Cement production line equipment fault diagnosis method based on FMEA and improved convolutional neural network

Abstract

The invention discloses a cement production line equipment fault diagnosis method based on FMEA and an improved convolutional neural network. The method comprises the following steps: (1) evaluating a rotary kiln, a coal mill system, a crusher and a warehousing elevator; (2) carrying out fault analysis on the mechanical equipment by using an FMEA table, and carrying out risk assessment and ranking by combining three evaluation standards of fault severity S, occurrence degree O and detection degree D; (3) collecting data and carrying out preprocessing to divide a training set and a test set; (4) constructing an improved CNN (Convolutional Neural Network) and training; (5) selecting ABC-based optimized CNN structure parameters, and testing the trained CNN model by using test set data to obtain a final output result; according to the invention, the technical precision of fault diagnosis is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing faults, and particularly relates to a method for diagnosing equipment faults in a cement production line based on FMEA and an improved convolutional neural network. Background Technique

[0002] With the advent of the big data era, big data analysis based on intelligent algorithms has begun to be applied to the fault diagnosis of rotary kilns. Existing technologies have proposed that there are many complex vibrations in crushers. Some vibrations play a very important role in the performance of the machine, such as characteristic frequencies; some vibrations will have a serious impact on the operation of the crusher, such as vibrations caused by component failures. By analyzing various vibrations of the crusher, it can provide a basis for the design improvement and fault diagnosis of the crusher. Scholars such as Wang Meirong have fully illustrated through diagnostic examples that using the change characteristics of vibration parameters, spectrograms, and time-domain dimensionless index analysis methods to analyze the vibration faults of rotating machinery is a simple and efficient fault diagnosis method, which has very important reference value for the fault diagnosis and maintenance plan formulation of rotating equipment. However, there are problems such as low accuracy of fault diagnosis technology and weak generalization ability, and it is unable to effectively handle the complex, uncertain, and concurrent faults of various equipment on the cement production line. Summary of the Invention

[0003] Object of the Invention: The object of the present invention is to provide a method for diagnosing equipment faults in a cement production line based on FMEA and an improved convolutional neural network, so as to solve the problems of low accuracy of fault diagnosis technology and weak generalization ability.

[0004] Technical Solution: A method for diagnosing equipment faults in a cement production line based on FMEA and an improved convolutional neural network according to the present invention includes the following steps:

[0005] (1) Evaluate the four rotating mechanical equipment, namely rotary kiln, coal mill system, crusher, and in-warehouse elevator, in terms of the production stoppage time, maintenance cost, quality, safety, and environmental protection affected by faults, and rate them into 3 grades of A, B, and C;

[0006] (2) Conduct fault analysis on the mechanical equipment using an FMEA form, and perform risk assessment and ranking in combination with three evaluation criteria of fault severity S, occurrence O, and detectability D;

[0007] (3) Collect data and perform preprocessing to divide the training set and the test set;

[0008] (4) Construct an improved CNN network and perform training;

[0009] (5) Select to optimize the CNN structure parameters based on ABC, and test the trained CNN model with the test set data to obtain the final output result.

[0010] Further, in step (2), the risk degree value with the highest ranking is 1, and then the ranking is carried out in turn backward; an FMEA analysis table is obtained. When making improvements, the equipment corresponding to the failure mode with the most forward ranking should be improved first; through the FMEA analysis table, the failure with the highest RPN ranking of each failure mode of the four types of equipment belongs to the bearing-related failure, that is, the bearing has the highest impact on each rotating mechanical equipment.

[0011] Further, in step (3), according to the results of the FMEA analysis, measuring points are installed on multiple rotating mechanical equipment on the production line of the cement production enterprise for monitoring, and the monitoring points are arranged on the bearings and related components of the six types of rotating mechanical equipment.

[0012] Further, the preprocessing in step (3) is specifically as follows: the acceleration data is restricted to the same magnitude and the same range and normalized; then, the preprocessed data is divided into a training set and a test set; furthermore, the one-dimensional time-domain signal sequence is converted into a two-dimensional time-frequency diagram through data transformation, and the fault diagnosis problem is converted into a processing similar to an image recognition classification task.

[0013] Further, in step (4), the improved CNNW network consists of a seven-layer convolutional neural network structure composed of an input layer, two convolutional layers, two pooling layers, a fully connected layer, and an output layer; pre-training is carried out using the training set. After pre-training, the relevant parameters of the network are optimized. After optimization, the above training operations are repeated until ideal parameters are obtained, and a CNN model reconstructed with optimal parameters is obtained.

