratio and abnormal recognition accuracy through adaptive denoising and optimized parameters.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of digital data processing. In particular, it relates to an emulsion pump flow monitoring method and system based on a recurrent neural network. BACKGROUND

[0002] The emulsion pump is the core power equipment of the fully mechanized coal face in the coal mine, responsible for providing stable high-pressure emulsion power for the hydraulic support, and the stability of its flow output is directly related to the support effect of the support and the safety of the mine production. Once the pump body has flow abnormalities (such as pipe blockage, leakage, etc.), it will quickly affect the response capability of the entire support system, and even induce major safety accidents such as support instability and roof collapse. Therefore, real-time and accurate flow monitoring of the emulsion pump is a key link to ensure safe and efficient production in the coal mine.

[0003] However, the underground environment is harsh, with strong electromagnetic interference and severe mechanical vibration, resulting in extremely low signal-to-noise ratio of the flow sensor signal. Especially in sudden changes of working conditions such as pump start-stop and support action, the real flow signal often contains key high-frequency sudden components, which seriously overlap with vibration noise in the frequency domain. Traditional fixed-parameter filtering methods are difficult to distinguish between noise and useful signals in the same frequency band: excessive filtering will smooth out the real working condition mutations, causing monitoring lag and false positives; insufficient filtering will make the noise drown out the signal, causing measurement errors and failing to meet the stringent requirements of high reliability in industrial sites.

[0004] At present, deep learning is introduced into flow monitoring, but most methods directly input the original noisy signal into the model, and there is a "distribution shift" between the training data and the real working conditions, and existing deep learning methods cannot effectively preprocess and feature extraction according to the physical characteristics of the emulsion pump and the noise environment, resulting in insufficient generalization ability of the model under complex working conditions, and difficulty in coping with the challenges of severe fluctuations, strong noise and frequent mutations in actual operation. SUMMARY

[0005] To solve the problem that existing deep learning methods do not effectively preprocess and feature extraction according to the characteristics of the emulsion pump, resulting in insufficient generalization ability under complex working conditions and difficulty in coping with severe fluctuations, strong noise and frequent mutations, the present application provides solutions in the following aspects.

[0006] In a first aspect, the emulsion pump flow monitoring method based on a recurrent neural network comprises: obtaining historical operation data and working condition state labels of an emulsion pump station, the historical operation data comprising: a flow signal, a pressure signal and a motor power signal, pre-processing the historical operation data to obtain a pre-processed sample segment; setting a range of modal number and bandwidth, combining the modal number and bandwidth parameters as a genetic individual, randomly generating an initial individual and constructing a population, and performing variational modal decomposition on each flow sample segment; according to the modal separation degree and the signal fidelity, evaluating the comprehensive optimization target value of each flow sample segment; based on the comprehensive optimization target value, iteratively optimizing, selecting the optimal parameter combination, constructing a historical database of mapping relationship, and based on the frequency of parameter combination, constructing a real-time parameter candidate set; based on the optimal parameter combination, performing modal decomposition to obtain a plurality of IMF components, determining the weight based on the correlation between the pressure signal and the motor power signal and each IMF component, and constructing a comprehensive synergy index; taking the proportion of the comprehensive synergy index of each flow sample segment as the reconstruction weight, weighting and reconstructing each IMF component of the optimal parameter combination, and completing the flow signal reconstruction; using the reconstructed flow signal data and the corresponding working condition state labels to train a preset LSTM model, outputting the probability value of each working condition state label, based on the parameter candidate set, selecting the optimal parameter combination for the real-time acquired flow signal data, completing the reconstruction and inputting into the trained LSTM model, outputting the probability of the corresponding label, and completing the real-time online monitoring of the emulsion pump flow.

[0007] Preferably, the modal separation degree calculation method comprises:

[0008] Taking any flow sample segment as a target segment, obtaining the IMF components after decomposition of the target segment, calculating the mutual information between each pair of IMF components, and accumulating and summing the mutual information between each pair of IMF components to obtain the modal separation degree of the target segment.

