Air compressor remote fault diagnosis system based on CNN and type-2 fuzzy algorithm

Abstract

The invention relates to an air compressor remote fault diagnosis system and method based on a CNN and a type-2 fuzzy algorithm, and belongs to the field of intelligent recognition, and the system comprises a signal acquisition system, a data transmission system, a remote monitoring system and the CNN and type-2 fuzzy algorithm. According to the system, parameters such as pressure, temperature, rotating speed, vibration frequency and amplitude of the air compressor are monitored in real time through the sensor and transmitted to the server for processing, fault diagnosis and prediction are conducted through the CNN and the type-2 fuzzy algorithm, and the function of remotely monitoring and diagnosing the faults of the air compressor is achieved. In this way, the feature extraction capability can be better improved, so that the classification performance is improved, and the accuracy of fault diagnosis is improved. Meanwhile, when sensor noise causes fluctuation of fault features, the type-2 fuzzy algorithm can adjust a classification decision according to an uncertainty range of a fuzzy membership function, so that interference of noise on fault classification is reduced, and better robustness is achieved.

Description

Technical Field

[0001] The present application relates to an air compressor remote fault diagnosis system and a diagnosis method based on CNN and type-2 fuzzy algorithm, belonging to the technical field of intelligent control. Background Technique

[0002] Traditional air compressor fault diagnosis systems rely on predefined rule sets to identify fault patterns, but may not be flexible and accurate enough for fault diagnosis of complex systems. Data-driven machine learning methods can learn fault patterns from a large amount of data, but require a large amount of labeled data and complex feature engineering. However, there are still some problems in the existing air compressor fault diagnosis.

[0003] Firstly, traditional methods often fail to accurately capture the fault patterns of the complex air compressor system, resulting in low diagnostic accuracy. Secondly, the existing technology does not support real-time monitoring and remote diagnosis sufficiently, and cannot detect and solve faults in time. In addition, traditional methods have limited capabilities in processing and analyzing large-scale data and are difficult to cope with the increasing data volume and complexity.

[0004] Therefore, a more efficient and accurate air compressor fault diagnosis technology is needed to improve the reliability and operation efficiency of the equipment. Summary of the Invention

[0005] In industrial production, the stable operation of the air compressor plays a crucial role in the entire production process. Traditional air compressor fault diagnosis methods have many defects and are difficult to meet the high requirements of modern industry for equipment reliability and operation efficiency. The air compressor remote fault diagnosis system based on CNN and type-2 fuzzy algorithm proposed in the present application brings new ideas and methods to solve these problems.

[0006] To achieve the above object, the solution of the present invention is as follows: An air compressor remote fault diagnosis system based on CNN and type-2 fuzzy algorithm, the diagnosis system includes a sensor group, a CNN module, an interval type-2 fuzzy inference subsystem, a result output module, and an alarm module.

[0007] Control flow between modules: In this system, the sensor group collects the operating data of the air compressor and transmits it to the CNN module to achieve preliminary information processing. The CNN module extracts features and classifies system-level faults from the input data, identifies potential system-level faults, and outputs the diagnostic results. Then, the interval type-2 fuzzy inference subsystem combines the output of the CNN module with the original data for more detailed component-level fault diagnosis to determine the specific fault type. The result output module is responsible for presenting the diagnostic results to the user in an intuitive visual form for easy understanding and decision-making. If a fault is detected in the system, the alarm module compares the fault diagnosis result with a preset threshold and triggers an alarm in a timely manner to remind the user to take necessary maintenance measures.

[0008] Control principle: The sensor group converts the operating state of the air compressor into electrical signals through physical or chemical effects, and these signals provide the necessary data basis for subsequent modules. The CNN module uses deep learning techniques such as convolution, pooling, and fully connected layers to extract key features and classify system-level faults, ensuring the accuracy and efficiency of fault detection. The interval type-2 fuzzy inference subsystem accurately judges the fault type through steps such as fuzzification, inference, defuzzification, and type reduction, thereby improving the reliability of diagnosis. The result output module clearly displays the diagnostic results using visualization tools, enabling users to quickly understand the fault situation. The alarm module is responsible for real-time monitoring of the fault status. By comparing the diagnostic results with a preset safety threshold, it issues an alarm in a timely manner to urge the user to take corresponding maintenance actions.

[0009] Through this modular design, the system is efficient and accurate. It can not only accurately diagnose the faults of centrifugal air compressors but also actively remind users based on real-time detection conditions to ensure the timely implementation of maintenance measures, thereby maximizing the normal operation of the equipment and extending its service life.

[0010] As an improvement of the present invention, the sensor group consists of a temperature sensor, a pressure sensor, and a vibration sensor, which are used to collect the operating data of the centrifugal air compressor. The data includes air filtration resistance (X1), first-stage shaft vibration (X2), second-stage shaft vibration (X3), third-stage shaft vibration (X4), lubricating oil temperature (X5), lubricating oil pressure (X6), oil filter pressure difference (X7), bearing temperature at the drive end (X8), motor current (X9), exhaust pressure (X10), exhaust flow rate (X11), ambient temperature (X12), inlet pre-cooler temperature (X13), and outlet pre-cooler temperature (X14).

[0011] As an improvement of the present invention, the input end of the CNN module is connected to a sensor group, which normalizes the collected 14-dimensional raw sample data to avoid the influence of different dimensions on the extraction of fault features. After operations of the convolutional layer and the pooling layer, feature vectors are extracted before the fully connected layer, and the system faults are classified into four types of system-level faults: normal, mechanical system fault, oil circuit system fault, and gas circuit system fault. Then, the interval type-2 fuzzy inference subsystem outputs 15-dimensional fault results (Y1 - Y15).