[0014] Further, step (5) is specifically as follows: the parameters to be optimized and determined are the number f 1 and size s 1 of the convolutional kernels of the first convolutional layer, 2 the number f 2 and size s

[0015] of the convolutional kernels of the second convolutional layer; the learning rate η and the training batch batch; these six parameters are mapped into the spatial vector positions of each honey source, and the position of the i-th honey source is expressed as: i x 1i =(f 2i , f i , s 2i , s i , batch i )

[0016] Among them, each honey source is a feasible solution. The quality of the honey source, that is, the quality of the feasible solution, is evaluated according to the amount of nectar of the honey source, that is, the fitness value. Through continuous iterative update search, the optimal honey source is found, which is the global optimal solution sought.

[0017] Furthermore, the output layer of the convolutional neural network uses the SoftMax function to convert the output signal into a distribution with a sum of probabilities of 1, and the fitness function calculates the average loss value of the i-th iteration; finally, after the bee colony searches and optimizes through N iterations, it finds the generation with the smallest average loss value, which corresponds to the globally optimal nectar source found.

[0018] A cement production line equipment fault diagnosis system based on FMEA and improved convolutional neural network according to the present invention includes:

[0019] Level module: used to evaluate the four rotating mechanical equipment, namely rotary kiln, coal mill system, crusher, and in-warehouse elevator, in terms of five aspects including production stop time, maintenance cost, quality, safety, and environmental protection affected by faults, and rate them into 3 levels: A, B, and C;

[0020] Evaluation module: used to perform fault analysis on mechanical equipment using the FMEA form, and conduct risk assessment and ranking in combination with three evaluation criteria: severity S, occurrence O, and detectability D;

[0021] Preprocessing module: used for data acquisition and preprocessing to divide the training set and the test set;

[0022] CNN module: used to construct an improved CNN network and perform training;

[0023] ABC module: used to select and optimize the CNN structure parameters based on ABC, and test the trained CNN model with the test set data to obtain the final output result.

[0024] An electronic device according to the present invention includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the computer program is loaded into the processor, it implements a cement production line equipment fault diagnosis method according to any one of the above.

[0025] A storage medium according to the present invention stores a computer program, and when the computer program is executed by a processor, it implements a cement production line equipment fault diagnosis method according to any one of the above.

[0026] Beneficial effects: Compared with the prior art, the present invention has the following remarkable advantages: applying the artificial bee colony algorithm to the parameter optimization of the CNN network, and optimizing six important structure parameters of the convolutional neural network. Compared with the mainstream algorithms, the convolutional neural network model optimized by the artificial bee colony algorithm has better effects, not only having higher recognition accuracy but also having a faster convergence speed. Description of the Drawings

[0027] Figure 1It is the FMEA analysis flow chart of the present invention;

[0028] Figure 2 It is the physical diagram of the measuring point layout of the coal mill of the present invention;

[0029] Figure 3 It is the data acquisition platform of the present invention;

[0030] Figure 4 It is the fault diagnosis flow chart of the present invention;

[0031] Figure 5 It is the fault diagnosis convolutional neural network architecture of the present invention;

[0032] Figure 6 It is the structure diagram of the seven-layer convolutional neural network of the present invention;

[0033] Figure 7 It is the flow chart of optimizing CNN by the ABC algorithm of the present invention. Specific implementation manner

[0034] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0035] The embodiment of the present invention provides a method for diagnosing faults in cement production line equipment based on FMEA and improved convolutional neural network, including the following steps:

[0036] Step 1: Evaluate the four rotating mechanical equipment, namely rotary kiln, coal grinding system, crusher and in-storage elevator of a certain cement company, in terms of five aspects: downtime, maintenance cost, quality, safety and environmental protection affected by faults, and rate them into 3 grades A, B, and C, as shown in Table 1;

[0037] Table 1 Classification table of cement production process equipment

[0038]

[0039] Step 2: Use the FMEA form to analyze the faults of the four rotating mechanical equipment, namely rotary kiln, coal grinding system, crusher and in-storage elevator on the cement production line of this cement company. The FMEA analysis process is as Figure 1As shown in the figure. Combining the three evaluation criteria of failure severity (S), occurrence (O), and detection (D) for risk assessment and ranking, the failure mode with the highest risk value is ranked 1, and then sorted in order. The FMEA analysis table is obtained. When making improvements, the equipment corresponding to the failure mode with the highest ranking should be improved first. From the FMEA analysis table, it can be seen that the failures with the highest risk rankings for each failure mode of these four pieces of equipment are all related to bearings, that is, the bearings have the greatest impact on each rotating mechanical equipment. Finding the root causes of failures and eliminating them for key components, and setting measurement points for key components for timely monitoring can ensure that the equipment status is reasonably adjusted, extend the service life of the equipment, ensure the normal operation of the cement production line to the greatest extent, reduce resource waste, and enable the enterprise to achieve maximum profits. Therefore, the subsequent research content of this paper will focus on the rolling bearings, the core components of the cement production line, as shown in Table 2.

