Intelligent fault diagnosis system for hydraulic device of cement roller press

Abstract

The application discloses an intelligent fault diagnosis system of a cement roller press hydraulic device, relates to the technical field of roller press monitoring, and comprises a data acquisition module, monitoring software and a fault diagnosis library; the data acquisition module monitors the related data of each component in the roller press hydraulic device in real time, and the monitored data is transmitted to the monitoring software; the data collected by the data acquisition module and the optimized algorithm are combined to establish a fault mode prediction of each component, the logical relationship between the fault phenomenon and the cause is inferred through an inference algorithm, the fault state is automatically classified, recognized and accurately judged to generate the fault diagnosis library; the data monitored by the subsequent data acquisition module is transmitted to the monitoring software, and the monitoring software generates statistical analysis reports and trend analysis; the application monitors the pressure in the hydraulic cylinder, judges the leakage in the hydraulic cylinder under actual working conditions, and has a good recognition ability, that is, the leakage in the hydraulic cylinder is effectively evaluated by using a BP neural network, and the accuracy is more than 90%.

Description

Technical Field

[0001] This invention relates to the field of roller press monitoring technology, and more specifically to an intelligent fault diagnosis system for the hydraulic device of a cement roller press. Background Technology

[0002] Roller presses are a new type of energy-saving cement grinding equipment developed internationally in the mid-1980s. They can replace the energy-intensive and inefficient ball mill pre-grinding system, reducing steel consumption and operating noise. Suitable for new plant construction and also for upgrading existing plants, they can increase system output by 30-50%. However, the hydraulic system in roller presses lacks a monitoring system, resulting in several problems:

[0003] (1) The hydraulic device has a high concentration of dust in the environment, which can easily cause serious oil pollution. The system also lacks effective means of monitoring and improving oil quality. Particulate matter clogs the damping holes and fitting clearances of key components, leading to a decline in component performance and system failure.

[0004] (2) Due to poor external heat dissipation conditions and lack of necessary oil level, oil temperature and pressure fluctuation monitoring and improvement methods, the system operating conditions change, resulting in soft faults such as the system actuators not working or long pressurization time. Due to the large number of components, the faults are difficult to diagnose in a short time.

[0005] (3) The hydraulic system of the roller press is a quantitative pressure limiting system. Due to uneven feeding and heavy load on site, internal leakage of hydraulic cylinders is caused, resulting in a short service life.

[0006] In view of the above-mentioned defects, the inventors of this invention have finally obtained this invention after a long period of research and practice. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent fault diagnosis system for the hydraulic device of a cement roller press, which solves the problem that the hydraulic device in the roller press lacks a monitoring system and cannot be effectively monitored.

[0008] This invention solves the aforementioned technical problems through the following technical solution: The invention includes a data acquisition module, monitoring software, and a fault diagnosis library. The data acquisition module monitors relevant data of various components in the hydraulic device of the roller press in real time, and the monitored data is transmitted to the monitoring software. Fault prediction models for each component are established based on the data collected by the data acquisition module and combined with optimization algorithms. Inference algorithms are used to infer the logical relationship between fault phenomena and causes, automatically classifying and accurately judging fault states to generate a fault diagnosis library. Subsequently, the data monitored by the data acquisition module is transmitted to the monitoring software, which generates statistical analysis reports and trend analyses, and compares them with parameters in the fault diagnosis library to achieve remote fault analysis and fault prediction.

[0009] Preferably, the data acquisition module involves installing corresponding monitoring sensors on each component of the hydraulic device of the roller press according to the parameters to be monitored.

[0010] Preferably, the fault diagnosis system includes monitoring of internal leakage in the hydraulic cylinder, monitoring of internal leakage in the hydraulic cylinder by a data acquisition module, and establishing a fault prediction model for internal leakage in the hydraulic cylinder by combining optimization algorithms. Multiple pressure sensors are used to monitor the hydraulic pressure in the rod chamber and rodless chamber of the hydraulic cylinder.

