Motor protector and control method thereof

Abstract

According to the motor protector and the control method thereof, the motor fault knowledge base is constructed through the knowledge graph, and data reference is provided for prediction and optimization of the neural network model. And constructing multiple classes of neural networks to evaluate and predict different types of faults of the motor, multiplying probability values of the same fault class output by the classification neural networks by corresponding weight coefficients, adding the probability values to obtain fused fault class probability values, and selecting the fault class with the maximum probability value as a prediction result. And the advantages of various neural networks in processing different types of data are fully utilized. And comprehensiveness and accuracy of fault diagnosis and protection of the motor are ensured by integrating the accuracy of the knowledge graph database and the capability of the neural network for processing high-dimensional complex data. And a self-adaptive control strategy instruction is generated based on the fusion analysis diagnosis result, the operation of the motor is controlled, the effect after the control instruction is executed is monitored in real time, and the flexibility and effectiveness of the execution strategy are ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of motor protection, in particular to a motor protector and a control method thereof. Background Art

[0002] As an energy conversion device that converts electrical energy into mechanical energy, electric motors are widely used in today's society, with an astonishing number of applications. Due to the advantages of simple overall structure, easy-to-understand manufacturing process, stable and reliable operation, and low price, electric motors are used in almost all fields.

[0003] Although the application of electric motors is quite common and has penetrated into all aspects of people's lives. However, in the actual application of electric motors, they are often affected by various factors, such as unstable city voltage caused by large power consumption, overload shock caused by excessive load, broken rotor bars caused by accidents, and harsh external environment (high temperature, dust, etc.), which lead to short circuit, overload, stall, three-phase imbalance and residual current in the motor, which bring great safety hazards to the normal operation of the motor. Once the motor fails, it will burn the motor at the least, or cause a fire in the workshop at the worst. Although the design and processing technology of electric motors have been greatly improved compared with the past, and the size of the motor has become smaller and smaller, and the appearance is more beautiful and environmentally friendly, it has led to more and more dense internal wiring, larger current density, and a significant decrease in the heat dissipation and heat resistance of the motor; and in order to meet the requirements of various tests and environments, the motor needs to be started and braked frequently, which greatly increases the possibility of accidents in the motor, which may cause the motor to stop suddenly, resulting in product scrapping, or causing personal injury. Therefore, whether from the perspective of safety, economy or environmental protection, the development of intelligent motor protectors with comprehensive protection functions has important practical significance.

[0004] The patent with authorization announcement number CN101714754B discloses a motor protector and a control method thereof, providing a simple and reliable method to realize the restart function of the motor protector. The patent with authorization announcement number CN102916399B discloses a motor protector and a method for realizing the motor protector applied to a dual-speed motor. The product has two sets of parameters, high speed and low speed, and uses one device to complete the protection and control of two devices, saving customer costs, saving installation space, and having strong targeted protection. With comprehensive measurement functions, it can measure current, voltage, residual current, frequency, power, and electric energy, and can realize protection functions such as overload, phase failure, imbalance, stall, blockage, start timeout, residual current, overvoltage, undervoltage, underpower, and external fault. However, these two schemes mainly focus on the detection and protection of motor faults, and cannot detect the hidden faults of the motor in advance, cannot analyze and determine the cause of the fault and locate the fault position after the motor fails, and cannot provide the optimized control scheme of the motor. Obviously, they cannot well meet the increasingly intelligent motor detection and control needs. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a motor protector and a control method thereof, so as to solve the problem that the existing motor protector cannot detect the hidden faults of the motor in advance, cannot analyze and determine the cause of the fault and locate the fault position after the motor fails, and cannot provide an optimized control solution for the motor.

[0006] To solve the above technical problems, an embodiment of the present invention provides the following technical solution: a control method of a motor protector, comprising the following steps:

[0007] S1. Knowledge graph construction, collecting knowledge data covering the basic principles, electrical characteristics, mechanical structure and characteristics, environmental influencing factors, and causes, development processes, common operation and maintenance and processing strategies of various types of faults of electric motors, removing duplicate and redundant parts of the knowledge data and classifying them. The knowledge data after classification includes multi-level node attributes, and the node attributes of each level include node labels, node types and node descriptions;

[0008] S2. Use the resource description framework based on the semantic web to formalize the organized knowledge data, define various entities, relationships and attributes, and construct a conceptual model that reflects the knowledge system in the field of electric motors;

[0009] S3. Create a complete electric motor knowledge graph with the help of Neo4j, a knowledge graph construction tool, and use the graph database to store and manage the knowledge graph;

[0010] S4, collecting fault type data and operation factor data of the motor during normal operation and fault operation as a data set, preprocessing the data of the data set, and performing multi-level classification on the preprocessed data set, and generating a label for each fault type; the protection measures corresponding to the fault type include but are not limited to start timeout protection, underload and overload protection, stall protection, short circuit protection, residual current protection, three-phase current imbalance protection, overvoltage and undervoltage protection, and temperature protection;

[0011] S5. Based on the multi-classified data set, a state assessment and prediction model based on multi-classification neural network fusion is constructed and trained;

[0012] S6. Collect the operating factor data corresponding to each fault category to be predicted in real time, and each classification neural network outputs its prediction probability for each fault category respectively. Multiply the probability value of the same fault category output by each classification neural network by the corresponding weight coefficient and add them together to obtain the fused fault category probability value, and select the fault category with the largest probability value as the prediction result;

[0013] S7, integrating the prediction result with the knowledge in the knowledge graph, using the knowledge stored in the knowledge graph to verify the output of the neural network, and at the same time feeding back the knowledge in the knowledge graph as prior information to the neural network for neural network optimization, and continuously looping the verification and optimization steps until the fusion analysis diagnosis result is obtained through the knowledge graph verification and the prediction is consistent;

[0014] S8, based on the fusion analysis and diagnosis results, combined with the real-time operating parameters of the motor and the operation and maintenance strategies and control rules stored in the knowledge graph, generate adaptive control strategy instructions;

[0015] S9. The generated adaptive control strategy instructions are applied to the motor and its related accessories through the corresponding actuators to realize the actual control of the motor operation. At the same time, the effect of the control instructions after execution is monitored in real time. After the motor operation factor data is collected again by the sensor and fed back to step S6, the process of S6-S9 is repeated.