[0040] Table 2 FMEA Failure Analysis Table

[0041]

[0042]

[0043]

[0044]

[0045]

[0046] Step 3: Data acquisition. According to the results of the FMEA analysis, measurement points are installed on multiple rotating mechanical equipment on the production line of this cement production enterprise for monitoring, and the monitoring points are arranged on the bearings and related components of these six rotating mechanical equipment. Taking the coal mill as an example to illustrate the specific layout of the measurement points, a total of 11 measurement points are arranged on the coal mill, one on the motor, and the rest on the gearbox. Figure 2 It is a physical diagram of the measurement point layout for the coal mill.

[0047] Step 4: The data collected by the acceleration sensor is one-dimensional acceleration data, while the input of the neural network we constructed is a two-dimensional matrix. Therefore, we need to convert the one-dimensional data into two-dimensional data. Before that, the original data needs to be processed first. The data acquisition platform is as Figure 3 shown:

[0048] Step1: Normalization processing

[0049] To prevent the influence of some extreme data in the original data on the overall vibration signal, this paper uses the normalization equation

[59] to map all acceleration data to the range of 0-1, ensuring that all data dimensions are the same, so that the algorithm can converge faster. The normalization equation is as follows:

[0050]

[0051] In this equation, X i represents each data after normalization, x i represents each data before normalization, and x min represents the smallest data in the sample, and x max represents the largest data in the sample.

[0052] Step2: Data transformation

[0053] The neural network we constructed is a convolutional neural network. Two-dimensional input can significantly improve the computational efficiency of the convolutional layer. Therefore, we need to transform the one-dimensional data. The original data shape is 1×400, and we transform it into 20×20. In this way, we get 3000 "images" of 20×20, which are used as the input of the neural network.

[0054] Step3: Division of the test set and the training set

[0055] For these ten different states, use the SoftMax classifier for classification. When classifying, a label needs to be assigned to each state of the bearing. Each label here is a column vector. For example, it represents a bearing in normal operation. According to the convention, the ratio of the training set to the test set is 4:1, that is, the training set for each bearing state is 240 "images", and the test set is 60 "images".

[0056] Step 5: Combining the relevant theories of convolutional neural networks, a rolling bearing fault diagnosis framework based on improved CNN is proposed. Use the training set for pre-training to enable the CNN model to complete feature extraction and classification evaluation. Then, the trained CNN model is tested with the test set data to obtain the final output result, as Figure 4 shown.

[0057] Step 5: Add a network optimization and CNN retraining process to the original architecture, as Figure 5 shown. After pre-training, optimize the relevant parameters of the network. After optimization, repeat the above training operations until ideal parameters are obtained, and obtain a CNN model with optimal parameter reconstruction.

[0058] Step 6: When used for fault diagnosis, the CNN often has an extremely complex structure. Relying solely on the method of randomly setting the network structure parameters is likely to fall into a local optimum. Therefore, the present invention proposes a seven-layer convolutional neural network structure including an input layer, two convolutional layers, two pooling layers, a fully connected layer, and an output layer, as Figure 6 shown.

[0059] Step 7: Select an algorithm for optimizing the CNN structure parameters based on ABC (ABC-CNN) to optimize the hyperparameters.

[0060] The parameters to be optimized and determined are the number f 1 and size s 1 of the convolutional kernels in the first convolutional layer, the number f 2 and size s 2 of the convolutional kernels in the second convolutional layer, the learning rate η, and the training batch size batch. Map these six parameters into the spatial vector position of each nectar source. The position of the i-th nectar source can be expressed as:

[0061] x i = (f 1i , f 2i , s i , s 2i , batch i , η i )

[0062] In the formula, each nectar source is a feasible solution. The quality of the nectar source, that is, the quality of the feasible solution, is evaluated according to the amount of nectar (fitness value) of the nectar source. We need to find the optimal nectar source through continuous iterative updates of the search, which is the global optimal solution we seek, as Figure 7 shown.