[0011] Preferably, a hydraulic cylinder internal leakage fault prediction model is established based on least squares fitting:

[0012] Establish a set of experimental data (X) i ,Y i The corresponding functional relationship between the independent and dependent variables; let the error of the experimental data be represented by a vector: δ=(δ0,δ1,…,δ m ) T To minimize the vector δ-norm ||δ||, the Euler norm ||δ||² is used; thus, the resulting fitting function is the least squares method.

[0013] Least squares fitting in a certain function space Find a function y = Φ * (x), which minimizes the sum of squared errors, i.e.:

[0014]

[0015] Assigning a certain weight to the error vector δ, its weighted norm form is:

[0016] Among them, the function system Linearly independent on the interval [a,b] containing node {xi}, any function in Φ It can be represented as:

[0017]

[0018]

[0019] Taking the partial derivative of each unknown parameter in the prediction equation and setting it equal to zero yields the result for a0, a1, ..., a n A system of equations, namely:

[0020]

[0021] Rearranging the above equations, we obtain the prediction partial derivative equations:

[0022]

[0023] Let vector express The function value at each node is a vector of components, in the form:

[0024]

[0025] Vector f is represented as y0, y1, ..., y i A vector of the form:

[0026] f = (y0, y1, ..., y m ) T

[0027] Let the vector dot product be:

[0028]

[0029] Therefore, the prediction partial derivative equation can be expressed in matrix form:

[0030]

[0031] This matrix is ​​a system of regular equations derived by the least squares method, a0, a1, ..., a n The unknown quantity is ρ(x). Since the dot product of vectors is symmetric, the coefficient matrix of this system of linear equations is clearly a symmetric matrix. When the function space is taken as a power function space, the weighted vector ρ(x) i If ) = 1, then the system of regular equations can be rewritten as:

[0032]

[0033] The fitted curve is:

[0034]

[0035] Therefore, by using second-order or higher polynomial curves for multiple fittings, the hydraulic cylinder internal leakage fault prediction model is as follows:

[0036]

[0037] Where x is the pressure difference between the rod-side chamber and the rodless chamber.

[0038] Preferably, a BP neural network is used to identify the leakage in the hydraulic cylinder. This requires establishing a BP neural network model and training the model to obtain a fault diagnosis library for hydraulic cylinder leakage.

[0039] Preferably, the BP neural network contains three types of nodes, located in the input layer, output layer, and hidden layer, respectively. The output layer and hidden layer require activation functions to process the data input to these nodes; the non-linear activation functions include the sigmoid function and the purelin function. The sigmoid function is further divided into the Log-sigmoid function and the Tan-sigmoid function. The expression for the Log-sigmoid function is:

[0040]

[0041] Taking the derivative with respect to the independent variable x, we get:

[0042]

[0043] The expression for the purelin function is:

[0044] S(x) = x.

[0045] Preferably, a BP neural network model is established:

[0046] Using the sum of squared errors as the loss function, the weight update formula is derived, and its sum of squared errors E(w) is defined as:

[0047]

[0048] In the above formula, e(n) represents the error, and its expression is: The actual output can be viewed as a function of the weight w. Taking the partial derivative of both sides of the above equation with respect to w, we can obtain the error partial derivative equation:

[0049]

[0050] Find the partial derivative of e(n) with respect to w:

[0051]

[0052] Substituting the above equation into the error partial derivative equation, we get:

[0053]

[0054] The change in weights Δw in the BP neural network is:

[0055]

[0056] Where η is the learning rate; therefore, the weight update formula can be written as:

[0057]

[0058] Substituting the Log-sigmoid function into the equation, we obtain the final weight update formula:

[0059]

[0060] Where t k For the expected output value, o k Let X be the ideal output value, and W be the input value. j Let be the weight vector of the j-th layer node, and let be the threshold value of -1 added to the end of both the input vector X and the weight vector.

[0061] Preferably, when using a BP neural network to perform pattern recognition on a target, the target vector is often represented by 0 and 1 encoding.