[0012] As an improvement of the present invention, the diagnostic system is provided with multiple interval type-2 fuzzy inference subsystems, which specifically correspond to the mechanical, oil circuit, and gas sub-systems of the air compressor. Each subsystem independently receives the corresponding feature vectors output by the CNN module, and through modules such as its own fuzzifier, inference engine, rule base, defuzzifier, and demodulator for comprehensive processing, and finally outputs a vector composed of accurate values to clearly judge the fault type, specifically as follows:

[0013] Mechanical interval type-2 fuzzy inference system: This system receives 3-dimensional inputs of the vibration of the first-stage shaft (X2), the vibration of the second-stage shaft (X3), and the vibration of the third-stage shaft (X4), and outputs 4-dimensional results (dynamic balance imbalance of the first-stage rotor or impeller, dynamic balance imbalance of the second-stage rotor or impeller, dynamic balance imbalance of the third-stage rotor or impeller, surge) to identify the fault characteristics of the mechanical part.

[0014] Oil circuit interval type-2 fuzzy inference system: This system processes 5-dimensional inputs of the lubricating oil temperature (X5), the lubricating oil pressure (X6), the oil filter pressure difference (X7), the ambient temperature (X12), and the temperature of the inlet pre-cooler (X13), and outputs 5-dimensional results (insufficient oil volume, oil filter blockage, too low ambient temperature, oil cooler failure, outlet oil pipe blockage) to analyze the relevant fault conditions of the oil circuit.

[0015] Gas circuit interval type-2 fuzzy inference system: This system focuses on 6-dimensional inputs of the air filtration resistance (X1), the motor current (X9), the exhaust pressure (X10), the ambient temperature (X12), the temperature of the inlet pre-cooler (X13), and the temperature of the outlet pre-cooler (X14), and outputs 6-dimensional results (intermediate cooler failure, pre-cooler failure, too high production ambient temperature, air filter blockage, air separation pipe blockage, power grid fluctuation) for accurate judgment of gas circuit faults.

[0016] Through this modular design, each subsystem can independently and efficiently perform fault diagnosis in its respective field, thus ensuring the overall accuracy and response speed of the entire air compressor remote fault diagnosis system, and is a powerful tool for realizing refined management and maintenance.

[0017] As an improvement of the present invention, the CNN module includes:

[0018] At least one convolutional layer with adaptively adjusted convolutional kernel parameters, which is used to extract features from the input data and capture local feature patterns in the data through convolutional operations;

[0019] At least one pooling layer, which uses max pooling or average pooling methods to perform dimensionality reduction on the output of the convolutional layer, reducing the computational amount and retaining the main features;

[0020] A fully connected layer, whose input end is connected to the output of the pooling layer or the feature map after flattening. The extracted feature vectors are mapped to a 15-dimensional fault result space through a weight matrix to achieve fault classification. And the weights of the fully connected layer are optimized and updated through the backpropagation algorithm during the training process. The training data includes a large number of labeled centrifugal air compressor fault samples and normal operation samples.

[0021] As an improvement of the present invention, the components of the interval type-2 fuzzy inference subsystem include:

[0022] Fuzzifier: Using trapezoidal function, z-type function (zmf), s-type function (smf), Gaussian function (gaussmf), double Gaussian mixture function (gauss2mf) as membership functions, it fuzzifies the input feature vector according to predefined parameters to determine the degree to which each input feature belongs to different fuzzy sets. And the membership function parameters of the fuzzifier are adjusted and optimized according to historical fault data and expert experience; for example, first, the input variables and value ranges need to be determined. For X10, X10 is the exhaust pressure, and three fuzzy sets are defined for it, namely L, N, and H. Each of these three fuzzy sets corresponds to an upper membership function and a lower membership function respectively. The actual value of the input variable is mapped to the corresponding membership function to obtain the membership degree interval of this value under different fuzzy sets.

[0023] Inference Engine: Based on the Mamdani inference structure or other inference structures suitable for fault diagnosis, it performs inference calculations according to the membership degrees of input features and the fuzzy rules extracted, simplified and merged from fault causes and the system fault tree, generating an output fuzzy set. The construction of the fuzzy rule base covers in-depth analysis of the causal relationships between various fault modes of centrifugal air compressors and sensor data features. First, it is necessary to systematically study the common fault modes of centrifugal air compressors, such as surge, bearing faults, and impeller wear, etc. Analyze the causal relationships between each fault mode and sensor data features. For example, the surge fault may be closely related to sensor data features such as large pressure fluctuations and abnormal flow rates. Based on the above causal relationships, extract corresponding fuzzy rules, such as: "If the pressure fluctuation is large and the flow rate is small, then a surge fault may occur." These rules will serve as the basis for inference in subsequent calculations. Simplify and merge rules: To improve the simplicity and efficiency of the rule base, simplify and merge the extracted fuzzy rules to remove redundant rules. Effective simplification and merging can reduce the computational complexity and improve the inference speed. Then match the membership degrees of input features with the rules in the fuzzy rule base. For example, if the membership degree of "large pressure fluctuation" corresponding to the current pressure data is 0.8, and the membership degree of "small flow rate" corresponding to the flow rate data is 0.6, then it is matched with the rule "If the pressure fluctuation is large and the flow rate is small, then a surge fault may occur." Next, calculate the rule activation strength. For the matched rules, calculate the activation strength of the rules according to the membership degrees of input features. In Mamdani inference, the method of taking the minimum value is usually adopted. In the above example, the activation strength of this rule is min(0.8, 0.6) = 0.6. Finally, generate the output fuzzy set according to the activated rules and their activation strengths. For each possible fault output (such as surge, bearing fault, etc.), determine its membership degree in the output fuzzy set according to the relevant activated rules and activation strengths. Suppose there are multiple rules related to the surge fault, calculate their activation strengths respectively, and then take the maximum value or perform other synthesis operations as the membership degree of the surge fault in the output fuzzy set.