[0009] Preferably, the signal fidelity calculation method comprises:

[0010] Taking any flow sample segment as a target segment, summing all IMF components of the target segment to obtain a reconstructed signal, calculating the difference between the original signal of each sampling point in the target segment and the corresponding sampling point reconstructed signal, calculating the square error by squaring the difference, and accumulating the square error of all sampling points in the target segment to obtain the signal fidelity of the target segment.

[0011] Preferably, the comprehensive optimization target value calculation method comprises:

[0012] Taking any flow sample segment as a target segment, calculating the difference between adjacent sampling points in the target segment, and taking the average value of the cumulative difference of all adjacent sampling points as the fluctuation intensity of the target segment.

[0013] The wave intensity is exponentially mapped using a negative exponential function, the reciprocal of 1 plus the mapped result is taken as the modal separation weight, 1 is subtracted from the modal separation weight as the weight of signal fidelity, and weighted summation is carried out to obtain a comprehensive optimization target value.

[0014] Preferably, the calculation method of the comprehensive synergy index comprises:

[0015] Taking any IMF component as an analysis component, the Pearson correlation coefficients between the analysis component and the pressure sample segments of the pressure signals and the Pearson correlation coefficients between the analysis component and the motor power sample segments of the motor power signals are calculated, and the sample segments of the pressure signals and the sample segments of the motor power signals are weighted and fused based on preset weights to obtain the comprehensive synergy index.

[0016] Preferably, the constructing of the real-time parameter candidate set comprises:

[0017] The occurrence frequencies of each parameter combination in the historical database are counted, the occurrence frequencies are sorted in descending order, and a preset number of parameter combinations are constructed into the real-time parameter candidate set, wherein the historical database comprises: a flow sample segment, a wave intensity, an optimal modal number and an optimal penalty factor.

[0018] Preferably, the working condition state label comprises: normal, pipe blockage and leakage.

[0019] Preferably, the preprocessing of the historical operation data comprises:

[0020] Based on a preset sliding window length and a preset sliding step, sample segments within the preset window length are obtained, 3σ criterion method is used for each sample segment to eliminate abnormal points, and linear interpolation method is used to fill in missing data.

[0021] In a second aspect, the emulsion pump flow monitoring system based on a recurrent neural network comprises a processor and a memory, and the memory stores computer program instructions.

[0022] The present application has the following effects:

[0023] 1、The present application effectively strips strong noise interference and retains the true working condition characteristics through adaptive variational mode decomposition and multi-sensor collaborative denoising.

[0024] 2、The application optimizes the parameters of variational modal decomposition by using a genetic algorithm, adapts to flow signals of different fluctuation intensities, avoids under-decomposition or over-decomposition problems caused by fixed parameters, and improves signal processing accuracy. The candidate set is generated by frequency statistics of historical parameters, and the real-time efficiency and parameter optimization accuracy of online monitoring are considered. Combined with the LSTM network, the long-term time sequence dependence of the flow signal is accurately captured, and the slowly changing or sudden abnormality has high recognition accuracy and low false alarm rate. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 is a method flowchart of steps S1-S5 in the emulsion pump flow monitoring method based on the recurrent neural network according to the embodiment of the application.

[0026] Figure 2 is a structural block diagram of the emulsion pump flow monitoring system based on the recurrent neural network according to the embodiment of the application. DETAILED DESCRIPTION

[0027] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are part of the embodiments of the application, rather than all the embodiments of the application.

[0028] REFERENCE Figure 1 The emulsion pump flow monitoring method based on the recurrent neural network includes steps S1-S5, and specifically as follows:

[0029] S1: Obtain historical running data and working condition labels of the emulsion pump station, and the historical running data includes flow signals, pressure signals and motor power signals. The historical running data is preprocessed to obtain a preprocessed sample segment.

[0030] In one embodiment, a high-frequency data acquisition system is used to synchronously acquire the historical running data of the emulsion pump station at a sampling frequency of 10 times per second. The acquired data specifically includes the emulsion pump total pipeline outlet flow signal, the pump outlet pressure signal and the motor power signal. The acquisition process comprehensively covers data under all typical working condition states of the pump station, such as normal, pipe blockage and leakage.

