Mechanical equipment fault diagnosis method based on convolutional neural network and bayesian network

Abstract

The present application relates to the technical field of fault diagnosis, in particular to a mechanical equipment fault diagnosis method based on convolutional neural network and Bayesian network. The method comprises the following steps: S1: monitoring various parameters of the emulsion pump through a sensor; S2: collecting various monitoring data when the emulsion pump occurs various faults, and performing normalization and standardization pretreatment on the data to obtain a trained emulsion pump fault diagnosis model; S3: building a diagnosis Bayesian network, including a fault layer and a reason layer; S4: embedding the emulsion pump fault diagnosis model composed of the trained lightweight convolutional neural network and the diagnosis Bayesian network into a cloud platform; S5: monitoring various data of the emulsion pump in real time, and inputting the data into the cloud platform; if a fault is diagnosed, an early warning is sent, and the diagnosis result is input into the Bayesian network for conditional probability reasoning, so that the probability of each fault reason is obtained, and fault positioning is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault diagnosis, in particular to a mechanical equipment fault diagnosis method based on convolutional neural network and Bayesian network. BACKGROUND

[0002] Coal is an important basic energy in China. According to the relevant report of the Chinese Academy of Engineering, by 2030, coal will account for 50% of China's energy structure, and by 2050, it will still be the top one with a proportion of 40%. Safe and efficient production of coal mines is one of the important foundations to ensure China's economic development and social stability. At present, coal is still the main energy in China, and the complex and harsh environment in the coal mine makes the coal mine equipment failure frequent, which seriously threatens the safety of coal mining and causes serious loss to the economic income of coal mines.

[0003] In the daily production process of fully mechanized coal mining face, various mechanical equipment plays a key role, but due to the complex underground working conditions and the harsh environment, the daily maintenance of the equipment is time-consuming and labor-intensive, and the on-site equipment staff has high technical requirements. When a fault occurs, it is difficult to achieve rapid and accurate fault diagnosis and fault positioning, which greatly prolongs the maintenance time, causes long-term shutdown, and aggravates the economic loss of coal mines. At present, there are few studies on mechanical equipment fault diagnosis methods, mostly through PLC and other online monitoring of mechanical equipment parameters, which cannot accurately prevent and diagnose. The fault mechanism of mechanical equipment is complex, including various fault phenomena, fault coupling, different causes, and difficult fault positioning, and traditional fault diagnosis methods cannot perform the task of mechanical equipment fault diagnosis.

[0004] Therefore, a mechanical equipment fault diagnosis method based on convolutional neural network and Bayesian network is proposed, which can monitor and diagnose the running data of mechanical equipment online, analyze and judge the running state of mechanical equipment, provide reference for daily maintenance of mechanical equipment, and when a fault occurs, it can use the diagnostic Bayesian network to realize rapid and accurate fault diagnosis and analysis of the causes, which has important significance in practical application. SUMMARY

[0005] The present application monitors various real-time data of mechanical equipment in actual operation through a lightweight convolutional neural network, performs real-time diagnosis, and can timely issue an alarm when a fault occurs, and uses Bayesian network to infer the fault cause, thereby providing a mechanical equipment fault diagnosis method based on convolutional neural network and Bayesian network.

[0006] The application adopts the following technical scheme: a mechanical equipment fault diagnosis method based on a convolutional neural network and a Bayesian network, comprising the following steps: S1: monitoring various parameters of an emulsion pump through a sensor; S2: collecting various monitoring data when the emulsion pump occurs various faults, performing normalization and standardization preprocessing on the monitoring data, training a light-weight convolutional neural network model using the preprocessed monitoring data, and obtaining a trained emulsion pump fault diagnosis model; S3: building a diagnostic Bayesian network, including a fault layer and a reason layer; S4: embedding the emulsion pump fault diagnosis model composed of the trained light-weight convolutional neural network and the diagnostic Bayesian network into a cloud platform; S5: monitoring various data of the emulsion pump in real time and inputting the data into the cloud platform, issuing a warning if a fault is diagnosed, and inputting the diagnosis result into the Bayesian network for conditional probability reasoning to obtain the probability of each fault reason and realize fault positioning.