[0024] Type-reducer: It adopts the EKM (enhanced Karnik-mendel) algorithm to convert the interval type-2 fuzzy set output by the inference engine into a type-1 fuzzy set through an efficient iterative calculation process. First, relevant parameters in the EKM algorithm need to be determined, such as the iteration termination condition. Generally, the iteration termination condition can be set as the difference between the results of two adjacent iterations being less than a certain minimum value, or the maximum number of iterations can be set, such as 100 times. Then, the centroid of the upper and lower boundaries is calculated. First, the domain of the output interval type-2 fuzzy set is discretized. Assuming the domain is, it can be equally divided into n points, that is,. Then, the centroid of the upper boundary and the centroid of the lower boundary are calculated according to the centroid formula respectively. The centroids of the upper and lower boundaries obtained from the first calculation are used as the initial values, and then iteration is carried out. In each iteration, according to the current centroids of the upper and lower boundaries, the centroids of the upper and lower boundaries are recalculated. If the number of iterations reaches the maximum number of iterations, the iteration stops; otherwise, it returns to the previous step and continues the iteration. It has high computational efficiency and accuracy when dealing with complex membership functions and dynamically changing fuzzy sets, and the calculation process of the type-reducer can be parallelized according to the real-time performance requirements of the system;

[0025] Defuzzifier: Its main function is to convert the output fuzzy set into specific fault diagnosis results, such as determining the probability of the current fault being surge or its severity, etc., so as to provide clear operation guidance for fault repair. To achieve efficient and accurate defuzzification, a method combining the centroid of area method and the threshold method is adopted. The specific steps are as follows:

[0026] First, the weighted average of the fuzzy vector is taken as the preliminary crisp value output by the centroid of area method, and then, according to the set real threshold λ∈[0,1], the corresponding values with membership function values greater than λ are selected as the final crisp value output. If there are multiple, the maximum value is taken as the final fault diagnosis result of the subsystem. The threshold λ of the defuzzifier can be dynamically adjusted according to the requirements of fault diagnosis accuracy and false alarm rate. According to the formula of the centroid of area method, the calculation of the preliminary crisp value is as follows: Here, the numerator is the sum of the products of each discrete point and its membership degree, and the denominator is the sum of all membership degrees. The preliminary crisp value output based on the centroid of area method is obtained through this formula. According to the requirements of the accuracy and false alarm rate of the air compressor fault diagnosis, a real threshold λ∈[0,1] is dynamically set, and then the discretized fuzzy set is traversed to find all the corresponding values with membership function values greater than the threshold. If only one value meets the conditions, then this value is the final fault diagnosis result of the subsystem. If there are multiple values that meet the conditions, the maximum value among them is taken as the final fault diagnosis result of the subsystem.

[0027] As an improvement of the present invention, the input end of the result output module is connected to the output end of the interval type-2 fuzzy inference subsystem, and the finally determined fault type is output to the user in the form of digital coding, text description or graphical interface. The output result includes the fault type, the fault severity level (which can be divided according to the output value size or predefined threshold range), and the probability estimate of the fault occurrence (reflected by the relative size of the exact value output by the interval type-2 fuzzy inference system).

[0028] An alarm module, connected to the result output module. When a fault is diagnosed and the fault severity reaches the preset alarm threshold, it sends alarm information to relevant maintenance personnel through sound and light alarms, text message notifications or remote communication protocols. The alarm threshold can be set according to the equipment operation safety requirements and maintenance plans.

[0029] A remote fault diagnosis method for air compressors based on CNN and type-2 fuzzy algorithm, the method includes the following steps:

[0030] Data preprocessing step: Normalize the original data collected by sensors so that its numerical range is between [0, 1] or [-1, 1], remove outliers and noise data, and expand the training data set through data augmentation techniques, including performing translation, scaling, and flipping operations on time series data.

[0031] CNN training step: Use the labeled training data to train the CNN module with stochastic gradient descent, Adagrad, Adadelta optimization algorithms, set appropriate learning rates, iteration times, and batch sizes. During the training process, monitor the performance of the model through cross-validation techniques to prevent overfitting, and adjust the model structure and parameters according to the verification results.

[0032] Interval type-2 fuzzy inference system training step: Based on historical fault cases and expert knowledge, determine the initial rules and membership function parameters of the fuzzy rule base, and use training data and intelligent optimization algorithms such as genetic algorithms and particle swarm optimization algorithms to optimize and adjust the fuzzy rules and membership functions to improve the accuracy and robustness of fault diagnosis. During the optimization process, measure the overall performance of the system through the combined evaluation index with the CNN module. Taking the particle swarm optimization algorithm as an example, the specific training steps are as follows:

[0033] Collect historical fault cases: Collect a large number of historical fault cases from channels such as the operation and maintenance records and fault repair reports of air compressors. These cases should contain various fault phenomena of the air compressor under different operating states, the corresponding fault causes, and relevant sensor data and other information.

[0034] Organize expert knowledge: Experts can summarize the relationships between some common fault modes and sensor data characteristics, as well as the mutual influences between different fault causes, etc. according to their professional knowledge and practical experience.

[0035] Construct the initial rules of the fuzzy rule base: Combine historical fault cases and expert knowledge to start constructing the initial rules of the fuzzy rule base. For example, if an expert points out that when the exhaust pressure of the air compressor is too high and the temperature rises abnormally, it is very likely that there is a fault in the intake valve. Then a fuzzy rule can be constructed: "If the exhaust pressure is 'high' and the temperature is 'high', then the probability of the fault cause being 'intake valve fault' is high."

[0036] Determine the membership function parameters: For each fuzzy variable, such as "exhaust pressure", "temperature", etc., according to its value range and physical meaning, determine the corresponding membership function type, such as triangular, trapezoidal, Gaussian, etc., and initially set its parameters. Taking "exhaust pressure" as an example, according to the normal working pressure range of the air compressor and the pressure fluctuation situation in historical data, determine the membership function parameters corresponding to fuzzy states such as "high", "medium", and "low".

[0037] Prepare the training data: Select a part of the data from the historical data as the training data. These data should cover various fault types and different operating conditions to ensure that the training data is representative and diverse. The training data includes input feature data (such as sensor measurement values) and corresponding output labels (actual fault types or fault degrees).