[0031] The acquired continuous long-time sequence multivariate data is synchronously divided and processed by using a sliding window method. Specifically, the window length is set to 10 seconds, and the sliding step is 1 second, so that the total number of sampling points in a single window is 100. Through the above division method, a total of multivariate samples are finally obtained. For the th sample (wherein ), the flow sample segment contained therein can be expressed as , the pressure sample segment can be expressed as , and the power sample segment can be expressed as wherein represents the total number of sampling points contained in a single sample.

[0032] Subsequently, for each sample segment, a data preprocessing operation is performed. Specifically, the 3σ criterion method is used to eliminate obvious outliers, and the linear interpolation method is used to complete a small amount of missing data to ensure the integrity and accuracy of the data.

[0033] In the strong noise environment downhole, the signal-to-noise ratio of the emulsion pump flow signal is extremely low, and due to the variable working conditions, the signal characteristics are unstable. The traditional fixed parameter filtering method is difficult to achieve good results in processing stationary and sudden conditions. As an adaptive signal decomposition method, variational mode decomposition can effectively process non-stationary signals, but its decomposition effect is severely dependent on the mode number and the penalty factor These two key parameters. In order to realize adaptive decomposition, this paper uses genetic algorithm to automatically optimize the parameter combination. The specific steps are as follows:

[0034] S2: Set the range of mode number and bandwidth, take the parameter combination of mode number and bandwidth as genetic individual, randomly generate initial individual and construct population, and perform variational mode decomposition on each flow sample segment. According to the mode separation degree and signal fidelity, evaluate the comprehensive optimization target value of each flow sample segment.

[0035] In one embodiment, the range of mode number is set to , and the bandwidth search range is Discretize search with a step size of 100 to balance the mode bandwidth constraint and reconstruction accuracy. Specifically, it can be adjusted according to specific circumstances.

[0036] Take any flow sample segment as the target segment, obtain the IMF (Intrinsic Mode Function) components after decomposition of the target segment, calculate the mutual information between each pair of IMF components, and accumulate the mutual information between each pair of IMF components to obtain the mode separation degree of the target segment.

[0037] Specifically, the mode separation degree satisfies the following relationship:

[0038] ;

[0039] wherein, represents the mode separation degree of the th flow sample segment, that is, the separation degree of each intrinsic mode function after decomposition, the smaller the value, the lower the statistical correlation between each modal component, the better the separation effect, and the effective avoidance of modal aliasing of noise and useful signal, represents the mode separation degree of the the number of modes of the variational modal decomposition of the flow sample segment, denotes the th IMF component, denotes the th intrinsic modal component, denotes the th intrinsic modal component, denotes the mutual information between . The modal separation degree is normalized by maximum and minimum value, and the value range of is limited between .

[0040] The mutual information is calculated to measure the statistical correlation between two random variables. In the context of variational modal decomposition, mutual information can be used to evaluate the correlation between different intrinsic modal components. The larger the value of mutual information, the higher the correlation between the two modal components. Therefore, by calculating the mutual information between all modal components, the separation degree between modal components can be quantified.

[0041] Taking any flow sample segment as the target segment, summing all IMF components of the target segment to obtain the reconstructed signal, calculating the difference between the original signal of each sampling point in the target segment and the corresponding sampling point reconstructed signal, calculating the square error by squaring the difference, and accumulating the square error of all sampling points in the target segment to obtain the signal fidelity of the target segment.

[0042] Specifically, the signal fidelity satisfies the following relationship:

[0043] ;

[0044] In the formula, denotes the signal fidelity of the th flow sample segment, initialized to 0, used to evaluate the degree of coincidence between the reconstructed signal of the modal component and the original signal, the smaller the value, the higher the fidelity of the reconstructed signal to the original signal, effectively avoiding the loss of effective signal components, denotes the original signal of the preprocessed th flow sample segment, denotes the number of modes of the variational modal decomposition of the flow sample segment, denotes the th IMF component, denotes the th intrinsic modal component, is the reconstructed signal of all modal components, ​This represents the mean square error function. The signal fidelity is normalized using an inverse maximum-minimum method, which... The range of values ​​is limited to Between these values, the minimum value of the reconstruction error is mapped to 1, and the maximum value is mapped to 0, so that samples with higher fidelity (i.e. smaller error) get higher normalization values.