[0007] Step S2 adopts the following method,

[0008] S21: 100 samples are taken for each normal state and each fault form, each sample including monitoring parameter data and corresponding labels, and the monitoring data in each sample is collected at a sampling frequency of 100 Hz for 1 minute, and finally all samples are divided into a training set, a validation set and a test set according to a ratio of 6:2:2;

[0009] S22: inputting the divided training set into the built light-weight convolutional neural network initial model for training, verifying with the validation set, then inputting the test set into the model for testing, and verifying whether the output result is consistent with the actual fault type; in the training process, an Adam learning rate adaptive algorithm is adopted, an optimal model is found through a cross-entropy loss function, the parameters and structure of the optimal model are saved, and thus a trained emulsion pump fault diagnosis model is obtained.

[0010] The light-weight convolutional neural network initial model comprises a two-dimensional separable convolutional layer, a two-dimensional maximum pooling layer, a global average pooling layer and a Softmax classifier.

[0011] Step S3 adopts the following method,

[0012] S31: the fault layer is composed of fault forms of the emulsion pump, and the reason layer is composed of reasons for causing each fault to occur;

[0013] S32: collecting fault maintenance records of the emulsion pump within a period of time, finding 50 cases of each fault form according to the records, and each case having a corresponding specific reason;

[0014] S33: According to the correspondence between each case fault form and reason, the connection relationship between the fault layer and the reason layer of the diagnostic Bayesian network structure is determined, so as to determine the structure of the Bayesian network;

[0015] S34: By combining the already built Bayesian network structure and the collected case data, the conditional probability between the fault nodes and the reason nodes in the structure of the Bayesian network and the prior probability of each fault occurrence are determined, and finally the conditional probability table of the Bayesian network is formed.

[0016] The detailed steps of step S34 conditional probability table determination are as follows: first, the relevant data of the reasons causing the fault are screened out from the cases of one fault form, quantitative analysis is performed, and the prior probability P 先验 of each reason is determined; second, the probability of fault occurrence caused by each reason is counted as the conditional probability of fault and reason; in this way, each fault is processed according to the above steps, and thus the conditional probability table P 条件 of the diagnostic Bayesian network is determined.

[0017] The specific process of step S5 Bayesian network conditional probability inference is as follows,

[0018] S51: The probability of fault occurrence caused by various reasons is calculated, and the calculation method is as follows,

[0019]

[0020] In the formula, P 原因 is the probability of the reason causing the fault, P 先验 is the prior probability of each reason, P 条件 is the conditional probability table of the diagnostic Bayesian network, P 故障 is the probability of fault occurrence; the fault form with the maximum probability value in P 原因 is taken as the diagnostic conclusion;

[0021] S52: According to the fault form diagnosed by the convolutional neural network as the input of the diagnostic Bayesian network model, the following three situations will occur:

[0022] 1. If it is found through maintenance that the fault is caused by the reason corresponding to the node, and this reason is a bottom event in the fault tree, then the reason is the cause of the fault;

[0023] 2. If the maintenance result shows that the fault is caused by the reason corresponding to the node, and this reason is a middle event of the fault tree, then it is caused by the bottom event, and further inference is needed; set the node to 100% occurrence, continue to deepen the inference, find the parent node with the maximum posterior probability, and then perform maintenance again;

[0024] 3、If the maintenance result shows that the fault is not caused by the reason corresponding to the node, it means that the reason represented by the node has not occurred, and the node needs to be set to 100% not to occur as evidence to continue reasoning to find the parent node with the maximum posterior probability for maintenance; corresponding to the maintenance result, the above related steps are repeated until the cause of the fault is found.

[0025] In step S1, the parameters of the emulsion pump include: motor current, motor torque, motor speed, emulsion pump system pressure, emulsion flow, emulsion pump vibration, emulsion concentration, emulsion temperature, emulsion level, emulsion oil level, lubricating oil temperature, injection pressure, winding temperature, bearing temperature, water tank water level and sound signal.