[0038] Define the fitness function: The fitness function is used to evaluate the quality of each particle. In the optimization of the interval type-2 fuzzy inference system, use indicators such as the accuracy of fault diagnosis, false alarm rate, and missed alarm rate to construct the fitness function. For example, the ratio of the number of correctly diagnosed samples to the total number of samples can be used as the main part of the fitness function, that is, fitness = number of correctly diagnosed samples / total number of samples. At the same time, penalty terms for false alarms and missed alarms can be considered to comprehensively evaluate the performance of the model.

[0039] Encoding and initializing the population: Represent the fuzzy rules and membership function parameters as the position vectors of the particles, and randomly initialize the positions and velocities of the particle swarm. Each particle represents a candidate solution for a set of fuzzy rules and membership function parameters.

[0040] Update the velocity and position: According to the current position, velocity, self-historical optimal position, and global optimal position of the group of the particle, update the velocity and position of each particle according to the velocity and position update formulas of the particle swarm optimization algorithm.

[0041] Evaluation and update: Calculate the fitness value of each particle, update the self-historical optimal position and global optimal position of the group of the particle. Judge whether the termination condition is met, such as reaching the maximum number of iterations, the fitness value converges, etc. If the termination condition is not met, continue to update the velocity and position of the particle until the termination condition is met.

[0042] Determine the optimized fuzzy rules and membership functions: After the intelligent optimization algorithm meets the termination condition, the fuzzy rules and membership function parameters corresponding to the optimal individual (genetic algorithm) or the global optimal particle (particle swarm optimization algorithm) are the optimized results. These optimized fuzzy rules and membership functions will be used to construct the final interval type-2 fuzzy inference system to improve the accuracy and reliability of air compressor fault diagnosis.

[0043] The sensor group is carefully equipped with various types of sensors such as temperature, pressure, vibration, and Hall effect, which can comprehensively collect 14-dimensional key data during the operation of the air compressor. These data are like the "health indicators" of the equipment, providing a rich information basis for subsequent fault diagnosis.

[0044] The CNN module undertakes the core data processing task in the whole system. Its convolutional layer uses an adaptively adjustable convolutional kernel, which can keenly capture local features in the data and is good at dealing with details. The pooling layer effectively reduces the data dimension through max pooling or average pooling, ensuring that important features are not missed while reducing the computational amount. The fully connected layer uses a weight matrix to accurately map the extracted feature vectors to a 15-dimensional fault result space, and through continuous optimization of the weights by the backpropagation algorithm under the training of a large number of labeled samples, it can quickly and accurately classify the system faults initially and screen out the possible problem subsystems.

[0045] The type-2 fuzzy inference subsystem is specially designed for different subsystem fault diagnoses. Each subsystem uses components such as a fuzzifier, an inference engine, a rule base, a defuzzifier, and a defuzzification unit to conduct in-depth fault analysis and judgment according to its unique input and output dimensions and specific fault types. The fuzzifier uses a variety of advanced membership functions, combines historical data and expert experience to optimize parameters, and fuzzifies the feature vectors output by the CNN, enabling the system to better handle the uncertainty of the data. The inference engine conducts precise inference calculations based on a rigorous inference structure and carefully extracted fuzzy rules. The defuzzifier uses an efficient EKM algorithm to ensure quick and accurate conversion to a type-1 fuzzy set when dealing with complex membership functions and dynamic fuzzy sets. The defuzzification unit finally obtains accurate fault diagnosis results through the ingenious combination of the center of area method and the threshold method.

[0046] In terms of the training and optimization of the system, there is also a set of scientific and rigorous methods. In the data preprocessing stage, by normalizing the original data, its numerical range is adjusted to a suitable interval, while removing outliers and noise data, and using data augmentation techniques to expand the training dataset, improving the generalization ability of the model. During the CNN training process, an optimization algorithm is reasonably selected, parameters such as the learning rate, number of iterations, and batch size are carefully set, and the cross-validation technique is used to monitor the model performance, effectively preventing overfitting. The training of the interval type-2 fuzzy inference system determines the initial rules and parameters based on historical fault cases and expert knowledge, and then uses intelligent optimization algorithms for further optimization and adjustment. Through the joint evaluation index with the CNN module, the overall performance of the system is continuously improved.

[0047] The designs of the result output module and the alarm module also fully consider the requirements of practical applications. The result output module presents the fault diagnosis results to the user in a variety of intuitive forms, including information such as the fault type, severity level, and probability of occurrence estimation, providing comprehensive decision-making basis for maintenance personnel. The alarm module, when the fault severity reaches the preset threshold, quickly sends alarm messages to relevant personnel through sound and light alarms, SMS notifications, or remote communication protocols, ensuring that the fault can be processed in a timely manner, minimizing the equipment downtime to the greatest extent, and guaranteeing the continuity of production.

[0048] The remote fault diagnosis method for air compressors proposed in this application performs excellently in the fault diagnosis of each system, not only improving the fuzziness of the system but also enhancing its interpretability. By simplifying the construction process of the rule base, we effectively reduce the complexity of the system. In addition, the recognition accuracy of this method has also been significantly improved, providing a solid guarantee for the effectiveness of fault diagnosis. Based on this, the developed improved centrifugal air compressor fault diagnosis system is particularly suitable for large industrial centrifugal air compressors. Through innovative architecture design, combined with advanced algorithm fusion technology and scientific training optimization methods, this system effectively solves many challenges faced in traditional air compressor fault diagnosis, providing reliable operation guarantee and efficient maintenance strategies for air compressors in industrial production. This system not only has great application value but also shows broad market prospects. With the accelerating progress of industrial automation and intelligence, our fault diagnosis technology can provide strong technical support for the safety and stability of various industrial equipment, helping enterprises achieve efficient production and sustainable development, with important application value and broad market prospects. Brief Description of the Drawings

[0049] The above and other objects, features, and advantages of the present application will become more apparent by describing the embodiments of the present application in more detail with reference to the accompanying drawings. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation to the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0050] Figure 1 is a schematic diagram of the overall system for remote fault diagnosis of the air compressor of the present application,

[0051] Figure 2 is a schematic diagram of the fault diagnosis system based on CNN and type-2 fuzzy of the present application,

[0052] Figure 3 is a schematic diagram of the fault decision result tree of the fault diagnosis system of the present application,

[0053] Figure 4 is a schematic diagram of the structure of the interval type-2 fuzzy inference system of the present application,

[0054] Figure 5 is a schematic diagram of the trapezoidal interval type-2 membership function of the present application. Detailed implementation manners

[0055] In view of the above-mentioned disadvantages of the prior art, the present application provides a remote fault diagnosis system for air compressors based on CNN and type-2 fuzzy algorithms to solve the above technical problems.