[0016] Furthermore, based on the data after multi-level classification, a state assessment and prediction model based on multi-classification neural network fusion is constructed and trained, specifically:

[0017] Divide the preprocessed data set into n training sets and n test sets, where the training set and the test set are in one-to-one correspondence, and n is an integer greater than 2;

[0018] For each training set, the corresponding neural network model is used for training to obtain a trained neural network model;

[0019] The fault factor data in each test set are respectively input into the corresponding neural network model, and m output result sets are obtained by classification through the Softmax function, where m is a positive integer;

[0020] According to the historical prediction performance evaluation indicators of each classifier on the training set and the validation set, a corresponding weight coefficient is assigned to each classifier, and the classifiers are weightedly fused in combination with the weight coefficient to obtain a state evaluation and prediction model of multi-classification neural network fusion; wherein the historical prediction performance evaluation indicators include accuracy and F1 value.

[0021] Furthermore, the neural network model includes a support vector machine classifier, a random forest classifier, and an LSTM neural network classifier:

[0022] The test set includes a first test set, a second test set and a third test set, the first test set corresponds to the support vector machine classifier, the second test set corresponds to the random forest classifier, and the third test set corresponds to the LSTM neural network classifier;

[0023] The fault factor data in each test set are respectively input into the corresponding neural network model to obtain m output result sets, where m is a positive integer, including:

[0024] Using the first test set as input of the support vector machine classifier, and extracting a first feature in a discriminator of the support vector machine classifier;

[0025] Using the second test set as input of the random forest classifier, and extracting a second feature in a hidden layer of the random forest classifier;

[0026] Using the third test set as input of the LSTM neural network classifier, and extracting a third feature in the convolutional layer of the LSTM neural network classifier;

[0027] The first feature, the second feature and the third feature are respectively input into a pre-trained multi-task learning neural network model to obtain m output result sets, where m is a positive integer, wherein the multi-task learning neural network model is used to implement learning tasks, and different learning tasks correspond to different output layers, and each output layer has a different classification target, and the multi-task learning neural network model includes a hidden layer shared with the support vector machine classifier, the random forest classifier, and the LSTM neural network classifier.

[0028] Furthermore, the training method of the multi-task learning neural network model includes:

[0029] Determine a network architecture of the multi-task learning neural network model, wherein the network architecture includes a hidden layer and m output layers, each of the output layers being used to output a corresponding learning task;

[0030] Creating a loss function, the loss function is used to integrate the loss of each perspective of the multi-task learning neural network model;

[0031] Constructing the multi-task learning neural network model according to the network architecture, and training the multi-task learning neural network model using a target training set;

[0032] When training the multi-task learning neural network model, optimizing the loss function in real time so that the learning performance of the multi-task learning neural network model in all the perspectives is optimal;

[0033] Wherein, the loss function includes:

[0034]

[0035] Where L is the sum of the losses of each perspective of the multi-task learning neural network model, L i is the loss of the i-th learning task, ω i is the weight corresponding to the loss of the i-th learning task.

[0036] Furthermore, the knowledge stored in the knowledge graph is used to verify the output of the neural network, specifically:

[0037] The knowledge graph stores the fault phenomenon data and fault factor data knowledge set of the motor during normal operation and fault operation, represented by nodes and edges, and is defined as follows: x , R, E y ) triples, where E x is the head entity, R is the relationship, E y is the tail entity;

[0038] The neural network defines a category label y∈{1, 2, ..., C}, and a fault factor data sample x corresponding to a category label is given. i , the predicted probability of a fault belonging to category c is P(y=c|x)

[0039] If the fault factor data sample x i and the head entity E in the knowledge graph x Correspondingly, the Pearson correlation coefficient is used to determine the relationship between category c and tail entity E. y The correlation coefficient r is used, and the predicted probability is the product of P(y=c|x) and the correlation coefficient as the final probability.

[0040] Furthermore, the knowledge in the knowledge graph is fed back to the neural network as prior information to optimize the neural network, specifically:

[0041] Convert entities and relationships in the knowledge graph into low-dimensional vector representations through a specific embedding algorithm;

[0042] Integrate the generated entity vectors and relationship vectors into the input layer or hidden layer of the neural network, so that the neural network can use the knowledge carried by these vectors during training and reasoning;

[0043] Logical rules are extracted from the knowledge graph and converted into constraints in the neural network training process.

[0044] Furthermore, the classification of the knowledge data of the knowledge graph and the multi-level classification of the preprocessed data set both use a decision tree algorithm:

[0045] The CART algorithm is selected for motor fault type classification. The Gini index is used as the feature selection standard when building the decision tree. The calculation formula of the Gini index is:

[0046] Where D represents the data set, K is the number of fault type categories, and p k is the proportion of samples belonging to the kth fault type in the data set to the total samples;

[0047] Calculate the Gini index of all available features in the data set, select the feature with the smallest Gini index as the basis for dividing the root node, and for feature A, divide the data set D into D1, D2, ..., Dn subsets according to its different values. Then the Gini index of feature A is

[0048] For each subset divided from the root node, repeat the above process of calculating the Gini index and selecting the minimum Gini index feature to construct child nodes, continuously branching and continuously dividing the data set;

[0049] Set the stopping condition. If all samples in a node belong to the same fault type, that is, the classification purity reaches the highest, then the node will be used as a leaf node and will not be divided further.

[0050] Furthermore, based on the fusion analysis and diagnosis results, combined with the real-time operating parameters of the motor and the operation and maintenance strategies and control rules stored in the knowledge graph, an adaptive control strategy instruction is generated, specifically:

[0051] According to the current running state of the motor and the fault diagnosis result, a comprehensive judgment is made on the current running state of the motor, wherein the running state includes the normal running state, the fault warning state, and the fault response state;

[0052] Search the knowledge graph for relevant knowledge that matches the current fault type and motor operating status, and perform operation and maintenance strategy matching and control rule matching. The operation and maintenance strategy matching includes generating operation and maintenance suggestions for specific faults and operation and maintenance adjustments based on the operating status. The control rule matching includes adjusting the control threshold under the fault state and generating control logic for different combinations of faults and operating states.

[0053] Generate adaptive control strategy instructions for different fault types, where the different fault types include electrical faults, mechanical faults, and other faults.

[0054] Furthermore, the adaptive control strategy instruction is specifically:

[0055] Normal operating state strategy: When the evaluation model determines that the motor is in a normal operating state, the control method monitors the changes of various operating parameters in real time, and dynamically adjusts the threshold range of each monitoring parameter through the fuzzy logic optimization algorithm based on historical operating data and current working conditions;

[0056] Fault warning status strategy: If the evaluation model determines that the motor is in a minor fault state, a fault warning signal is issued and the motor's operating parameters are automatically adjusted;

[0057] Fault response status strategy: When the motor is judged to be in a moderate or severe fault state, the corresponding protection action is triggered immediately, and the key information of various operating parameters and fault types when the fault occurs is recorded.