[0045] Taking any flow sample segment as the target segment, calculate the difference between adjacent sampling points within the target segment, and take the average of the sum of the differences between all adjacent sampling points as the fluctuation intensity of the target segment.

[0046] The fluctuation intensity is mapped exponentially using a negative exponential function. The reciprocal of the mapping result is added to 1 and used as the mode separation weight. The mode separation weight is subtracted from 1 and used as the signal fidelity weight. The weighted sum is then calculated to obtain the comprehensive optimization target value.

[0047] The step of using a negative exponential function to perform exponential mapping on the fluctuation intensity, and then adding 1 to the mapping result and taking the reciprocal, is a normalization method using the Sigmoid function to map the input value to... Within the range, it exhibits smooth transition characteristics. As the fluctuation intensity increases, it indicates greater fluctuation in the flow signal, raising the risk of noise and useful signal aliasing. At this point, the mode separation weight gradually approaches 1, tilting towards mode separation and preferentially stripping away interference to ensure effective separation of noise and useful signal even with large fluctuations.

[0048] Specifically, the wave intensity satisfies the following relationship:

[0049] ;

[0050] In the formula, Indicates the first The fluctuation intensity of each flow sample segment A larger value indicates more drastic fluctuations in the traffic sample segment, while a smaller value indicates a more stable traffic signal. This indicates the total number of sampling points contained in each flow sample segment. Indicates the sequence number of the traffic sample segment. Indicates the first In the traffic sample segment, the first The instantaneous detection value at each sampling time. The sampling point number has a value range of 1. .

[0051] S3: Iterative optimization is performed based on the comprehensive optimization objective value to select the optimal parameter combination, a historical database of mapping relationships is constructed, and a real-time parameter candidate set is constructed based on the frequency of occurrence of parameter combinations.

[0052] Specifically, an individual is selected based on a comprehensive optimization target value using a roulette wheel selection method, wherein the higher the comprehensive optimization target value, the higher the probability of being selected. The individual with a high comprehensive optimization target value is retained by the roulette wheel selection method to form a new generation population. Through iterative crossover operation and mutation operation, the individual with the maximum comprehensive optimization target value is selected from the population, which is the optimal parameter combination.

[0053] The crossover operation and mutation operation are well-known techniques in the art and will not be described in detail.

[0054] The optimal parameter combination , the fluctuation intensity , and the flow sample segment are stored in the historical database.

[0055] The frequency of each parameter combination in the historical database is counted, and the frequencies are sorted in descending order. A real-time parameter candidate set is constructed from the parameter combinations with the highest frequencies. The historical database includes flow sample segments, fluctuation intensity, optimal modal number, and optimal penalty factor.

[0056] Furthermore, the top 20% of the parameter combinations are selected to construct the real-time parameter candidate set. In real-time monitoring, searching for the optimal parameter combination directly from all possible parameter combinations would require a large amount of computation, especially in high-frequency data acquisition and real-time processing scenarios. By constructing a real-time parameter candidate set, the parameter search range is limited to the top 20% of the parameter combinations that perform best in historical data, significantly reducing computation time and improving real-time processing speed. This improves the efficiency and accuracy of real-time monitoring while reducing the consumption of computing resources.

[0057] S4: Perform modal decomposition based on the optimal parameter combination to obtain several IMF components. Determine the weight based on the correlation between the pressure signal and motor power signal and each IMF component, and construct a comprehensive synergy index. Use the proportion of the comprehensive synergy index of each flow sample segment as the reconstruction weight to weight and reconstruct the IMF components of the optimal parameter combination, and complete the flow signal reconstruction.