[0056] Embodiment 1: The present application provides a method for remote fault diagnosis of an air compressor, as Figure 1 shown, the main contents include:

[0057] Collect parameters such as the pressure, temperature, rotational speed, and vibration frequency and amplitude of the air compressor through sensors.

[0058] Transmit the collected data to the remote monitoring system through the network to achieve real-time sharing of data. Considering the reliability, real-time performance, and security of data transmission, a wireless network based on Wi-Fi 6 is selected as the network communication method.

[0059] Build a monitoring platform in the remote monitoring system to display, store, and analyze the operation data of the air compressor in real time. When a fault occurs, the system can send out an alarm message in a timely manner.

[0060] In order to facilitate users to monitor the air compressor, a Web page is used as the UI interface of the monitoring platform, and the Web page can achieve flexible layout design and interactivity, which is suitable for real-time display and storage of the operation data of the air compressor.

[0061] For the convenience of real-time data storage and retrieval, MySQL relational data is used for data storage.

[0062] Given the high accuracy of CNN, its fast response in processing a large amount of data in a short time, and the ability of type-2 fuzzy algorithm to handle fuzzy, uncertain and incomplete information, CNN and type-2 fuzzy algorithm are combined in data analysis to achieve efficient and accurate fault diagnosis.

[0063] Contents collected by sensors: Pressure sensors are used to detect parameters such as intake pressure and exhaust pressure, and judge whether there is abnormal pressure in the air compressor by obtaining pressure data in real time.

[0064] Temperature sensors are used to measure the temperatures of key parts of the air compressor, such as cylinder temperature, lubricating oil temperature, exhaust temperature, etc. An abnormal increase in temperature is often a precursor to a fault.

[0065] Vibration sensors are used to detect information such as vibration amplitude, frequency and phase. When the rotor of the air compressor is unbalanced, misaligned or mechanical components are loose, the vibration characteristics will change significantly.

[0066] Hall effect sensors are used for speed detection. By detecting the magnetic field change on the rotating part to measure the speed and calculating the frequency or pulse number of the signal, the speed of the rotating part can be determined.

[0067] The content of fault diagnosis is as Figure 2 shown: The data of each item of the air compressor collected are input into the CNN model to obtain the system-level fault diagnosis results, which are divided into normal, mechanical system fault, oil circuit system fault, and gas circuit system fault. After determining the subsystem fault type, the feature vectors of each sample data extracted from the CNN are input into the interval type-2 fuzzy inference system of the corresponding subsystem. After passing through the five modules of fuzzifier, inference engine, rule base, defuzzifier, and de-fuzzifier, a vector composed of accurate values can be obtained. The fault type number corresponding to the largest accurate value in the vector is the finally determined fault.

[0068] Figure 3 For the fault decision result, the data collected by the sensors or images are classified into three types of system faults by CNN: mechanical system, oil circuit system, and gas circuit system. The original sample data has 14 dimensions (X1-X14), and the output fault result has 15 dimensions (Y1-Y15). After passing through the CNN system, it is divided into four categories: normal, mechanical system fault, oil circuit system fault, and gas circuit system fault.

[0069] As shown in Table 1 and Table 2, the correct category does not need to be discriminated anymore; the mechanical interval type-2 fuzzy system has only 3-dimensional inputs (X2, X3, X4), namely (primary shaft vibration, secondary shaft vibration, tertiary shaft vibration), and 4-dimensional outputs (Y1, Y2, Y3, Y4), namely (unbalance of dynamic balance of primary rotor or impeller, unbalance of dynamic balance of secondary rotor or impeller, unbalance of dynamic balance of tertiary rotor or impeller, surge); the oil circuit interval type-2 fuzzy system has only 5-dimensional inputs (X5, X6, X7, X12, X13), namely (lubricating oil temperature, lubricating oil pressure, oil filter pressure difference, ambient temperature, inlet pre-cooler temperature), and 5-dimensional outputs (Y5, Y6, Y7, Y8, Y9), namely (insufficient oil quantity, oil filter blockage, too low ambient temperature, oil cooler failure, outlet oil pipe blockage); the gas circuit interval type-2 fuzzy system has only 6-dimensional inputs (X1, X9, X10, X12, X13, X14), namely (air filtration resistance, motor current, exhaust pressure, ambient temperature, inlet pre-cooler temperature, outlet pre-cooler temperature), and 6-dimensional outputs (Y10, Y11, Y12, Y13, Y14, Y15), namely (intermediate cooler failure, pre-cooler failure, too high production ambient temperature, air filter blockage, air separation pipe blockage, power grid fluctuation).

[0070] The specific process of the CNN system is as follows: First, compress the data collected by the sensor into N x 14, where N is the number of samples and 14 is the dimension, and label the compressed data with the four categories of normal, mechanical system, oil circuit system, and gas circuit system to obtain the original data set.

[0071] In the preprocessing part, divide the data set into a training set and a test set according to the ratio of 2:8, and use the cross-validation method for model evaluation.

[0072] The CNN network trains the preprocessed input data, and the activation function is selected as ELU. It is recommended to select a convolution kernel of size 3x3, and the pooling layer window is 2x2. Other parameters such as batch size, number of training times, and learning rate can be set by the user. During the training process, use the Adam optimizer for training optimization, and use the cross-entropy loss function for calculating the loss function part.