[0058] The present invention also proposes a motor protector corresponding to the above control method, comprising: a controller, an interface circuit module, a signal acquisition module, a signal processing module, a human-machine interface module, an execution module, and a power module, wherein the controller is electrically connected to the interface circuit module, the interface circuit module, the human-machine interface module, the execution module, and the power module, the input end of the signal processing module is electrically connected to the signal acquisition module, the power module is also electrically connected to the signal processing module and the execution module, and the interface circuit module is communicatively connected to a computer; wherein the software modules running on the computer include:

[0059] The knowledge graph construction module is used to collect knowledge data covering electric motors, remove duplicate and redundant parts of the knowledge data and classify them. The knowledge data after classification includes multi-level node attributes, and the node attributes of each level include node labels, node types and node descriptions; the sorted knowledge data is formally represented by using a resource description framework based on the semantic web, and various entities, relationships and attributes are defined to construct a conceptual model reflecting the knowledge system in the field of electric motors; a complete knowledge graph of electric motors is created with the help of the knowledge graph construction tool Neo4j, and the knowledge graph is stored and managed using a graph database;

[0060] A data set construction module is used to collect fault type data and operating factor data during normal operation and fault operation of the motor as a data set, preprocess the data of the data set, perform multi-level classification on the preprocessed data set, and generate a label for each fault type; the protection measures corresponding to the fault type include but are not limited to start timeout protection, underload and overload protection, stall protection, short circuit protection, residual current protection, three-phase current imbalance protection, overvoltage and undervoltage protection, and temperature protection;

[0061] The state assessment and prediction model building module is used to build and train a state assessment and prediction model based on multi-classification neural network fusion based on the multi-classification data set;

[0062] The fault probability prediction module is used to collect the operating factor data corresponding to each fault category to be predicted in real time. Each classification neural network outputs its predicted probability for each fault category respectively. The probability values ​​of the same fault category output by each classification neural network are multiplied by the corresponding weight coefficient and then added to obtain the fused fault category probability value. The fault category with the largest probability value is selected as the prediction result.

[0063] An analysis and diagnosis module is used to fuse the prediction results with the knowledge in the knowledge graph, use the knowledge stored in the knowledge graph to verify the output of the neural network, and feed back the knowledge in the knowledge graph as prior information to the neural network for neural network optimization, and continuously cycle the verification and optimization steps until the fusion analysis and diagnosis results are obtained through the knowledge graph verification and the prediction are consistent;

[0064] The control instruction generation module is used to generate adaptive control strategy instructions based on the fusion analysis and diagnosis results, combined with the real-time operating parameters of the motor and the operation and maintenance strategies and control rules stored in the knowledge graph;

[0065] The instruction control module is used to apply the generated adaptive control strategy instructions to the motor and its related accessories through the corresponding actuators to achieve actual control of the motor operation; at the same time, it monitors the effect of the control instructions in real time, collects the motor's operating factor data again through sensors, and feeds it back to the fault probability prediction module, and then repeats the subsequent process.

[0066] The beneficial effects of the above technical solution of the present invention are as follows:

[0067] 1. The present invention constructs a motor fault knowledge base through a knowledge graph to provide data reference for the prediction and optimization of the neural network model. Multiple types of neural networks are constructed to evaluate and predict different types of motor faults, and the probability values ​​of the same fault category output by each classification neural network are multiplied by the corresponding weight coefficient and then added to obtain the fused fault category probability value. The fault category with the largest probability value is selected as the prediction result, making full use of the advantages of various neural networks in processing different types of data.

[0068] 2. The present invention integrates the prediction results with the knowledge in the knowledge graph, uses the knowledge stored in the knowledge graph to verify the output of the neural network, and feeds back the knowledge in the knowledge graph as prior information to the neural network for neural network optimization, until the fusion analysis and diagnosis results are consistent with the prediction through knowledge graph verification, and combines the accuracy of the knowledge graph database and the ability of the neural network to process high-dimensional complex data to ensure the comprehensiveness and accuracy of motor fault diagnosis and protection.

[0069] 3. The present invention generates adaptive control strategy instructions based on the fusion analysis and diagnosis results, combined with the real-time operating parameters of the motor and the operation and maintenance strategies and control rules stored in the knowledge graph. The generated adaptive control strategy instructions are applied to the motor and its related accessories through the corresponding actuators to achieve actual control of the motor operation. At the same time, the effect of the control instructions after execution is monitored in real time to ensure the flexibility and effectiveness of the execution strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 A flow chart of a control method for a motor protector of the present invention;

[0071] Figure 2 It is the principle block diagram of the motor protector of the present invention;

[0072] Figure 3 The voltage signal conditioning circuit diagram of the present invention;

[0073] Figure 4 The current signal conditioning circuit diagram of the present invention;

[0074] Figure 5 It is a temperature signal processing circuit diagram of the present invention;

[0075] Figure 6 It is the tripping circuit diagram of the present invention. DETAILED DESCRIPTION

[0076] In order to make the technical problems, technical solutions and advantages to be solved by the present invention more clear, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0077] Example 1

[0078] like Figure 1 As shown, the present invention proposes a control method for a motor protector, comprising the following steps:

[0079] S1. Knowledge graph construction, collecting knowledge data covering the basic principles, electrical characteristics, mechanical structure and characteristics, environmental influencing factors, and the causes, development process, common operation and maintenance and processing strategies of various types of faults of electric motors. Repetitive and redundant parts of knowledge data are removed and classified. After classification, knowledge data includes multi-level node attributes, and each level of node attributes includes node label, node type and node description;

[0080] S2. Use the resource description framework based on the semantic web to formalize the organized knowledge data, define various entities, relationships and attributes, and construct a conceptual model that reflects the knowledge system in the field of electric motors;

[0081] S3. Create a complete electric motor knowledge graph with the help of Neo4j, a knowledge graph construction tool, and use the graph database to store and manage the knowledge graph;

[0082] S4, collecting fault type data and operation factor data of the motor during normal operation and fault operation as a data set, preprocessing the data of the data set, and performing multi-level classification on the preprocessed data set, and generating a label for each fault type; the protection measures corresponding to the fault type include but are not limited to start timeout protection, underload and overload protection, stall protection, short circuit protection, residual current protection, three-phase current imbalance protection, overvoltage and undervoltage protection, and temperature protection;