[0062] The hidden layer nodes of the BP neural network use the Log-sigmoid function as the activation function to map the input signal to the range of 0-1, and the output layer function uses the purelin function to achieve equivalent output.

[0063] The number of hidden layer nodes is:

[0064]

[0065] In the formula, m is the number of input layer nodes and n is the number of output layer nodes.

[0066] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0067] 1. The data acquisition module can monitor the relevant data of each component in the hydraulic device of the roller press in real time, and optimize the monitoring data to predict fault models and generate a fault diagnosis library. When monitoring each component in the future, the data obtained by the data acquisition module is compared with the parameters in the fault diagnosis library, thereby realizing remote fault analysis and fault prediction. This can effectively monitor various parameters in the hydraulic device of the roller press and quickly diagnose faulty components.

[0068] 2. The pressure inside the hydraulic cylinder is monitored, and the leakage situation inside the hydraulic cylinder under actual working conditions is evaluated with an accuracy rate of over 90%. It has good identification ability, that is, the BP neural network is used to effectively evaluate the leakage situation inside the hydraulic cylinder. Attached Figure Description

[0069] Figure 1This is a distribution map of monitoring points in Example 1;

[0070] Figure 2 This is a graph showing the relationship between the internal leakage coefficient and the pressure difference between the two rod chambers. Detailed Implementation

[0071] The above-mentioned and other technical features and advantages of the present invention will be described in more detail below with reference to the accompanying drawings.

[0072] Example 1

[0073] This embodiment provides a technical solution: an intelligent fault diagnosis system for the hydraulic device of a cement roller press, including a data acquisition module, monitoring software, and a fault diagnosis library;

[0074] The data acquisition module monitors relevant data from various components in the hydraulic system of the roller press in real time, and the monitored data is transmitted to the monitoring software; for components whose fault status can be measured by sensors, such as... Figure 1 The diagram shows the distribution of monitoring points in this embodiment. Each monitoring point uses a corresponding sensor to monitor and obtain corresponding parameters. The parameters to be monitored for hydraulic station 1 include: oil temperature, liquid level, contamination level, viscosity, water content, and density; the parameters to be monitored for motor 2 include: vibration, shell temperature, and speed; the parameters to be monitored for pump 3 include: vibration and pump body temperature; the parameters to be monitored for the main pipelines 10 of the two hydraulic cylinders include: temperature and vibration; in addition, the pressure of overflow valve 4, the two side pipelines 5, accumulator 6, accumulator nitrogen filling 7, the rod chamber 8 of the hydraulic cylinder, and the rodless chamber 9 of the hydraulic cylinder needs to be monitored (only the components that need to be monitored are shown in the figure, and the number of components is not limited); the data obtained from the monitoring can be provided with a data interface to connect to a third-party monitoring platform for real-time monitoring, enabling monitoring on PC and mobile devices, and has alarm function, trend analysis function, and historical data query function.

[0075] Information that cannot be directly measured by sensors is used to build fault models through mathematical algorithms, such as predicting faults like internal leakage in hydraulic cylinders and abnormal conditions of hydraulic pumps. Fault models for each component are built by collecting data through data acquisition modules and combining them with optimization algorithms. The logical relationship between fault phenomena and causes is inferred through reasoning algorithms, and fault states are automatically classified, identified, and accurately judged to generate a fault diagnosis library.

[0076] The data collected by the subsequent data acquisition module is transmitted to the monitoring software. The monitoring software generates statistical analysis reports and trend analysis, and compares them with the parameters in the fault diagnosis database. It automatically classifies and accurately judges the fault status, realizes remote fault analysis and fault prediction, and finally provides maintenance and fault solutions.

[0077] The monitoring software displays curves for various parameters, along with the start and end times of each curve. Time selection buttons allow users to access data for specific periods, facilitating data analysis and printing of the various curves generated during the process. Furthermore, the historical curve interface clearly displays all monitored data in list and curve formats, categorizing and accurately displaying the data values. Visual C# controls are used to easily achieve this. The software also records fault information during operation, displaying each current and previous fault in a list format with real-time updates. Considering users may retain many fault records, the program utilizes database technology, employing ADO.NET in C# for dynamic database operations. The database is pre-built using SQL Server, leveraging its powerful features to easily pre-define and manipulate the database's format, data types, and content.