[0073] Predict the results output by the CNN, and use indicators such as classification accuracy and confusion matrix to evaluate the classification effect of the model. Finally, obtain one result in the four-classification. If it is one of the three results other than normal, continue to input the feature vector obtained after pooling into the type-2 fuzzy system for further fault diagnosis.

[0074] Table 1 Total system inputs

[0075]

[0076] The feature vectors obtained after the data collected by the sensor pass through the convolutional layer and the pooling layer can be used as the input of the interval type-2 fuzzy system. Figure 4 As shown in the structural schematic diagram of the interval type-2 fuzzy inference system, it can be divided into five modules: fuzzifier, inference engine, rule base, defuzzifier, and de-fuzzifier.

[0077] The interval type-2 fuzzy system fuzzifies these inputs according to predefined fuzzy rules and membership functions. The upper and lower boundaries of the uncertainty domain are two type-1 membership functions, denoted as the upper membership function UpperMF and the lower membership function LowerMF respectively. Taking the trapezoidal function as an example, as Figure 5 shown. Finally, through the fuzzy inference engine for inference, according to the membership degree of the input features and the fuzzy rules, the output fuzzy set is calculated.

[0078] Table 2 Fault types corresponding to the system output serial numbers

[0079]

[0080] In order to enhance the fuzziness of the system, the following four functions are mainly selected for the membership functions in this paper: z-type function (zmf), s-type function (smf), Gaussian function (gaussmf), and double Gaussian mixture function (gauss2mf). Taking the design of the fuzzy set and membership function of the gas path interval type-2 fuzzy system as an example, as shown in Table 3.

[0081] Table 3 Design of the membership function of the gas path type-2 system

[0082]

[0083] Rule base: Analyze and extract the corresponding system fault rules from the summarized fault causes and the system fault tree, simplify and merge the fuzzy rules, and convert them into fuzzy languages that the fuzzy system can recognize. When the fuzzy system is of Mamdani structure and the input variables are x = (x1, x2,..., x n ) T ∈R n ), the i-th fuzzy rule is expressed as:

[0084] Rule i : If x1 is Α1 i and x2 is A2 i and...and x n is A n i

[0085] Then y = B i

[0086] where A i 、Bi They represent the type-2 fuzzy sets in the antecedent and consequent of the rule respectively.

[0087] For example, the rule R is extracted from the fault tree of the gas path system: If the pre-cooler fails (Y11), then the gas cannot reach the predetermined temperature reduction effect after entering the pre-cooler, the temperature of the gas leaving the pre-cooler is too high, and finally the exhaust gas temperature is too high. After converting it into a fuzzy rule, it is:

[0088] R:If X1 is L,X9 is Ν,X10 is Ν,X12 is L,X13 is L,X14 is Η

[0089] then y=Y11

[0090] Defuzzifier: Compared with the traditional type-1 fuzzy system, the interval type-2 fuzzy system adds a defuzzifier module. Since the output result of the inference engine is an interval type-2 fuzzy set, it is necessary to convert it into a type-1 fuzzy set through the defuzzifier and then pass it to the defuzzifier to convert it into an exact value as the final output of the system. Three centroid defuzzification algorithms, such as KM (Karnik-mendel), MKM (Modified Karnik-mendel), and EKM (enhanced Karnik-mendel), can be used for the defuzzifier. Here we use the EKM algorithm.

[0091] The KM algorithm is a classic method for calculating the centroid in fuzzy systems. It played an important role in early fuzzy control and fuzzy inference systems and provided a basis for the development of subsequent algorithms. The KM algorithm can effectively calculate the centroid of a fuzzy set to obtain a suitable control quantity. However, its computational efficiency is low and the accuracy is limited. When the number of elements in the fuzzy set is large or the shape of the membership function is complex, a large number of sorting and interpolation calculations are required.

[0092] The MKM algorithm improves the KM algorithm and reduces the computational amount to a certain extent. Through some optimization strategies, it avoids some repeated calculation steps in the KM algorithm, can reduce the number of sorting and interpolation, and thus improves the calculation speed. Compared with the KM algorithm, the MKM algorithm also has improved accuracy, can consider more detailed information, is more sensitive to changes in the membership function, can better adapt to changes in the shape of the membership function, and obtain a more accurate centroid.

[0093] Compared with the KM and MKM algorithms, the EKM algorithm has higher computational efficiency, fewer iteration times and may not require sorting; it is superior in terms of accuracy, can handle complex membership functions and adapt to dynamic changes; it has good adaptability and robustness, is suitable for various fuzzy systems and is easy to combine with other algorithms, and has significant advantages in multi-field applications. Therefore, we adopt the EKM algorithm for defuzzification.

[0094] After the defuzzifier is calculated by the system inference engine and the type-reduction unit, the fuzzy output y(μ R (y1), μ R (y2),..., μ R (y n )) is obtained. However, it is a type-1 fuzzy set and needs to be converted into a crisp value before it can be output and passed to the user. Therefore, a defuzzifier is required to defuzzify the fuzzy vector.

[0095] Here, we use a defuzzification method that combines the centroid of area method and the threshold method to obtain a deterministic output, that is, the fault diagnosis result.

[0096] Centroid of area method: Take the weighted average y′ of the fuzzy vector y as the crisp value output.

[0097]

[0098] Threshold method: In a fuzzy set y of the universe of discourse X, setting an arbitrary real number threshold λ ∈ [0, 1], we can get:

[0099] y λ ={x|μ R (x)≥λ, x ∈ X}

[0100] Take out all the membership function values in the fuzzy vector y. If there are values greater than λ, then judge the corresponding yi as the crisp value and output it.

[0101] Taking the pneumatic circuit system as an example, for each output Yi, first perform type-reduction on it through the centroid of area method, and the obtained type-1 fuzzy set can be converted into an accurate value Yi. Then form the output array Y = (Y10’, Y11’, Y12’, Y13’, Y14’, Y15’), where Y is the output value of the type-2 fuzzy system, and Yi′ represents the possibility of the occurrence of the i-th fault.