[0083] S5. Based on the multi-classified data set, a state assessment and prediction model based on multi-classification neural network fusion is constructed and trained;

[0084] S6. Collect the operating factor data corresponding to each fault category to be predicted in real time, and each classification neural network outputs its prediction probability for each fault category respectively. Multiply the probability value of the same fault category output by each classification neural network by the corresponding weight coefficient and add them together to obtain the fused fault category probability value, and select the fault category with the largest probability value as the prediction result;

[0085] S7. Fusion the prediction results with the knowledge in the knowledge graph, using the knowledge stored in the knowledge graph to verify the output of the neural network, and at the same time feeding the knowledge in the knowledge graph as prior information to the neural network for neural network optimization, and continuously looping the verification and optimization steps until the fusion analysis diagnosis result is obtained through the knowledge graph verification and the prediction is consistent;

[0086] S8, based on the fusion analysis and diagnosis results, combined with the real-time operating parameters of the motor and the operation and maintenance strategies and control rules stored in the knowledge graph, generate adaptive control strategy instructions;

[0087] S9. The generated adaptive control strategy instructions are applied to the motor and its related accessories through the corresponding actuators to realize the actual control of the motor operation. At the same time, the effect of the control instructions after execution is monitored in real time. After the motor operation factor data is collected again by the sensor and fed back to step S6, the process of S6-S9 is repeated.

[0088] Among them, steps S1-S3 are the construction process of the knowledge graph, specifically:

[0089] 1. Knowledge graph construction steps:

[0090] We collect motor-related knowledge from multiple channels, including but not limited to professional motor textbooks, industry research reports, motor equipment manufacturing manuals, long-term operation and maintenance records, and practical experience of experts in the field. The knowledge collected covers the basic principles of motors (such as the electromagnetic induction principle of three-phase asynchronous motors, the working mechanism of synchronous motors, etc.), electrical characteristics (the current, voltage, power factor range and change law during normal operation, the abnormal performance of electrical parameters corresponding to various electrical faults, etc.), mechanical structure and characteristics (the detailed structure and function of components such as stators, rotors, bearings, and the vibration, noise, and temperature change characteristics under different mechanical fault conditions, etc.), environmental influencing factors (the influence of environmental conditions such as temperature, humidity, dust, and corrosive gases on motor performance and life), as well as the causes, development process, common operation and maintenance and processing strategies of various faults. We sort out and classify the massive amount of knowledge collected, remove duplicate and redundant parts, and form a clear and structured knowledge collection.

[0091] 2. Knowledge representation and modeling:

[0092] Using the knowledge representation method based on the Resource Description Framework (RDF) of the semantic web, the organized knowledge is formally represented, and various entities (such as individual motors, different component names, various fault types, environmental parameter categories, etc.) are defined, relationships (such as "composed of..." leading to "triggered under..." conditions", "corresponding to..." operation and maintenance strategies, etc.) and attributes (such as the specific values ​​of each electrical parameter, the severity level of the fault, the value range of environmental parameters, etc.) are defined to construct a conceptual model that can accurately reflect the knowledge system in the field of motors. For example, it is expressed in the form of triples as "bearing wear fault-leads to-increase in the vibration amplitude of the motor", "high temperature and high humidity environment-affects-degradation of the insulation performance of the motor", etc., which clearly presents the internal connection between knowledge.

[0093] 3. Knowledge graph creation and storage:

[0094] With the help of professional knowledge graph construction tools (such as Neo4j, Apache Jena, etc.), various knowledge entities, relationships and attributes are entered according to the knowledge representation model constructed above to create a complete motor knowledge graph. The knowledge graph is stored and managed using a graph database to facilitate subsequent efficient knowledge query, reasoning and update operations. During the construction process, the graph structure is continuously optimized to ensure the integrity and accuracy of the knowledge and the rationality of the logical relationship between the elements, making it a comprehensive and reliable motor knowledge resource library.

[0095] Step S4 collects the fault type data and operating factor data of the motor during normal operation and fault operation as a data set. First, the fault type data target is clarified, and the common fault types of the motor are focused on, such as bearing failure (including wear, rupture, etc.), winding failure (short circuit, open circuit, insulation damage, etc.), rotor failure (unbalance, broken bar, etc.), cooling system failure (coolant leakage, cooling fan damage, etc.) and electrical connection failure (loose wiring, poor contact, etc.). After clarifying these fault types, they can be marked and recorded in a targeted manner during the data collection process, which is convenient for the subsequent construction of a complete fault type data set. Then, determine the various operating factors that affect the operating state of the motor and need to be collected, mainly covering the following aspects:

[0096] Electrical parameters: such as three-phase current, three-phase voltage, power factor, motor input power, etc. These parameters can reflect the motor's electrical load and energy efficiency, and are of great significance for determining whether the motor is overloaded, missing a phase, or has an electrical fault.

[0097] Mechanical parameters: including the motor's speed, vibration amplitude, vibration frequency, axial movement distance, etc. The speed can reflect whether the motor's operating speed meets the requirements, while vibration-related parameters are critical for detecting whether there are faults in mechanical parts such as bearings and rotors. Abnormal axial movement distance may also indicate problems with the motor's internal structure.

[0098] Temperature parameters: Focus on collecting the temperature values ​​of key parts of the motor (such as windings, bearings, housing, etc.). Excessive temperature is often an important sign of motor failure. For example, winding overheating may be caused by overload, insulation damage leading to short circuit, etc., and abnormally high bearing temperature may indicate poor lubrication or increased wear.

[0099] Environmental parameters: Consider the environmental factors of the motor, such as ambient temperature, humidity, dust content, etc. Harsh environmental conditions may affect the heat dissipation and insulation performance of the motor, and thus cause failures. Collecting this data helps analyze the potential impact of the environment on the operation of the motor.

[0100] Finally, protection measures corresponding to the fault type are taken, including but not limited to start timeout protection, underload and overload protection, stall protection, short circuit protection, residual current protection, three-phase current imbalance protection, overvoltage and undervoltage protection, and temperature protection.