[0058] It should be noted that modal decomposition is a well-known technique in the art and will not be described in detail. In addition, it should be noted that there is an inherent physical relationship between the flow signal, pump outlet pressure signal, and motor power signal of the emulsion pump. An increase in flow usually leads to a change in pump outlet pressure. For example, when the flow increases, the pump needs higher pressure to maintain output. An increase in flow usually leads to an increase in motor power, and the pump needs more energy to transport more liquid.

[0059] A modal component has a higher correlation with the pressure and power signals, it is more likely to be an effective signal containing useful information; conversely, if a modal component has a lower correlation with the pressure and power signals, it is more likely to be noise. Therefore, the comprehensive synergy index of the synchronously collected pump outlet pressure signal and motor power signal is calculated, and the specific steps are as follows:

[0060] With any IMF component as the to-be-analyzed component, the Pearson correlation coefficient between the to-be-analyzed component and each pressure sample segment of the pressure signal is calculated, the Pearson correlation coefficient between the to-be-analyzed component and each motor power sample segment of the motor power signal is calculated, and the sample segments of the pressure signal and the sample segments of the motor power signal are weighted and fused based on preset weights to obtain the comprehensive synergy index.

[0061] For example, the preset weights of the sample segments of the pressure signal and the sample segments of the motor power signal are 0.5 respectively, which can be adjusted according to actual conditions.

[0062] S5: Using the reconstructed flow signal data and the corresponding working condition state label, the preset LSTM model is trained to output the probability value of each working condition state label, the optimal parameter combination is selected based on the parameter candidate set for the real-time acquired flow signal data, the reconstructed and input trained LSTM model is completed, the probability of the corresponding label is output, and the real-time online monitoring of the emulsion pump flow is completed.

[0063] Among them, the corresponding working condition state labels are: normal , blockage , leakage .

[0064] The obtained sample set is divided into training set, validation set and test set in time sequence according to the ratio of 8:1:1. The model is built, which includes input layer, LSTM (Long Short-Term Memory, long short-term memory network) layer, Dropout (Dropout Regularization, dropout regularization) layer and output layer. The input layer receives an input like (batch size, sequence length 100, feature dimension 1). The LSTM layer is configured as a single layer, and the number of neurons is 64 according to the situation. The dropout rate of the Dropout layer is set to 0.3, which is designed to prevent overfitting. The output layer adopts full connection configuration, operates with Softmax activation function, and outputs the probability covering 3 categories, completing the LSTM model training.

[0065] In one embodiment, the real-time data is subjected to the above steps, the optimal parameter combination is obtained based on the real-time parameter candidate set, the decomposition is carried out according to the optimal parameter combination, the flow signal reconstruction is completed, and the trained LSTM model is input to output the probability belonging to each working condition state, and the real-time online monitoring and abnormal identification of the emulsion pump flow are completed.

[0066] The present application also provides a circulating neural network-based emulsion pump flow monitoring system. As shown in the figure, the system comprises a processor and a memory, the memory stores computer program instructions, when the computer program instructions are executed by the processor, the circulating neural network-based emulsion pump flow monitoring method according to the first aspect of the present application is realized. The system also includes a communication bus and a communication interface and other components well known to those skilled in the art, the setting and function of which are known in the art, and therefore will not be described here. Figure 2

[0067] It should be pointed out that for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which all belong to the protection scope of the present application. Therefore, the protection scope of the present application patent should be subject to the appended claims.​