[0102] Then identify the fault through the threshold method. Set the judgment threshold λ, compare and take out the output value Yi′ greater than the threshold as the output (if there are multiple values greater than the threshold λ at the same time, take the largest value among them as the output result). Finally, the fault type Yi corresponding to Yi′ is the final fault diagnosis result of the pneumatic circuit interval type-2 fuzzy system.

[0103] It should be noted that the above embodiments are not used to limit the protection scope of the present invention. Equivalent transformations or substitutions made on the basis of the above technical solutions all fall within the protection scope of the claims of the present invention.

Claims

1. The air compressor remote fault diagnosis system based on CNN and type-II fuzzy algorithm is characterized by: The diagnostic system includes a sensor group, a CNN module, an interval type-2 fuzzy reasoning subsystem, a result output module and an alarm module. The sensor group collects the operating data of the air compressor and transmits it to the CNN module to realize the preliminary processing of the information. The CNN module extracts features and classifies system-level faults on the input data, identifies potential system-level faults, and outputs the diagnosis results. Then, the interval type-2 fuzzy reasoning subsystem combines the output of the CNN module with the original data to perform more detailed device-level fault diagnosis to determine the specific fault type. The result output module is responsible for presenting the diagnosis results to the user in an intuitive visual form to facilitate understanding and decision-making. If the system detects a fault, the alarm module will compare the fault diagnosis results with the preset threshold and trigger the alarm in time to remind the user to take necessary maintenance measures.

2. The air compressor remote fault diagnosis system based on CNN and type-II fuzzy algorithm according to claim 1 is characterized in that: The sensor group consists of a temperature sensor, a pressure sensor and a vibration sensor, and is used to collect the operating data of the centrifugal air compressor. The data includes air filtration resistance (X1), primary shaft vibration (X2), secondary shaft vibration (X3), tertiary shaft vibration (X4), lubricating oil temperature (X5), lubricating oil pressure (X6), oil filter pressure difference (X7), bearing temperature drive end (X8), motor current (X9), exhaust pressure (X10), exhaust flow (X11), ambient temperature (X12), pre-cooler inlet temperature (X13), and pre-cooler outlet temperature (X14).

3. The air compressor remote fault diagnosis system based on CNN and type-II fuzzy algorithm according to claim 2 is characterized in that: The CNN module, whose input end is connected to the sensor group, performs normalization processing on the collected 14-dimensional original sample data to avoid the influence of different dimensions on the fault feature extraction, extracts the feature vector before the full connection layer after the convolution layer and the pooling layer operation, and divides the system fault into four types of system-level faults: normal, mechanical system fault, oil system fault, and gas system fault. Then, the interval type-2 fuzzy reasoning subsystem outputs 15-dimensional fault results (Y1-Y15), which are respectively the first-stage rotor or impeller dynamic balance imbalance, the second-stage rotor or impeller dynamic balance imbalance, the third-stage rotor or impeller dynamic balance imbalance, surge, insufficient oil, oil filter blockage, low ambient temperature, oil cooler failure, oil outlet pipe blockage, intercooler failure, precooler failure, high production environment temperature, air filter blockage, air distribution pipe blockage, and power grid fluctuation.

4. The air compressor remote fault diagnosis system based on CNN and type-II fuzzy algorithm according to claim 1 is characterized in that: The diagnostic system is equipped with multiple interval type-2 fuzzy reasoning subsystems. Each subsystem independently receives the corresponding feature vector output by the CNN module, and performs comprehensive processing through its own fuzzifier, inference engine, rule base, reducer and defuzzifier modules, and finally outputs a vector composed of precise values ​​to clearly determine the fault type, as follows: Mechanical interval type II fuzzy reasoning system: The system failure is mainly affected by the primary shaft vibration (X2), the secondary shaft vibration (X3), and the tertiary shaft vibration (X4). The specific failure causes can be divided into four types, namely, primary rotor or impeller dynamic balance imbalance, secondary rotor or impeller dynamic balance imbalance, tertiary rotor or impeller dynamic balance imbalance, and surge. The mechanical interval type II fuzzy reasoning system only needs to pay attention to the 3D input, namely, primary shaft vibration (X2), secondary shaft vibration (X3), tertiary shaft vibration (X4), and output 4D results (Y1, Y2, Y3, Y4), namely (primary rotor or impeller dynamic balance imbalance, secondary rotor or impeller dynamic balance imbalance, tertiary rotor or impeller dynamic balance imbalance, and surge), which is used to identify the fault characteristics of the mechanical part. Oil circuit interval type II fuzzy reasoning system: The system failure is mainly affected by lubricating oil temperature (X5), lubricating oil pressure (X6), oil filter pressure difference (X7), ambient temperature (X12), and pre-cooler inlet temperature (X13). The specific failure causes can be divided into five types, namely insufficient oil, oil filter blockage, low ambient temperature, oil cooler failure, and oil outlet pipe blockage. The oil circuit interval type II fuzzy reasoning system only needs to process 5-dimensional input, namely lubricating oil temperature (X5), lubricating oil pressure (X6), oil filter pressure difference (X7), ambient temperature (X12), and pre-cooler inlet temperature (X13), and output 5-dimensional results (Y5, Y6, Y7, Y8, Y9), namely (insufficient oil, oil filter blockage, low ambient temperature, oil cooler failure, and oil outlet pipe blockage), which are used to analyze related fault conditions of the oil circuit. Gas path interval type II fuzzy reasoning system: The failure of this system is mainly affected by the air filter resistance (X1), motor current (X9), exhaust pressure (X10), ambient temperature (X12), precooler inlet temperature (X13), and precooler outlet temperature (X14). The specific failure causes of this system can be divided into six types: intercooler failure (Y10), precooler failure (Y11), excessive ambient temperature (Y12), air filter blockage (Y13), air separation pipeline blockage (Y14), and power grid fluctuation (Y15). Therefore, the gas path interval type II fuzzy reasoning system only needs to pay attention to the inputs X1, X9, X10, X12, X13, and X14 detected by the sensor, and the output parts Y10, Y11, Y12, Y13, Y14, and Y15 for accurate judgment of gas path failures.