[0101] After data collection, the data set is preprocessed, mainly including: (1) data cleaning and verification: first, the integrity and rationality of the collected raw data are verified, and obvious outliers and duplicate data caused by sensor failure, communication interference, etc. are eliminated. At the same time, the missing values ​​are filled using data interpolation algorithms; (2) feature extraction and engineering, mining the potential feature information in the raw data, carrying out feature extraction work for different types of sensor data, and using feature engineering technology to combine, transform, and reduce the dimension of the extracted raw features; (3) data normalization and standardization, using normalization and standardization methods to unify the data scale.

[0102] Step S5 constructs and trains a state assessment and prediction model based on multi-classification neural network fusion based on the multi-classification data, specifically:

[0103] (1) Divide the preprocessed data set into n training sets and n test sets, where the training set and the test set are in a one-to-one correspondence, and n is an integer greater than 2; because it is assumed that the neural network training model of the present invention has multiple different models, and the type, attribute and quantity of the data set trained by each model are different.

[0104] (2) For each training set, the corresponding neural network model is used for training to obtain the trained neural network model; First, the present invention selects support vector machine classifier, random forest classifier, and LSTM neural network classifier, each of which has its own advantages. The present invention is used to achieve complementary advantages:

[0105] Support vector machine classifier The relationship between motor fault-related features and fault types is often not a simple linear relationship. The association between features such as vibration amplitude and frequency and different fault types may present complex nonlinearity. SVM can map the original feature space to a high-dimensional space with the help of kernel functions (such as the commonly used Gaussian kernel function), making the originally nonlinear data linearly separable, thereby effectively processing these complex nonlinear relationships and accurately diagnosing faults. The motor operation data may have certain noise interference (such as sensor measurement errors, etc.). SVM is trained based on the principle of structural risk minimization. It is relatively insensitive to noise and outliers in the data. It can maintain a relatively stable fault diagnosis performance in a data environment with interference, and will not cause excessive deviations in the diagnosis results due to a small amount of abnormal data.

[0106] The random forest classifier constructs multiple decision trees and allows these decision trees to classify the input motor operation data (also characteristic data such as temperature, vibration, current, etc.) (distinguishing different fault types and normal states), and then combines the results of each decision tree (usually using majority voting) to determine the final fault diagnosis result. During the construction process, each decision tree uses the self-service sampling method to sample the training data and randomly selects some features to grow the tree, so that there are differences between different decision trees to avoid overfitting. Motor operation data is usually more complex and is affected by many factors, and local feature fluctuations are prone to occur. By integrating multiple decision trees, random forests can effectively overcome the overfitting problem that may occur in a single decision tree, and maintain good adaptability to both training data and new test data, making the fault diagnosis results relatively stable and reliable. Even when there is a certain amount of noise in the data or some features are not accurate, a more reasonable diagnosis can be given.

[0107] The monitoring of the motor operating status involves numerous parameter indicators and high data dimensions. The random feature selection mechanism within the random forest can automatically screen out features that are important for fault diagnosis. There is no need for manual complex feature engineering screening in advance. It reduces the data dimension while focusing on key features, which helps to improve the accuracy and efficiency of fault diagnosis. It can also directly handle multi-category fault classification problems and can adapt well to the diagnosis scenarios of various motor fault types.

[0108] As a neural network that is good at processing sequence data, LSTM can organize the continuously collected motor operation data in time series (such as a series of temperature, vibration, current and other parameter values ​​arranged in time order) as input for motor fault diagnosis. It can capture long-term dependencies in data sequences. For example, the failure of a motor component may be caused by the gradual change and accumulation of related parameters over a period of time. LSTM can memorize the characteristic information of these changes. By learning the sequence data, it outputs the corresponding fault type probability and other results in the output layer to determine whether the motor has a fault and what kind of fault it is. The running state of the motor changes dynamically over time. Many faults will have some early signs reflected in the time series changes of the operating parameters before they occur. LSTM can make good use of these time series features to mine the long-term dependencies. For example, the motor bearing wear fault may have a slow increase in vibration amplitude and a gradual increase in temperature in the early stage. LSTM can accurately determine the occurrence and type of fault based on the data features accumulated over a long period of time. Compared with some methods that only focus on the data features of a single time point, the diagnosis is more forward-looking and accurate.

[0109] (3) After selecting the neural network architecture, some specific structural parameters need to be determined, such as:

[0110] The number of input layer nodes: depends on the number of features of the input data. For example, if the input is a 100-dimensional vector, the number of input layer nodes is set to 100.

[0111] Number of hidden layers and number of nodes in each hidden layer: More complex tasks may require more hidden layers and nodes, but too many may lead to overfitting. You can gradually increase the number of hidden layers and nodes and make adjustments by observing the performance of the model on the validation set.

[0112] The number of output layer nodes is related to the number of targets to be predicted. For example, if the present invention performs a multi-classification task (divided into 5 categories), the number of output layer nodes is usually set to 5; if the regression task predicts a value, the number of output layer nodes is 1.

[0113] Activation function selection: Select appropriate activation functions in the hidden layer and output layer to introduce nonlinear factors, so that the neural network can handle complex nonlinear relationships. The present invention inputs the fault factor data in each test set into the corresponding neural network model, and obtains m output result sets through Softmax function classification, where m is a positive integer.

[0114] (4) Select an optimization algorithm. Commonly used optimization algorithms include stochastic gradient descent (SGD) and its variants. The purpose of these optimization algorithms is to continuously adjust the weights and bias parameters of the neural network so that the value of the loss function (a function that measures the difference between the model's predicted results and the actual results) continues to decrease.

[0115] (5) Setting the loss function: Since the present invention is a multi-classification task, the cross entropy loss function is used. It measures the difference between the probability distribution of the model output and the probability distribution of the true label, which can make the model output closer to the true category probability distribution.

[0116] (6) Start training, input the training set data into the constructed neural network, and adjust the network weights and bias parameters through multiple iterations (i.e., multiple traversals of the training set data) according to the selected optimization algorithm and loss function, so that the value of the loss function gradually decreases. Each iteration process is called a training cycle (Epoch). During the training process, the loss value of each cycle and the changes in the evaluation indicators (such as accuracy, recall rate, mean square error, etc.) on the validation set are usually recorded to observe the training progress and performance changes of the model and to promptly detect whether problems such as overfitting occur.

[0117] (7) Finally, according to the historical prediction performance evaluation indicators of each classifier on the training set and the validation set, a corresponding weight coefficient is assigned to each classifier, and the classifiers are weightedly fused with the weight coefficient to obtain the state evaluation and prediction model of the multi-classification neural network fusion; among them, the historical prediction performance evaluation indicators include accuracy and F1 value.