[0078] Example 2

[0079] This embodiment is a further optimization based on Embodiment 1. The parts that are the same as those in the aforementioned technical solutions will not be repeated here. Furthermore, in order to better realize the present invention, the following configuration is adopted: the fault diagnosis system includes monitoring the internal leakage of the hydraulic cylinder, the data acquisition module monitors the internal leakage of the hydraulic cylinder, and the hydraulic cylinder internal leakage fault prediction model is established by combining optimization algorithms. Multiple pressure sensors are used to monitor the hydraulic pressure in the rod chamber and the rodless chamber of the hydraulic cylinder.

[0080] Establish a set of experimental data (X) i ,Y i The corresponding functional relationship between the independent and dependent variables; experimental data will have certain errors, therefore it is not required that the established F(x) pass through all data points, only that the error is minimized according to a certain rule. Therefore, it is necessary to establish a prediction model for the internal leakage of the hydraulic cylinder based on least squares fitting; let its error be represented by a vector as: δ=(δ0,δ1,…,δ m ) T The above requirement is to minimize the vector δ norm ||δ|| according to a certain rule. For ease of calculation, the Euler norm ||δ||2 is often used; that is, the resulting fitting function is the least squares method.

[0081] Least squares fitting needs to be implemented in a certain function space. Find a function y = Φ * (x), which minimizes the sum of squared errors, i.e.:

[0082]

[0083] Considering the different weights of the data, the error vector δ is assigned a certain weight, and its weighted norm form is as follows:

[0084] Among them, the function system Linearly independent on the interval [a,b] containing node {xi}, any function in Φ It can be represented as:

[0085]

[0086]

[0087] When the combination coefficients a0, a1, ... a n Different selected values ​​correspond to different function classes.

[0088] In this embodiment, Φ = P n =Span(1,x) 2 ,……,x n The function subspace generated by ) yields a polynomial fitting curve.

[0089] Based on the above analysis, the sum of squared errors of the curve is essentially a0, a1, ..., a n Therefore, the data fitting problem is transformed into finding the multivariate function G(a0, a0…a…). n The problem of finding the minimum value can be solved by taking the partial derivative of each unknown parameter in the prediction equation and setting it equal to zero, thus obtaining the value for a0, a1, ... a n A system of equations, namely:

[0090]

[0091] Rearranging the above equations, we obtain the prediction partial derivative equations:

[0092]

[0093] Let vector express The function value at each node is a vector of components, in the form:

[0094]

[0095] Vector f is represented as y0, y1, ..., y i A vector of the form:

[0096] f = (y0, y1, ..., y m ) T (7)

[0097] Let the vector dot product be:

[0098]

[0099] Therefore, Equation 5 can be written in matrix form:

[0100]

[0101] This matrix is ​​a system of regular equations derived by the least squares method, a0, a1, ..., a n The unknown quantity is ρ(x). Since the dot product of vectors is symmetric, the coefficient matrix of this system of linear equations is clearly a symmetric matrix. When the function space is taken as a power function space, the weighted vector ρ(x) i If ) = 1, then the system of regular equations can be rewritten as:

[0102]

[0103] At this point, the fitted curve is:

[0104]

[0105] according to Figure 2 The curve trend indicates that as the internal leakage coefficient increases, the pressure difference between the two rod chambers decreases, which is consistent with the theoretical analysis. This trend is nonlinear, so a second-order or higher polynomial curve is used for multiple fittings.

[0106] Through multi-order fitting analysis, the fitting error no longer decreases when the order exceeds 5; therefore, a 5th-order polynomial can better express the relationship between the internal leakage coefficient and the pressure difference; the final hydraulic cylinder internal leakage fault prediction model is as follows:

[0107]

[0108] Among them, x is the pressure difference of the hydraulic pressure in the rod-side chamber and the rodless chamber.