5. The air compressor remote fault diagnosis system based on CNN and type-II fuzzy algorithm according to claim 1 is characterized in that ,CNN modules include: At least one convolution layer, with convolution kernel parameters adaptively adjusted, for extracting features from input data and capturing local feature patterns in the data through convolution operations; At least one pooling layer, using maximum pooling or average pooling to reduce the dimension of the convolutional layer output, reduce the amount of calculation and retain the main features; The fully connected layer, whose input is connected to the output of the pooling layer or the feature map after flattening, maps the extracted feature vector to the 15-dimensional fault result space through the weight matrix to realize fault classification. The weight of the fully connected layer is optimized and updated through the back propagation algorithm during the training process. The training data contains a large number of labeled centrifugal air compressor fault samples and normal operation samples.

6. The air compressor remote fault diagnosis system based on CNN and type II fuzzy algorithm according to claim 1 is characterized in that ,The components of interval type-2 fuzzy reasoning subsystem include: Fuzzy machine: trapezoidal function, z-type function (zmf), s-type function (smf), Gaussian function (gaussmf), and double Gaussian mixture function (gauss2mf) are used as membership functions. The input feature vector is fuzzified according to predefined parameters to determine the degree to which each input feature belongs to different fuzzy sets. The membership function parameters of the fuzzifier are adjusted and optimized based on historical fault data and expert experience. Inference engine: Based on the Mamdani inference structure, the inference calculation is performed according to the membership of the input features and the fuzzy rules extracted and simplified from the fault causes and system fault trees to generate output fuzzy sets. First, it is necessary to systematically study the common failure modes of centrifugal air compressors and analyze the causal relationship between each failure mode and the sensor data features. Based on the above causal relationship, the corresponding fuzzy rules are extracted. Reducer: The EKM (enhanced Karnik-mendel) algorithm is used to convert the interval type-2 fuzzy set output by the inference engine into a type-1 fuzzy set through an efficient iterative calculation process. First, the relevant parameters in the EKM algorithm need to be determined, and then the upper and lower boundary centroids are calculated. First, the domain of the output interval type-2 fuzzy set is discretized. Assuming the domain is [a, b], it can be divided into n points at equal intervals, namely x1, x2, …, x n , then calculate the upper boundary centroid and the lower boundary centroid respectively according to the centroid formula, take the upper and lower boundary centroids obtained by the first calculation as the initial value, and then iterate. In each iteration, recalculate the upper and lower boundary centroids according to the current upper and lower boundary centroids. If the number of iterations reaches the maximum number of iterations, stop the iteration; otherwise, return to the previous step to continue the iteration. It has high computational efficiency and accuracy when processing complex membership functions and dynamically changing fuzzy sets, and the calculation process of the reducer can be parallelized according to the real-time performance requirements of the system; Defuzzifier: The main function is to convert the output fuzzy set into specific fault diagnosis results, such as determining the probability or severity of the current fault being surge, so as to provide clear operational guidance for fault maintenance. In order to achieve efficient and accurate defuzzification, the area centroid method and the threshold method are combined. The specific steps are as follows: First, the weighted average of the fuzzy vector is taken as the initial clarity value output through the area centroid method, and then according to the set real number threshold λ∈[0,1], the corresponding value with a membership function value greater than λ is screened out as the final clarity value output. If there are multiple values, the largest value is taken as the final fault diagnosis result of the subsystem. The threshold λ of the defuzzifier can be dynamically adjusted according to the accuracy and false alarm rate requirements of the fault diagnosis. According to the formula of the area centroid method, the calculation of the initial clarity value y1 is as follows: The numerator is the sum of the products of each discrete point and its membership, and the denominator is the sum of all memberships. The y1 obtained by this formula is the preliminary clear value output based on the area centroid method. According to the accuracy and false alarm rate requirements of air compressor fault diagnosis, a real number threshold λ∈[0,1] is dynamically set, and then the discretized fuzzy set is traversed to find all the corresponding x values ​​whose membership function values ​​are greater than the threshold. i Value, if the filtered x meets the conditions i If there is only one, then this value is the final fault diagnosis result of the subsystem. If there are multiple x that meet the conditions i , then take the largest value as the final fault diagnosis result of the subsystem.

7. The air compressor remote fault diagnosis system based on CNN and type II fuzzy algorithm according to claim 1 is characterized in that ,The input end of the result output module is connected to the output end of the interval type-2 fuzzy ,reasoning subsystem, and the final determined fault type is output to the ,user in the form of digital code, text description or graphical interface. ,The output result includes the fault type, fault severity level and the ,probability estimation of the fault occurrence; The alarm module is connected to the result output module. When a fault is diagnosed and the fault severity reaches the preset alarm threshold, an alarm message is sent to the relevant maintenance personnel through sound and light alarm, SMS notification or remote communication protocol. The alarm threshold can be set according to the equipment operation safety requirements and maintenance plan.

8. The remote fault diagnosis method of air compressor based on CNN and type-II fuzzy algorithm is characterized by: The air compressor remote fault diagnosis system based on CNN and type-II fuzzy algorithm according to any one of claims 1 to 7 is adopted, and the method comprises the following steps: Data preprocessing step: normalize the raw data collected by the sensor to make its value range between [0,1] or [-1,1], remove outliers and noise data, and use data enhancement technology to translate, scale and flip the time series data to expand the training data set; CNN training steps: Use the labeled training data to train the CNN module using stochastic gradient descent and Adadelta optimization algorithms, set appropriate learning rate, number of iterations, and batch size, monitor the performance of the model during training through cross-validation technology to prevent overfitting, and adjust the model structure and parameters based on the validation results; Training steps of interval type-II fuzzy inference system: determine the initial rules and membership function parameters of the fuzzy rule base based on historical fault cases and expert knowledge, optimize and adjust the fuzzy rules and membership functions using training data and particle swarm optimization algorithm intelligent optimization algorithm, and measure the overall system performance through joint evaluation indicators with CNN module during the optimization process.

Images