[0118] It should be noted that the test set for training the above-mentioned neural network model adopted in the present invention includes a first test set, a second test set and a third test set. The first test set corresponds to a support vector machine classifier, the second test set corresponds to a random forest classifier, and the third test set corresponds to an LSTM neural network classifier.

[0119] The fault factor data in each test set are input into the corresponding neural network model to obtain m output result sets, where m is a positive integer, including:

[0120] Using the first test set as an input of the support vector machine classifier, and extracting a first feature in a discriminator of the support vector machine classifier;

[0121] The second test set is used as the input of the random forest classifier, and the second feature is extracted in the hidden layer of the random forest classifier;

[0122] The third test set is used as the input of the LSTM neural network classifier, and the third feature is extracted in the convolution layer of the LSTM neural network classifier;

[0123] The first feature, the second feature and the third feature are respectively input into the pre-trained multi-task learning neural network model to obtain m output result sets, where m is a positive integer, wherein the multi-task learning neural network model is used to implement learning tasks, and different learning tasks correspond to different output layers, and each output layer has different classification targets. The multi-task learning neural network model includes a hidden layer shared with a support vector machine classifier, a random forest classifier, and an LSTM neural network classifier. The training method of the multi-task learning neural network model includes:

[0124] Determine a network architecture of a multi-task learning neural network model, wherein the network architecture includes a hidden layer and m output layers, each output layer being used to output a corresponding learning task;

[0125] Create a loss function that integrates the loss of each view of the multi-task learning neural network model;

[0126] Construct a multi-task learning neural network model according to the network architecture, and use the target training set to train the multi-task learning neural network model;

[0127] When training a multi-task learning neural network model, the loss function is optimized in real time to optimize the learning performance of the multi-task learning neural network model in all perspectives;

[0128] Among them, the loss function includes:

[0129]

[0130] Where L is the sum of the losses of each view of the multi-task learning neural network model, L i is the loss of the i-th learning task, ω i is the weight corresponding to the loss of the i-th learning task.

[0131] After the state assessment and prediction model based on the fusion of multiple classification neural networks is trained in step S5, the operation factor data corresponding to each fault category to be predicted is collected in real time in step S6, and each classification neural network outputs its prediction probability for each fault category respectively. The probability values ​​of the same fault category output by each classification neural network are multiplied by the corresponding weight coefficient and then added to obtain the fused fault category probability value, and the fault category with the largest probability value is selected as the prediction result;

[0132] First, assume that there are multiple different types of classification neural networks involved in this fusion process, such as the support vector machine classifier mentioned above (which can be regarded as an extension of the output probability form, and its discrimination results are converted into probability form through some means), the random forest classifier (which can itself output category probability estimates) and the LSTM neural network classifier.

[0133] These classification neural networks each target the task of motor fault diagnosis. Based on the input motor operating status related data (such as vectors or time series data composed of parameter features such as temperature, vibration, current, etc., depending on the specific network's requirements for data format), they output the probability values ​​corresponding to each fault category. For example, assuming there are three types of motor faults, namely "bearing fault", "winding fault", "rotor imbalance fault" and the category of normal operation, then each classification neural network will output a probability result similar to the following after receiving new data to be diagnosed:

[0134] Support vector machine classifier output: bearing failure probability is 0.2, winding failure probability is 0.3, rotor imbalance failure probability is 0.1, and normal operation probability is 0.4;

[0135] Random forest classifier output: bearing failure probability is 0.3, winding failure probability is 0.25, rotor imbalance failure probability is 0.15, and normal operation probability is 0.3;

[0136] LSTM neural network classifier output: bearing failure probability is 0.25, winding failure probability is 0.35, rotor imbalance failure probability is 0.1, and normal operation probability is 0.3.

[0137] The weight coefficient is used to measure the importance of each classification neural network in the final fusion decision. Different classification neural networks contribute to different degrees to the final result due to their own characteristics, performance on specific data sets, and application scenario requirements, so it is necessary to set appropriate weight coefficients for them respectively. The present invention uses historical motor fault diagnosis data to assign weight coefficients, divides the data set into training set, verification set and test set, and uses each classification neural network for training and evaluation (for example, using evaluation indicators such as accuracy, recall rate, F1 value to measure their performance on the verification set and test set). The weight coefficient is assigned according to the performance of each network on these evaluation indicators. For example, if the random forest classifier has a high accuracy rate in past tests and stable overall performance, it can be given a relatively high weight coefficient; while the support vector machine classifier has advantages in processing small samples, but its accuracy rate is slightly lower on the current relatively large sample motor fault data set, so it can be assigned a relatively small weight coefficient.

[0138] Taking the category of "bearing fault" as an example, we will show how to obtain the fused fault category probability value. The calculation method for other fault categories ("winding fault", "rotor imbalance fault" and "normal operation") is the same.

[0139] The fused "bearing failure" probability value is calculated as follows:

[0140] P (bearing fault) = Psvm (bearing fault) × Wsvm + Prr (bearing fault) × WrF + PLstm (bearing fault) × WLSTM = 0.2x0.3 + 0.3x0.4 + 0.25x0.3 = 0.06 + 0.12 + 0.075 = 0.255

[0141] Similarly, the fused "winding fault" probability value can be calculated:

[0142] P(winding fault)=Psvm(winding fault)×Wsvm+Prr(winding fault)×WrF+PLsrm(winding fault)×WLSTM=0.3x0.3+0.25x0.4+0.35x0.3=0.09+0.1+0.105=0.295

[0143] Probability value of "rotor unbalance fault":

[0144] P (rotor unbalance fault) = Psvm (rotor unbalance fault) × WsvM + Prr (rotor unbalance fault) × WrF + Pusrm (rotor unbalance fault) × WLSTM = = 0.1x0.3 + 0.15x0.4 + 0.1x0.3 = 0.03 + 0.06 + 0.03 = 0.12

[0145] "Uptime" probability value:

[0146] P (normal operation) = Psvm (normal operation) × Wsvm + Prr (normal operation) × WrF + Pusrm (normal operation) × WLSTM = 0.4x0.3 + 0.3x0.4 + 0.3x0.3 = 0.12 + 0.12 + 0.09 = 0.33

[0147] In this example, the probability value of "normal operation" is 0.33, which is the largest, so it can be determined that the current motor is in normal operation.