[0109] In actual operating conditions, hydraulic system failures are caused by a combination of multiple factors. For hydraulic cylinder failures, internal leakage causes pressure variations in the two rod chambers as the piston rod moves through different positions. As the leakage increases, the pressure difference between the rod-side and rodless chambers also increases. In actual operating conditions, the piston rod is subjected to a constant load, therefore the pressure difference between the two chambers should be a constant value. However, simply monitoring the pressure difference between the rod-side and rodless chambers of the hydraulic cylinder is insufficient for accurate diagnosis of internal leakage. Factors such as load fluctuations, system temperature changes, decreased oil quality leading to increased oil particles, and valve spool blockage can also produce symptoms similar to those of internal leakage failures in hydraulic cylinders. In conclusion, this paper simplifies the actual problem to establish a hydraulic cylinder leakage prediction model, making the model more consistent with practical engineering applications and possessing good practicality and feasibility.

[0110] This embodiment uses comparative analysis to add certain prerequisites to the hydraulic cylinder fault model, ensuring that the analysis and diagnosis of the hydraulic cylinder internal leakage mode are performed in accordance with the following principles:

[0111] (1) The temperature and oil quality of the hydraulic device are within the normal range.

[0112] (2) The system pressure increment is close to the pressure increment of the high-pressure chamber of the faulty cylinder, but does not exceed the set alarm threshold.

[0113] (3) For the hydraulic cylinders connected to the same multi-way proportional valve, when the deviation of the pressure difference between the rodless and non-rod chambers of a certain hydraulic cylinder increases while the other cylinders are within the normal range.

[0114] When a backpropagation (BP) neural network is used for pattern recognition, the network contains three types of nodes, located in the input layer, output layer, and hidden layers. Except for the input layer, each layer requires an activation function to process the data input to that node. Commonly used non-linear activation functions include the sigmoid function and the purelin function. The sigmoid function is further divided into the Log-sigmoid function and the Tan-sigmoid function. The expression for the Log-sigmoid function is:

[0115]

[0116] Taking the derivative with respect to the independent variable x, we get:

[0117]

[0118] The expression for the purelin function is:

[0119] S(x)=x (16)

[0120] Establishing a BP neural network model:

[0121] The learning rule for a backpropagation (BP) neural network uses the sum of squared errors as the loss function to derive the weight update formula; its sum of squared errors E(w) is defined as:

[0122]

[0123] Wherein, e(n) represents the error, and its expression is: The actual output can be viewed as a function of the weight w. Taking the partial derivative of both sides of the above equation with respect to w, we can obtain the error partial derivative equation:

[0124]

[0125] Find the partial derivative of e(n) with respect to w:

[0126]

[0127] Substituting the above equation into equation 18, we get:

[0128]

[0129] The core idea of ​​BP neural network weight update is gradient descent, that is, the change in weight Δw is:

[0130]

[0131] Where η is the learning rate; therefore, the weight update formula can be written as:

[0132]

[0133] Substituting the Log-sigmoid function into the equation, we obtain the final weight update formula:

[0134]

[0135] Where t k For the expected output value, o k Let X be the ideal output value, and W be the input value. j Let be the weight vector of the j-th layer node, and let be the threshold value of -1 added to the end of both the input vector X and the weight vector.

[0136] Training the BP neural network:

[0137] The input vector of the BP neural network is a feature quantity representing different degrees of internal leakage in the hydraulic cylinder. Based on actual working conditions, the piston rod is subjected to a constant load during its forward movement, resulting in pressure fluctuations during the experiment. In this embodiment, the pressure difference between the rodless and rod-side chambers is used as a feature quantity. The hydraulic cylinder is tested using a prototype to operate under different leakage coefficients. Pressure signals from the rodless and rod-side chambers are collected at ten key locations and the difference is calculated, which is X. Four types of internal leakage are defined based on the leakage coefficient: zero leakage, small internal leakage, medium internal leakage, and large internal leakage, each corresponding to a specific leakage coefficient. The leakage coefficients in this embodiment are shown in the table below.