[0026] In step S2, the fault forms of the emulsion pump include: pump cannot start, pump starts without pressure or pressure cannot be increased, pressure pulsation, insufficient flow or no flow, crankcase temperature is too high, system pressure cannot be increased and slowly decreases, pump pressure suddenly rises above the relief valve set pressure, pump noise during operation, unloading valve frequently acts when the support stops supplying liquid, pump has no discharge during operation, the matching concentration cannot be increased, and a total of 11 fault forms of the motor fault.

[0027] Compared with the prior art, the advantages of the present application are that the convolutional neural network is combined with the Bayesian network, the advantages of each are fully utilized, the separable convolution is adopted in the convolutional layer, compared with the ordinary convolution, a large number of parameters are reduced, the diagnostic efficiency of the model is greatly improved, and the input data is a fusion of multi-source data, which can further improve the diagnostic accuracy. At the same time, the global average pooling is adopted in the last pooling layer, compared with the fully connected layer, the problem of model overfitting caused by the explosion of parameters in the fully connected layer is avoided. After fault diagnosis, the Bayesian network is used for uncertainty reasoning of the fault reason, compared with the expert system, not only all possible fault reasons can be reasoned out, but also the probability of occurrence of each reason can be obtained. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 is a flowchart of the present application;

[0029] Figure 2 is a multi-source monitoring data fusion flowchart of the present application (taking an emulsion pump as an example);

[0030] Figure 3 is a structure diagram of an ordinary convolutional layer;

[0031] Figure 4 is a separable convolutional layer structure diagram of the specific embodiment of the present application;

[0032] Figure 5 is a lightweight convolutional neural network structure diagram provided by the present application;

[0033] Figure 6 is a light weight convolutional neural network training and testing flowchart provided by the present application;

[0034] Figure 7 is a diagnostic Bayesian network model structure diagram (taking an emulsion pump as an example) provided by the present application;

[0035] Figure 8 is a diagnostic Bayesian network model flowchart provided by the present application. DETAILED DESCRIPTION

[0036] The present application will be described in detail below with specific embodiments. The following embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any form. It should be pointed out that, for those skilled in the art, without departing from the concept of the present application, a number of changes and improvements can be made. These all belong to the protection scope of the present application.

[0037] Referring to Figure 1 The present embodiment discloses a mechanical equipment fault diagnosis method based on convolutional neural network and Bayesian network, which is described in detail below taking an emulsion pump as an example, mainly including the following steps:

[0038] S1: In order to comprehensively and accurately diagnose the emulsion pump, a multi-monitoring parameter fusion method is used to monitor various parameters of the emulsion pump through various sensors.

[0039] S2: Collecting various monitoring data when the emulsion pump fails is artificially collected, such as collecting a certain amount of data when the emulsion pump has no pressure after starting, collecting a certain amount of data when the emulsion pump crankcase temperature is too high, collecting a certain amount of data when the emulsion pump system pressure cannot be adjusted and slowly decreases, and so on, plus 12 kinds of normal state failure modes.

[0040] For data collection, as shown in Table 1, 100 samples are taken for each failure mode, including normal state, a total of 1200 samples, A, B... M in the table are 16 monitoring parameters, each sample includes the 16 monitoring parameters, and the monitoring data in each sample is collected at a sampling frequency of 100 Hz, the collection time is 1 minute, and the matrix composed of the time sequence data of each sample becomes the input of the convolutional neural network model.

[0041] Table 1

[0042]

[0043] The collected data is normalized and standardized, and the training set, the validation set and the test set are set according to the number of the collected data set in the ratio of 6:2:2. The initial light-weight convolutional neural network model is trained by using the preprocessed training set, and the validation set is used for verification to prevent model overfitting. Then the test set is input into the model for testing, and whether the output result is consistent with the actual fault type is verified. In the training process, the Adam learning rate adaptive algorithm is used, and the optimal model is found through the cross-entropy loss function. The parameters and structure of the optimal model are saved, so that a good emulsion pump fault diagnosis model is obtained.

[0044] Among them, as Figure 5 , the initial model of the light-weight convolutional neural network includes: a two-dimensional separable convolution layer, a two-dimensional maximum pooling layer, a global average pooling layer and a Softmax classifier. The number of convolution layers, pooling layers and global average pooling layers in the convolutional neural network model, the size and number of convolution kernels, and the operation step length and other hyperparameters are set. The appropriate model training parameters are selected, including training batch, batch size and initial learning rate, and the data in S2 is used for training.