[0063] Step 8: Initialization of nectar sources

[0064] Set the number of nectar sources N = 30, the maximum number of failure times to 40, and the number of iterations to 400. Therefore, the number of employed bees and the number of follower bees are both N. First, the ABC algorithm randomly generates an initial population (nectar sources) containing N solutions. Each solution X i (i = 1, 2,..., N) is represented by a 6-dimensional vector x i = (x i1 , x i2 ,..., x i6 ) T The initial solution is generated according to the formula:

[0065] x ij = lb i + (ub j - lb j ) * rand(0, 1)

[0066] Due to the different requirements of the six parameters, some constraints need to be imposed after generation according to the formula. The number, size of the convolutional kernel, and training batch need to be rounded, while the learning rate does not. After initialization, the fitness value under the current parameters is fed back to the bee colony, and the position of the nectar source is updated. The search process is executed step by step until the optimal nectar source, that is, the optimal convolutional neural network structure parameters, are found. In this paper, the cross-entropy function is used to test the regression loss, and the fitness function formula is as follows:

[0067] y i j = f(f 1i , f 2i , s i , s 2i , batch i , η i )

[0068]

[0069] Among them, f uses the position of the i-th nectar source as the convolutional neural network structure parameters; y i j represents the fully connected result of the j-th group of input data in the i-th iteration and serves as the input to the SoftMax classifier; q i j represents the output classification label of the j-th group of input data in the i-th iteration; p represents the actual classification label of the training set; L i j represents the loss value of the j-th group in the i-th iteration. For example, assume there is a three-class classification problem, and the correct labels of the training set samples are (1, 0, 0), (0, 1, 0), (0, 0, 1). In the first iteration, the probability vector with a sum of 1 output by the third group of samples after passing through the SoftMax classifier is (0.096, 0.03, 0.874), and the actual classification label of the third group of samples is the third class. Therefore, the loss value calculation formula for the third group of sample data in the first iteration is as follows:

[0070] L 1 3 = -(0×lg0.096 + 0×lg0.03 + 1×lg0.874) ≈ 0.058

[0071] Step 9: Search process

[0072] Step1: Employed bees conduct neighborhood search

[0073] Employed bees are used to conduct neighborhood search. Since the upper and lower limits of the six parameters are not all the same, and there are both integers and decimals, some constraints need to be imposed after generating using the formula. The spatial position vector of the newly generated nectar source is used as the structural parameter of the convolutional neural network, and it is trained on the training set, the mean square error value is calculated, and the fitness value of each nectar source is recorded.

[0074] Step2: Position update

[0075] If the nectar source after search is better than the current nectar source, the current nectar source is replaced by the nectar source after search. If a certain nectar source after search is globally optimal, the original global optimal solution is replaced by this nectar source. If the fitness of the nectar source after search is inferior to that of the current nectar source, the invalid times of this nectar source are incremented by one.

[0076] Step3: Follow-up of follower bees

[0077] The follower bees select a nectar source to follow according to the obtained nectar source information according to the formula and execute Step2.

[0078] Step4: Global search of scout bees

[0079] Judge whether each nectar source has reached the maximum failure times. If it has reached the maximum failure times, the employed bees become scout bees and randomly generate a new nectar source according to the formula to replace the original nectar source.

[0080] Step5: End condition

[0081] Judge whether the maximum number of iterations has been reached. If the maximum number of iterations has not been reached, return to Step1 to continue the iteration; otherwise, end the global search.

[0082] Step 10: Fault classification

[0083] Since the spatial positions of each nectar source x i =(f 1i ,f 2i ,s i ,s 2i ,batch i ,η i ) are different, that is, the structural parameters of the convolutional neural network will change each time training is performed. For different structural parameters of the convolutional neural network, the same training set and test set are used for training and testing, where the training set has known fault category labels and the test set has unknown labels. The output layer of the convolutional neural network uses the SoftMax function to convert the output signal into a distribution with a sum of probabilities of 1, and the average loss value of the i-th iteration is calculated by the fitness function. Finally, after the bee colony searches and optimizes through N iterations, the generation with the smallest average loss value is found, and the corresponding globally optimal nectar source is found.

[0084] Step 11: Model reconstruction

[0085] After obtaining the global optimal nectar source x k =(f 1k , f 2k , s k , s 2k , batch k , η k ), use it to reconstruct the convolutional neural network with the six parameters of the optimal number of convolutional kernels, size, training batch, and learning rate, and then test it on the test set. At this time, the convolutional neural network fault diagnosis model has the highest classification accuracy.