[0138] Small internal leak ≤0.2 ≤2％ Internal leakage 0.2-0.5 2％-7％ Internal leak 0.5-1 7％-10％

[0139] Adjust the leakage level of the hydraulic cylinder, divide the piston rod's forward stroke into 8 positions, and record the differential pressure signal at each position as X1, X2, X3, X4, X5, X7, X8; record 8 differential pressure signals collected for each forward stroke of the piston rod as one sample, set four leakage states: zero leakage, small internal leakage, medium internal leakage, and large internal leakage, and record 10 samples for each state; randomly divide the samples collected for the same state into two groups to establish a training set and a test set for the neural network model.

[0140] When using a backpropagation (BP) neural network for pattern recognition, the target vector is often represented by 0 and 1, where 1 corresponds to different valid states at different positions; for example, (1, 0, 0, 0). TThe normal state is represented by the data, and the state information is digitized to evaluate the four states of the hydraulic cylinder. For ease of representation, y1, y2, y3, and y4 represent the outputs of each node in the network. The expected output values ​​are shown in the table below:

[0141]

[0142]

[0143] The hidden layer nodes of the BP neural network use the Log-sigmoid function to map the input signal to the range of 0-1, and the output layer function uses the purelin function to achieve an equivalent output.

[0144] The number of hidden layer nodes should be neither too many nor too few. Too few nodes will make weight adjustment difficult and reduce recognition accuracy; too many layers and nodes will not only affect training speed and increase training time, but also easily lead to overfitting. An empirical formula is used for setting hidden layer nodes:

[0145]

[0146] In the formula, m is the number of input layer nodes and n is the number of output layer nodes.

[0147] Analysis of BP neural network recognition results:

[0148] In this embodiment, the collected samples are used to train the BP neural network model, keeping the weights and threshold matrices of each layer unchanged, and pattern recognition is performed on the test samples; the training results of the BP neural network model are shown in the table below:

[0149]

[0150]

[0151] The recognition results show that the model can accurately identify hydraulic cylinders when the threshold is set to 0.5. Specifically, if the output layer result is greater than 0.5, it is considered a correct identification; otherwise, it is considered an incorrect identification. Using this model to evaluate internal leakage in hydraulic cylinders under actual working conditions, the accuracy rate exceeds 90%, demonstrating excellent recognition capabilities. Therefore, a BP neural network can be used to effectively assess internal leakage in hydraulic cylinders.

[0152] The above description is merely a preferred embodiment of the present invention and is illustrative rather than restrictive. Those skilled in the art will understand that many changes, modifications, and even equivalents can be made within the spirit and scope defined by the claims of the present invention, all of which will fall within the protection scope of the present invention.