[0045] Convolution layer: the convolution layer uses a convolution kernel to perform convolution operation on the local area of the input signal and generates corresponding features. The convolution layer has the characteristic of weight sharing, that is, the same convolution kernel will traverse the input once with a fixed step. The present application selects a separable convolution. The separable convolution can be regarded as dividing the ordinary convolution into two parts: spatial convolution and channel convolution, as shown in Figure 4 , first, the spatial convolution is performed through the convolution kernel of , and then the channel convolution is performed through the convolution kernel of , and finally the same result as the ordinary convolution of Figure 3 is output. By calculating the number of parameters of ordinary convolution and separable convolution, it can be obtained that the parameter amount of ordinary convolution is:

[0046] (1)

[0047] The parameter amount of separable convolution is:

[0048] (2)

[0049] In equations (1) and (2), is the input layer size, is the convolution layer size, is the input feature channel number, is the number of convolution kernels.

[0050] By comparing the parameter amount of separable convolution and ordinary convolution:

[0051] (3)

[0052] It can be concluded that for the same input layer and output layer, the separable convolution can greatly reduce the parameters compared with the ordinary convolution, thereby shortening the training time and improving the training efficiency.

[0053] In the calculation of the convolution layer, the convolution adopts point multiplication of the input matrix and the corresponding points in the convolution kernel, and then sums up, and adds a bias, and the calculation formula is:

[0054] (4)

[0055] In the formula, K represents the number of channels, M is the number of rows of the convolution kernel of each channel, and N is the number of columns of the convolution kernel of each channel. n y represents the convolution output result, b n a represents the bias in linear calculation, and a i,j,k is the weighting coefficient in linear operation, and x i,j,k represents the feature element value of the original input or the output result of the previous convolution layer.

[0056] Pooling layer: the maximum pooling layer is selected in the application, which mainly functions as downsampling, dimensionality reduction, removing redundant information, compressing features, simplifying network complexity, reducing calculation amount, reducing memory consumption, and also having the characteristics of realizing nonlinearity, expanding the perception field, realizing invariance (translation invariance, rotation invariance and scale invariance) and the like. , the pooling output is:

[0057] (5)

[0058] In the formula, n represents a part of the input vector division, represents the i-th feature tensor, W represents the size of the pooling window, and S represents the step length.

[0059] Global average pooling (GAP): the global average pooling layer is used to replace the fully connected layer after the convolution layer, because: the global average pooling is simpler and more natural in the conversion between the feature map and the final classification, and the global average pooling does not need as many training parameters as the fully connected layer, so that the space parameters make the model more robust, and the anti-overfitting effect is better.

[0060] Classifier layer: the activation function of the output layer is the Softmax function. Through the action of the Softmax function, it is mapped into a value of (0, 1), and the sum of these values is 1 (satisfying the nature of probability), so it can be understood as probability. When selecting the output node at the end, the node with the maximum probability (that is, the value corresponding to the maximum) can be selected as the final prediction target.

[0061] (6)

[0062] Loss function: cross-entropy loss function is adopted in the application:

[0063] (7)

[0064] In the formula: E is the objective function, n is the sample number, y is the true value, and t is the predicted value.

[0065] S3: build a diagnostic Bayesian network, including a fault layer and a reason layer. Collect historical records of the causes of emulsion pump failure to make a statistical table, determine the fault layer nodes and the reason layer nodes, further establish the connection relationship between the fault layer and the reason layer, and determine the conditional probability table of the Bayesian network according to the collected historical records, and complete the above steps, that is, the construction of the diagnostic Bayesian network is completed.

[0066] Bayesian network is a kind of probability network, which is a graphical network based on probability reasoning. Bayesian network is a mathematical model based on probability reasoning. Probability reasoning is a process of obtaining other probability information through some variable information. Bayesian network based on probability reasoning is proposed to solve the problems of uncertainty and incompleteness, and has good advantages in solving the faults caused by uncertainty and correlation of complex equipment, and is widely used in many fields.

[0067] The step S3 further comprises:

[0068] S31: the fault layer is composed of 16 fault forms of the emulsion pump, and the reason layer is composed of reasons causing each fault to occur.