[0148] In this embodiment, the knowledge stored in the knowledge graph is used to verify the output of the neural network, specifically:

[0149] The knowledge graph stores the fault phenomenon data and fault factor data of the motor during normal operation and fault operation, which are represented by nodes and edges and are defined as follows: x , R, E y ) triples, where E x is the head entity, R is the relationship, E y is the tail entity;

[0150] In the neural network, the category label y∈{1, 2, …, C} is defined. Given a fault factor data sample x corresponding to a category label, i , the predicted probability of a fault belonging to category c is P(y=c|x)

[0151] If the fault factor data sample x i and the head entity E in the knowledge graph x Correspondingly, the Pearson correlation coefficient is used to determine the relationship between category c and tail entity E. y The correlation coefficient r is used, and the predicted probability is the product of P(y=c|x) and the correlation coefficient as the final probability.

[0152] For example, if the output of the neural network is "bearing failure", the characteristics of the bearing failure are found in the knowledge graph as "the vibration amplitude exceeds the threshold, and the vibration frequency is concentrated in a certain range, accompanied by a slow increase in temperature", and then check whether the actual operation data input to the neural network meets these characteristic descriptions. If the vibration amplitude in the actual motor operation data does exceed the set threshold, the vibration frequency is also in the corresponding range, and the temperature rises slowly, then in this respect, the output of the neural network is consistent with the knowledge graph knowledge and passes the verification; otherwise, there may be a problem, and further analysis is needed to determine whether it is due to the neural network's misjudgment or the inaccurate knowledge of the knowledge graph.

[0153] The Pearson correlation coefficient is often used to measure the degree of linear correlation between two variables, and its value range is between [-1,1]. When the correlation coefficient is 1, it means that the two variables are completely positively linearly correlated; when it is -1, it means that they are completely negatively linearly correlated; when it is 0, it means that there is no linear correlation. In the context of using knowledge graphs to verify the consistency of neural network outputs, we can use it to measure the degree of correlation between the output of the neural network and the variables involved in the relevant knowledge in the knowledge graph to determine whether they are consistent.

[0154] For the present invention, from the perspective of neural network output: assuming that the neural network outputs probability values ​​corresponding to different fault types (for example, outputting bearing fault probability P1, winding fault probability P2, etc.), the probability of a specific fault type is selected as a variable. For example, the probability of "bearing fault" P1 is focused on.

[0155] From the perspective of knowledge graph: determine the relevant variables based on the characteristic knowledge corresponding to the fault type (such as bearing fault) in the knowledge graph. For example, the knowledge graph shows that bearing fault will increase the vibration amplitude, so we can select the actual collected motor vibration amplitude X (after appropriate data preprocessing, such as normalization) as another variable.

[0156] Collect multiple sets of neural network output results and the corresponding motor actual operation data (i.e., data related to the features in the knowledge graph). For example, there are cases where the motor is monitored and neural network diagnosed at different times, and each time the bearing failure probability P output by the neural network is recorded. 1i (i=1,2,……n, representing different moments) and the corresponding motor vibration amplitude X i , forming two sets of data sequences {P 11 , P 12 , ..., P 1n} and {X1, X2, ... X n}, the Pearson correlation coefficient is calculated using the following formula:

[0157]

[0158] in, is the mean value of the bearing failure probability P1, is the mean of the amplitude values ​​X.

[0159] In this embodiment, the knowledge in the knowledge graph is fed back to the neural network as prior information to optimize the neural network, specifically:

[0160] Convert entities and relationships in the knowledge graph into low-dimensional vector representations through a specific embedding algorithm;

[0161] Integrate the generated entity vectors and relationship vectors into the input layer or hidden layer of the neural network, so that the neural network can use the knowledge carried by these vectors during training and reasoning;

[0162] Logical rules are extracted from the knowledge graph and converted into constraints in the neural network training process.

[0163] In this embodiment, the decision tree algorithm is used to classify the knowledge data of the knowledge graph and to perform multi-level classification on the preprocessed data set:

[0164] The CART algorithm is selected for motor fault type classification. The Gini index is used as the feature selection standard when building the decision tree. The calculation formula of the Gini index is:

[0165] Where D represents the data set, K is the number of fault type categories, and p k is the proportion of samples belonging to the kth fault type in the data set to the total samples;

[0166] Calculate the Gini index of all available features in the data set, select the feature with the smallest Gini index as the basis for dividing the root node, and for feature A, divide the data set D into D1, D2, ..., Dn subsets according to its different values. Then the Gini index of feature A is

[0167] For each subset divided from the root node, repeat the above process of calculating the Gini index and selecting the minimum Gini index feature to construct child nodes, continuously branching and continuously dividing the data set;

[0168] Set the stopping condition. If all samples in a node belong to the same fault type, that is, the classification purity reaches the highest, then the node will be used as a leaf node and will not be divided further.

[0169] In this embodiment, based on the fusion analysis and diagnosis results, combined with the real-time operating parameters of the motor and the operation and maintenance strategies and control rules stored in the knowledge graph, adaptive control strategy instructions are generated, specifically:

[0170] According to the current running state of the motor and the fault diagnosis result, a comprehensive judgment is made on the current running state of the motor, wherein the running state includes the normal running state, the fault warning state, and the fault response state;

[0171] Search the knowledge graph for relevant knowledge that matches the current fault type and motor operating status, and perform operation and maintenance strategy matching and control rule matching. The operation and maintenance strategy matching includes generating operation and maintenance suggestions for specific faults and operation and maintenance adjustments based on the operating status. The control rule matching includes adjusting the control threshold under the fault state and generating control logic for different combinations of faults and operating states.

[0172] Generate adaptive control strategy instructions for different fault types, where the different fault types include electrical faults, mechanical faults, and other types of faults (such as cooling system faults and control circuit faults).

[0173] The real-time operating parameters of the above motors include current, voltage, speed, temperature, etc. When the fusion analysis and diagnosis results diagnose the following parameter faults, specifically:

[0174] Current: The change in current after a fault occurs. For example, a short circuit in the winding may cause a sharp increase in current. A stuck mechanical transmission component may cause the motor to stall, and the current will instantly soar to the stall current value. Comparing the current curve during normal operation with the current current value can help determine the severity of the fault and the extent of damage to the motor.

[0175] Voltage: Partial voltage failure (such as loose electrical connections causing increased contact resistance) may cause voltage drop, affecting the normal operating voltage range of the motor, and further affecting the motor's output torque and speed.

[0176] Speed: Rotor imbalance, mechanical transmission problems, etc. may cause speed instability, fluctuations, declines, or even failure to reach the rated speed. The control strategy related to the speed needs to be formulated based on the specific performance of the current speed.