Claims

1. An intelligent fault diagnosis system for the hydraulic device of a cement roller press, characterized in that, It includes a data acquisition module, monitoring software, and a fault diagnosis library. The data acquisition module monitors the relevant data of each component in the hydraulic device of the roller press in real time, and the monitored data is transmitted to the monitoring software. Based on the data collected by the data acquisition module and combined with the least squares method, a fault model prediction model for each component is established. Through inference algorithms, the logical relationship between fault phenomena and causes is inferred, and the fault status is automatically classified, identified, and accurately judged to generate a fault diagnosis library. Subsequently, the data monitored by the data acquisition module is transmitted to the monitoring software, which generates statistical analysis reports and trend analysis, and compares them with the parameters in the fault diagnosis library to realize remote fault analysis and fault prediction. Data acquisition module: Based on the parameters to be monitored, corresponding monitoring sensors are installed on each component of the hydraulic device of the roller press; The fault diagnosis system includes monitoring of internal leakage in hydraulic cylinders, monitoring of internal leakage in hydraulic cylinders by a data acquisition module, and establishing a fault prediction model for internal leakage in hydraulic cylinders by combining the least squares method. Multiple pressure sensors are used to monitor the hydraulic pressure in the rod chamber and rodless chamber of the hydraulic cylinder. A hydraulic cylinder internal leakage fault prediction model was established based on least squares fitting: Establish a set of experimental data ( The corresponding functional relationship between the independent and dependent variables; let the error of the experimental data be represented by a vector: ; Requires vector norm To achieve the minimum, the Euler norm is used. The resulting fitted function is the least squares method. Least squares fitting in a certain function space Find a function This minimizes the sum of squared errors. ； Error vector Assigning a certain weight, its weighted norm form; Among them, the function system The nodes are linearly independent on the interval [a,b] containing the node {xi}. any function It can be represented as: ； ； Taking the partial derivative of each unknown parameter in the prediction equation and setting it equal to zero yields the result regarding... A system of equations, namely: ； Rearranging the above equations, we obtain the prediction partial derivative equations: ； Let vector express The function value at each node is a vector of components, in the form: ； vector Indicates to A vector of the form: ； Let the vector dot product be: ， ； Therefore, the prediction partial derivative equation can be written in matrix form: ； This matrix represents a system of regular equations derived using the least squares method. The unknowns are: Since the dot product of vectors is symmetric, the coefficient matrix of this system of linear equations is clearly symmetric; when the function space is taken as a power function space, the weighted vectors... Then the canonical system of equations can be rewritten as: ； The fitted curve is: ； Therefore, by using second-order or higher polynomial curves for multiple fittings, the hydraulic cylinder internal leakage fault prediction model is as follows: ； in, This represents the pressure difference between the hydraulic fluid in the rod-side chamber and the rodless chamber.

2. The intelligent fault diagnosis system for the hydraulic device of a cement roller press as described in claim 1, characterized in that, To identify internal leakage in hydraulic cylinders using a BP neural network, a BP neural network model needs to be established and trained to obtain a fault diagnosis library for internal leakage in hydraulic cylinders.

3. The intelligent fault diagnosis system for the hydraulic device of a cement roller press as described in claim 2, characterized in that, A backpropagation (BP) neural network contains three types of nodes, located in the input layer, output layer, and hidden layer. The output and hidden layers require activation functions to process the data input to these nodes; these functions are non-linear and include the sigmoid function and the purelin function. The sigmoid function is further divided into Log-sigmoid and Tan-sigmoid functions. The expression for the Log-sigmoid function is as follows:

4. Regarding the independent variable Taking the derivative, we get: ； The expression for the purelin function is: 。 5. The intelligent fault diagnosis system for the hydraulic device of a cement roller press as described in claim 3, characterized in that, Establish a BP neural network model: Using the sum of squared errors as the loss function, the weight update formula is derived, and its sum of squared errors is defined. for: ； In the above formula For the error, its expression is: The actual output can be viewed as weights. A function, both sides of the above equation Taking the partial derivatives, we can obtain the error partial derivative equation: ； beg about Partial derivative: ； Substituting the above equation into the error partial derivative equation, we get: ； Changes in weights of a BP neural network for: ； Where, in the formula The learning rate is used; therefore, the weight update formula can be written as: ； Substituting the Log-sigmoid function into the equation, we obtain the final weight update formula: ； Where, in the formula For the expected output value, For the ideal output value, For input quantity, Let j be the weight vector of the node at layer j, and the input vector. A threshold of -1 is added to the end of each vector.

6. The intelligent fault diagnosis system for the hydraulic device of a cement roller press as described in claim 2, characterized in that, When using a BP neural network to perform pattern recognition on a target, the target vector is often represented by 0 and 1 encoding. The hidden layer nodes of the BP neural network use the Log-sigmoid function as the activation function to map the input signal to the range of 0-1, and the output layer function uses the purelin function to achieve equivalent output. The number of hidden layer nodes is: ； Where, in the formula The number of nodes in the input layer. This represents the number of nodes in the output layer.

Images