[0069] S32: collect the fault maintenance records of the emulsion pump in a period of time, find 50 cases of each fault form according to the records, a total of 550 cases, and each case has its corresponding specific reason.

[0070] S33: according to the corresponding relationship between each case fault form and reason, determine the connection relationship between the fault layer and the reason layer of the diagnostic Bayesian network structure, and thus determine the structure of the Bayesian network.

[0071] S34: By combining the already built Bayesian network structure and the collected case data, the conditional probability between the fault nodes and the reason nodes in the structure of the Bayesian network and the prior probability of each reason leading to the occurrence of the fault are determined, and finally the conditional probability table of the Bayesian network is formed. The detailed steps for determining the conditional probability table are as follows: for example, for the fault of no pressure or pressure unable to be adjusted after the pump is started, first, the relevant data of the reasons causing this fault are screened out from the collected cases, quantitative analysis is performed, and the prior probability P 先验 of each reason is determined; secondly, the probability of the occurrence of the fault due to each reason is also counted, which is taken as the conditional probability of the fault and the reason; in this way, each fault is processed according to the above steps, and thus the conditional probability table P 条件 of the diagnostic Bayesian network is determined.

[0072] S4: Embed the emulsion pump fault diagnosis model composed of the trained lightweight convolutional neural network and the Bayesian network into the intelligent cloud platform of the fully mechanized coal mining face.

[0073] S5: Real-time monitoring is performed on each data during the operation of the emulsion pump, and the data is input to the intelligent cloud platform of the fully mechanized coal mining face; if it is diagnosed as normal, the monitoring will continue; if it is diagnosed as a fault, a warning will be issued, and the fault diagnosis result is input to the Bayesian network for conditional probability reasoning, so as to obtain the probability of each fault reason, and the reason corresponding to the node with the largest posterior probability value is taken as the diagnosis result, thereby realizing fault positioning and helping the front-line equipment maintenance personnel to perform rapid maintenance.

[0074] Application of the diagnostic Bayesian network. According to the fault form diagnosed by the convolutional neural network as the input of the diagnostic Bayesian network model, it is represented that the fault node occurs 100%. The probability calculation method due to various reasons leading to the occurrence of the fault is shown in formula 8, and the maximum probability value is taken as the diagnostic conclusion.

[0075] (8)

[0076] In the formula, P 原因 is the probability of the reason causing the fault, P 先验 is the prior probability of each reason, P 条件 is the conditional probability table of the diagnostic Bayesian network, and P 故障 is the probability of the occurrence of the fault.

[0077] The present application uses GeNIe 2.0 software to build a diagnostic Bayesian network (BN). 11 kinds of faults are taken as the child nodes of the diagnostic Bayesian network, and the reasons for the occurrence of each fault are taken as the parent nodes of the diagnostic Bayesian network. For example, the fault diagnosis Bayesian network of the pump starting without pressure or the pressure unable to be adjusted is as follows Figure 7As shown, there are 17 nodes in total, and the structure diagram of other failure forms and causes is similar to this.

[0078] The specific reasoning process is as follows Figure 7 Taking the failure of no pressure or pressure unable to be adjusted after the pump is started as an example, when the Bayesian network receives the failure form, the reason corresponding to the node with the maximum posterior probability value is taken as the diagnosis result, and the repair is performed, and the following three cases will occur:

[0079] 1. Through the repair, it is found that the failure is caused by the reason corresponding to the node, and the reason is located in the bottom event in the fault tree, so the reason is the cause of the failure.

[0080] 2. If the repair result shows that the failure is caused by the reason corresponding to the node, and the reason is a middle event of the fault tree, it is indicated that the occurrence is caused by the bottom event, and further reasoning is needed; the node is set to 100% occurrence, the reasoning is continued, the parent node with the maximum posterior probability is found, and the repair is performed again.

[0081] 3. If the repair result shows that the failure is not caused by the reason corresponding to the node, it is indicated that the reason represented by the node does not occur, and then the node is set to 100% non-occurrence as evidence input to continue reasoning, the parent node with the maximum posterior probability is found, and the repair is performed.

[0082] According to the repair result, the above related steps are repeated until the cause of the failure is repaired.