Claims

1. A method for monitoring the flow rate of an emulsion pump based on a recurrent neural network, characterized in that, The method comprises the following steps: acquire historical running data and working condition state labels of the emulsion pump station, the historical running data comprising: flow signal, pressure signal and motor power signal, pre-process the flow signal to obtain a pre-processed flow sample segment; set the range of modal number and bandwidth, combine the parameters of modal number and bandwidth as a genetic individual, randomly generate initial individuals and construct a population, and perform variational modal decomposition on each flow sample segment, and evaluate the comprehensive optimization target value of each flow sample segment according to the modal separation degree and signal fidelity; based on the comprehensive optimization target value, perform iterative optimization, select the optimal parameter combination, construct a historical database of mapping relationship, and construct a real-time parameter candidate set based on the frequency of parameter combination; based on the optimal parameter combination, perform modal decomposition to obtain a plurality of IMF components, determine the weight based on the correlation between the pressure signal and the motor power signal and each IMF component, and construct a comprehensive synergy index; take the proportion of the comprehensive synergy index of each flow sample segment as the reconstruction weight, and perform weighted reconstruction on the IMF components of the optimal parameter combination to complete flow signal reconstruction; use the reconstructed flow signal data and the corresponding working condition state labels to train a preset LSTM model, output the probability value of each working condition state label, select the optimal parameter combination based on the parameter candidate set for the real-time acquired flow signal data, complete reconstruction and input into the trained LSTM model, output the probability of the corresponding label, and complete real-time online monitoring of the emulsion pump flow.

2. The Recurrent Neural Network based emulsion pump flow monitoring method of claim 1, wherein, The modal separation degree calculation method comprises: taking any flow sample segment as a target segment, obtaining the IMF components after decomposition of the target segment, calculating the mutual information between each pair of IMF components, and accumulating and summing the mutual information between each pair of IMF components to obtain the modal separation degree of the target segment.

3. The Recurrent Neural Network based emulsion pump flow monitoring method of claim 1, wherein, The signal fidelity calculation method comprises: taking any flow sample segment as a target segment, summing all IMF components of the target segment to obtain a reconstructed signal, calculating the difference between the original signal of each sampling point in the target segment and the corresponding sampling point reconstructed signal, calculating the square error of the difference, and accumulating the square error of all sampling points in the target segment to obtain the signal fidelity of the target segment.

4. The Recurrent Neural Network based emulsion pump flow monitoring method of claim 1, wherein, The comprehensive optimization target value calculation method comprises: taking any flow sample segment as a target segment, calculating the difference between adjacent sampling points in the target segment, and accumulating the average value of all adjacent sampling point differences as the fluctuation intensity of the target segment; performing exponential mapping on the fluctuation intensity using a negative exponential function, taking the reciprocal of 1 plus the mapping result as the modal separation weight, taking 1 minus the modal separation weight as the weight of the signal fidelity, and performing weighted sum to obtain the comprehensive optimization target value.

5. The recurrent neural network-based emulsion pump flow rate monitoring method of claim 1, wherein, The comprehensive synergy index calculation method comprises: taking any IMF component as an analysis component, calculating the Pearson correlation coefficient between the analysis component and each pressure signal sample segment, calculating the Pearson correlation coefficient between the analysis component and each motor power signal sample segment, and performing weighted fusion on the sample segments of the pressure signal and the motor power signal based on a preset weight to obtain the comprehensive synergy index.

6. The recurrent neural network-based emulsion pump flow rate monitoring method of claim 1, wherein, The constructing the real-time parameter candidate set comprises: The occurrence frequency of each parameter combination in the historical database is counted, the occurrence frequencies are sorted in descending order, and a preset number of parameter combinations are constructed into the real-time parameter candidate set, wherein the historical database comprises: a flow sample segment, a fluctuation intensity, an optimal modal number, and an optimal penalty factor.

7. The recurrent neural network-based emulsion pump flow rate monitoring method of claim 1, wherein, The working condition state label comprises: normal, blocked pipe, and leakage.

8. The recurrent neural network-based emulsion pump flow rate monitoring method of claim 1, wherein, The preprocessing of the flow signal comprises: Based on a preset sliding window length and a preset sliding step, a sample segment within the preset sliding window length is obtained, 3σ criterion method is used for each sample segment to eliminate abnormal points, and linear interpolation method is used to fill in missing data.

9. An emulsion pump flow monitoring system based on a recurrent neural network, characterized in that, Comprise: A processor and a memory, the memory stores computer program instructions, when the computer program instructions are executed by the processor, the method for monitoring the flow of the emulsion pump based on the recurrent neural network is realized according to any one of claims 1-8.

Images