Claims

1. A cement production line equipment fault diagnosis method based on FMEA and improved convolutional neural network, characterized in that: The following steps are involved: (1) The four rotating mechanical equipment, namely the rotary kiln, coal mill system, crusher and storage elevator, were evaluated according to the five aspects of downtime caused by failure, maintenance cost, quality, safety and environmental protection, and were rated A, B and C; (2) Use FMEA forms to analyze mechanical equipment failures, and conduct risk assessment and ranking based on the three evaluation criteria of failure severity S, occurrence O, and detection D. (3) Data collection and preprocessing to divide the data into training and test sets; (4) Build an improved CNN network and train it; (5) Select ABC-based optimization of CNN structural parameters, test the trained CNN model with the test set data, and obtain the final output result.

2. According to claim 1, a cement production line equipment fault diagnosis method based on FMEA and improved convolutional neural network is characterized in that: In step (2), the highest risk value is ranked 1, and then ranked backwards; the FMEA analysis table is obtained. When making improvements, the equipment corresponding to the fault mode with the highest ranking should be improved first; through the FMEA analysis table, the faults with the highest risk RPN ranking of each fault mode of the four types of equipment are obtained. The faults related to bearings, that is, the bearings have the greatest impact on various rotating mechanical equipment.

3. According to claim 1, a cement production line equipment fault diagnosis method based on FMEA and improved convolutional neural network is characterized in that: In step (3), based on the results of FMEA analysis, multiple rotating mechanical equipment on the production line of the cement production enterprise are monitored by installing measuring points, and the monitoring points are arranged on the bearings and related components of the six types of rotating mechanical equipment.

4. According to claim 1, a cement production line equipment fault diagnosis method based on FMEA and improved convolutional neural network is characterized in that: The specific preprocessing of step (3) is as follows: the acceleration data is restricted to the same magnitude and the same range and normalized; then, the preprocessed data is divided into a training set and a test set; and the one-dimensional time domain signal sequence is converted into a two-dimensional time-frequency diagram through data deformation, and the fault diagnosis problem is converted into a similar image recognition and classification task for processing.

5. According to claim 1, a cement production line equipment fault diagnosis method based on FMEA and improved convolutional neural network is characterized in that: In step (4), the improved CNNW network consists of a seven-layer convolutional neural network structure consisting of an input layer, two convolutional layers, two pooling layers, a fully connected layer and an output layer; pre-training is performed using a training set, and after pre-training, the relevant parameters of the network are optimized, and the above training operation is repeated after optimization until the ideal parameters are obtained, thereby obtaining a CNN model reconstructed with the optimal parameters.

6. The method for fault diagnosis of cement production line equipment based on FMEA and improved convolutional neural network according to claim 1 is characterized in that: Step (5) is as follows: the parameters to be optimized are the number f1 and size s1 of the convolution kernels in the first convolution layer, the number f2 and size s2 of the convolution kernels in the second convolution layer; the learning rate η and the training batch batch; these six parameters are mapped into the spatial vector position of each nectar source, and the position of the i-th nectar source is expressed as: x i =(f 1i ,f 2i ,s i ,s 2i ,batch i ,η i ) Among them, each nectar source is a feasible solution. The quality of the nectar source, that is, the quality of the feasible solution, is evaluated according to the amount of nectar in the nectar source, that is, the fitness value. Through continuous iterative update search, the optimal nectar source is found, which is the global optimal solution.

7. The method for fault diagnosis of cement production line equipment based on FMEA and improved convolutional neural network according to claim 6 is characterized in that: The output layer of the convolutional neural network uses the SoftMax function to convert the output signal into a distribution with a probability sum of 1, and the fitness function calculates the average loss value of the i-th iteration; finally, after N iterative searches, the bee colony finds the generation with the smallest average loss value, which corresponds to the global optimal nectar source.

8. A cement production line equipment fault diagnosis system based on FMEA and improved convolutional neural network, characterized in that: include: Grade module: used to evaluate the four rotating mechanical equipment, namely rotary kiln, coal mill system, crusher and storage elevator, according to the five aspects of downtime caused by failure, maintenance cost, quality, safety and environmental protection, and grade them into three grades: A, B and C; Evaluation module: used to analyze mechanical equipment using FMEA forms, and to conduct risk assessment and ranking based on the three evaluation criteria of fault severity S, occurrence O and detection D. Preprocessing module: used for data collection and preprocessing to divide training sets and test sets; CNN module: used to build and train improved CNN networks; ABC module: used to select CNN structural parameters based on ABC optimization, test the trained CNN model with test set data, and obtain the final output results.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the computer program is loaded into the processor, a cement production line equipment fault diagnosis method based on FMEA and improved convolutional neural network is implemented according to any one of claims 1 to 7.

10. A storage medium storing a computer program, characterized in that: When the computer program is executed by the processor, a cement production line equipment fault diagnosis method based on FMEA and improved convolutional neural network is implemented according to any one of claims 1 to 7.

Images