[0083] Taking the convolutional neural network recognition as an example, the failure of no pressure or pressure unable to be adjusted after the emulsion pump is started is taken as an example, as follows Figure 7As shown, firstly, the convolutional neural network output is taken as the Bayesian network input, that is, the "no pressure or pressure cannot be increased after pump start" node is set to 100% occurrence; secondly, the properties of each node in the BN are defined. "Observations" means observation nodes, generally indicating top events, accident types, fault types and the like; "Targets" means target nodes, generally indicating nodes whose posterior probabilities are expected to be obtained through evidence input. Herein, the "no pressure or pressure cannot be increased after pump start" node is defined as "Observations", and the remaining nodes are defined as "Targets"; finally, ClusteringAlgorithm (joint tree algorithm) in GeNIe2.0 software is selected to update the BN, and the inference result is completed, and the posterior probabilities of each target node are obtained. By comparing the posterior probabilities of each target node, it is assumed that the "main valve fault" with the maximum posterior probability value is selected as the diagnostic result, and maintenance is performed. The maintenance result shows that there is indeed a main valve fault, but since it corresponds to an intermediate event of the fault tree, it must be caused by the bottom event, so further diagnosis is still required. The main valve fault is input into the diagnostic Bayesian network as known information, that is, the "main valve fault" node is set to 100% occurrence, and the BN is updated again. It is assumed that the "main valve jam" with the maximum posterior probability value is selected as the diagnostic result, and maintenance is performed. The maintenance result shows that there is indeed a main valve jam, and since it is located at the bottom event of the fault tree, the fault diagnosis is completed. According to the diagnosed fault cause, maintenance is performed, so that the equipment can work normally.

[0084] The main principle of the present application is that firstly, a lightweight convolutional neural network model is used to identify the fault type, which not only takes advantage of the powerful data feature extraction capability of the convolutional neural network (CNN), but also reduces the parameters by using separable convolution layers, thereby improving the training efficiency of the model; then, a global average pooling layer is used in the diagnostic model, which also reduces the parameters and avoids model overfitting; finally, the diagnostic Bayesian network is used for uncertainty reasoning of the fault cause, and the probability of each fault cause is obtained. Compared with the expert system, the diagnostic efficiency is improved, the interference of human factors is reduced, and the intelligent level of the diagnosis is improved as a whole.

[0085] The specific embodiments of the present application are described above. It should be understood that the present application is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essential content of the present application. The embodiments of the present application and the features in the embodiments can be arbitrarily combined without conflict.

Claims

1. A method for mechanical equipment fault diagnosis based on convolutional neural network and Bayesian network, characterized in that: The method comprises the following steps, S1: monitoring various parameters of the emulsion pump by a sensor; S2: collecting various monitoring data when the emulsion pump occurs various faults, normalizing and standardizing the monitoring data, training a light-weight convolutional neural network model by using the monitoring data, and obtaining a trained emulsion pump fault diagnosis model; The light-weight convolutional neural network initial model comprises a two-dimensional separable convolutional layer, a two-dimensional maximum pooling layer, a global average pooling layer, and a Softmax classifier; S3: building a diagnosis Bayesian network, including a fault layer and a reason layer; The fault layer is composed of fault forms of the emulsion pump, and the reason layer is composed of reasons for causing each fault to occur; S4: embedding the emulsion pump fault diagnosis model composed of the trained light-weight convolutional neural network and the diagnosis Bayesian network into a cloud platform; S5: monitoring various data of the emulsion pump in real time, inputting the data into the light-weight convolutional neural network, issuing a warning if a fault is diagnosed, inputting the diagnosis result into the Bayesian network for conditional probability reasoning, obtaining probabilities of various fault reasons, and realizing fault positioning; The specific process of the conditional probability reasoning of the Bayesian network in step S5 is, S51: calculating the probability of causing the fault due to various reasons, and the calculation method is as follows, where P 原因 is the probability of a fault occurring due to a cause 先验 is the prior probability of a cause 条件 is the conditional probability table of the Bayesian network for diagnosis 故障 is the probability of a fault occurring; P 原因 is the probability of a fault occurring; P 原因 is the probability of a fault occurring; P 原因 is the probability of a fault occurring; P 原因 is the probability of a fault occurring; P 原因 is the probability of a fault occurring; P 原因 is the probability of a fault occurring; P 原因 is the probability of a fault occurring; P 原因 is the probability of a fault occurring; P 原因 is the probability of a fault occurring; S52: taking the fault form diagnosed by the convolutional neural network as the input of the diagnosis Bayesian network model, and the following three situations will occur: If it is found through maintenance that the fault is caused by the reason corresponding to the node, and the reason is a bottom event in the fault tree, it is indicated that the reason is the cause of the fault; If the maintenance result shows that the fault is caused by the reason corresponding to the node, and the reason is a middle event of the fault tree, it is indicated that the occurrence is caused by a bottom event, and further reasoning is required; the node is set to 100% occurrence, and further reasoning is continued to find the parent node with the maximum posterior probability for maintenance; If the maintenance result shows that the fault is not caused by the reason corresponding to the node, it is indicated that the reason represented by the node does not occur, and the node is set to 100% non-occurrence as evidence for further reasoning to find the parent node with the maximum posterior probability for maintenance; corresponding to the maintenance result, the steps S51 and S52 are repeated until the cause of the fault is found.

2. The method of claim 1, wherein the method comprises: The step S2 adopts the following method, S21: 100 samples are taken for each fault form and normal state, each sample includes monitoring parameter data and corresponding labels, and the monitoring data in each sample is collected at a sampling frequency of 100 Hz for 1 minute; finally, all samples are divided into a training set, a verification set and a test set according to a ratio of 6:2:

2. S22: input the divided training set into the built lightweight convolutional neural network initial model for training, and use the validation set for verification, then input the test set into the model for testing, and verify whether the output result is consistent with the actual fault type; in the training process, the Adam learning rate adaptive algorithm is adopted, the optimal model is found through the cross entropy loss function, the parameters and structure of the optimal model are saved, and thus a trained emulsion pump fault diagnosis model is obtained.

3. The method of claim 1, wherein the method comprises: The step S3 adopts the following method, S31: the fault layer is composed of the fault forms of the emulsion pump, and the reason layer is composed of the reasons for causing each fault to occur; S32: collect the fault maintenance records of the emulsion pump in a period of time, find 50 cases of each fault form according to the records, and each case has a corresponding specific reason; S33: according to the corresponding relationship between each case fault form and reason, determine the connection relationship between the fault layer and the reason layer of the diagnostic Bayesian network structure, so as to determine the structure of the Bayesian network; S34: by combining the already built Bayesian network structure and the collected case data, the conditional probability between the fault nodes and the reason nodes in the Bayesian network structure and the prior probability of each fault occurrence are determined, and finally the conditional probability table of the Bayesian network is formed.

4. The method of claim 3, wherein the method further comprises: The detailed steps of the step S34 of determining the conditional probability table are as follows: firstly, the relevant data of the causes of the fault are screened out from the cases collected from a fault form, quantitative analysis is performed, and the prior probability P of each cause is determined 先验 ; secondly, the probability of the fault occurrence caused by each cause is counted and taken as the conditional probability of the fault and the cause; and in this way, each fault is processed according to the steps, and the conditional probability table P of the diagnostic Bayesian network is determined 条件 .

5. The method of claim 1, wherein the method further comprises: In the step S1, the parameters of the emulsion pump include: motor current, motor torque, motor speed, emulsion pump system pressure, emulsion flow, emulsion pump vibration, emulsion concentration, emulsion temperature, emulsion level, emulsion oil level, lubricating oil temperature, injection pressure, winding temperature, bearing temperature, water tank water level and sound signal.

6. The method of claim 5, wherein the method further comprises: In the step S2, the fault forms of the emulsion pump include: pump cannot start, pump starts without pressure or pressure cannot be raised, pressure pulsation, insufficient flow or no flow, crankcase temperature is too high, system pressure cannot be raised and slowly decreases, pump pressure suddenly rises above the unloading valve set pressure, pump runs with loud noise, unloading valve frequently acts when support stops supplying liquid, pump has no discharge during operation, matching concentration cannot be raised, and 11 fault forms of motor failure.

Images