[0177] Temperature: Temperature data of key parts of the motor (windings, bearings, etc.). Faults are often accompanied by abnormal temperature increases. For example, severe bearing wear generates excessive friction heat, and heat dissipation system failure causes heat to be unable to be dissipated in time, which will cause the temperature to exceed the normal range. Excessive temperature may further cause other secondary faults, so temperature parameters are very important for judging the urgency of the control strategy.

[0178] Then, combined with the above comprehensive evaluation of the operating status, according to the changes in the above operating parameters and the fault diagnosis results, a comprehensive judgment is made on the current operating status of the motor, such as whether it can temporarily operate with a fault but needs to be closely monitored, or whether it is an emergency state that must be stopped immediately to avoid more serious damage.

[0179] Next, search the knowledge graph for relevant knowledge that matches the current fault type and motor operating status, including:

[0180] 1. Operation and maintenance strategy matching:

[0181] (1) Operation and maintenance suggestions for specific faults:

[0182] For a winding short-circuit fault, if it is a minor short circuit and the motor can still run, the knowledge graph may recommend regular inspections of the resistance and temperature changes of the short-circuited winding, and recording of relevant data to keep track of the fault development trend; if it is a serious short circuit, it will prompt the need to arrange professional maintenance personnel to repair or replace the winding as soon as possible, and give the approximate time period and precautions for maintenance.

[0183] In terms of bearing wear, depending on the degree of wear, the operation and maintenance strategy covers different operational suggestions from simple grease replenishment to bearing replacement, and also indicates the timing and specific process of the corresponding operation. For example, when the bearing wear reaches a certain degree, the bearing should be replaced when the motor is next shut down for maintenance, and the preparations required before replacing the bearing should be explained in detail (such as the correct steps for removing the motor end cover, etc.).

[0184] (2) Operation and maintenance adjustment based on operating status:

[0185] If the motor is in a state where it can temporarily run with a fault but needs to be monitored, the knowledge graph will provide a list of items that require key maintenance during this period. For example, the frequency of testing relevant operating parameters (current, temperature, etc.) can be increased from the conventional once per minute to once per minute. At the same time, for some vulnerable parts (such as belts, couplings, etc., if any are involved), spare parts should be prepared in advance so that they can be replaced in time if further problems arise.

[0186] 2 Control rule matching:

[0187] (1) Control threshold adjustment under fault conditions:

[0188] Taking into account that the existence of a fault may cause the motor operating parameters to exceed the normal range, the corresponding temporary control thresholds will be set in the knowledge graph. For example, during normal operation, a current exceeding 120% of the rated current is considered an overload and measures must be taken. However, in the event of a winding short-circuit fault and temporary operation with a fault, this overload current threshold may be appropriately relaxed to 150% (the specific value needs to be comprehensively judged based on the actual tolerance of the motor and the impact of the fault). At the same time, other parameters of the motor are closely monitored. Once the newly set threshold is reached, the corresponding control instructions are executed.

[0189] (2) Control logic for different combinations of faults and operating states:

[0190] For example, when there is a bearing wear failure and the motor speed fluctuates unstably and the temperature rises slightly, the control rule may be to first try to reduce the pressure on the bearing by adjusting the load of the motor, and observe whether the speed and temperature can return to stability. If there is no improvement within the specified time, an early warning message will be issued and preparations will be made to shut down for maintenance. For example, in the case of a cooling system failure causing the temperature to continue to rise, the control rule will guide the start of backup cooling measures (if any), such as switching to a backup cooling fan or opening an auxiliary water cooling circuit (for water-cooled motors), and gradually reducing the motor load power at the rhythm of each liter of altitude until the temperature drops to a safe range or reaches the temperature critical value that requires shutdown.

[0191] Through the above systematic steps, by making full use of fault diagnosis information, the current operating status of the motor and the professional knowledge in the knowledge graph, it is possible to generate control strategy instructions that fit the actual situation and have adaptive characteristics, so as to maximize the safety and stable operation of the motor and reduce the losses and impacts caused by faults.

[0192] The above-mentioned adaptive control strategy instructions are specifically:

[0193] Normal operating state strategy: When the evaluation model determines that the motor is in a normal operating state, the control method monitors the changes of various operating parameters in real time, and dynamically adjusts the threshold range of each monitoring parameter through the fuzzy logic optimization algorithm based on historical operating data and current working conditions;

[0194] Fault warning status strategy: If the evaluation model determines that the motor is in a minor fault state, a fault warning signal is issued and the motor's operating parameters are automatically adjusted;

[0195] Fault response status strategy: When the motor is judged to be in a moderate or severe fault state, the corresponding protection action is triggered immediately, and the key information of various operating parameters and fault types when the fault occurs is recorded.

[0196] Example 2

[0197] The present invention also proposes a motor protector corresponding to the above control method, comprising: a controller, an interface circuit module, a signal acquisition module, a signal processing module, a human-machine interface module, an execution module, and a power module, wherein the controller is electrically connected to the interface circuit module, the interface circuit module, the human-machine interface module, the execution module, and the power module, the input end of the signal processing module is electrically connected to the signal acquisition module, the power module is also electrically connected to the signal processing module and the execution module, and the interface circuit module is communicatively connected to a computer;

[0198] like Figure 3The figure shows the voltage signal conditioning circuit diagram. Before the signal is sent to the controller (single-chip microcomputer), it must be processed by the processing circuit. The processing circuit needs to complete the following tasks: amplify or reduce the collected electrical signal, filter the signal carrying harmonics, and rectify the AC signal. When the signal that the controller performs AD conversion is a voltage signal, the voltage range is 0-3.3V. In order to ensure that there is no over-range situation during design, the maximum voltage can be designed to be less than 3.3V.

[0199] The processing principle of current signal is basically the same as that of voltage signal. The only difference is in the processing of short-circuit current. When a short-circuit fault occurs, the phase current in the stator winding will increase rapidly. At this time, the voltage value may exceed the maximum range of the AD converter, and the controller may be burned if it is sent. In order to limit the current and avoid voltage overrange, the following circuit is designed: Figure 4 The current signal conditioning circuit diagram is shown.

[0200] The temperature signal processing circuit is different from the current and voltage processing circuits. It uses a negative temperature coefficient thermistor NTC as a sensor. The software is used to set the output of the pin of OPT2 close to the resistor R14 to high before determining that temperature protection is needed. + and T - When it is suspended, U3A outputs a high level, OPT2 is turned on, and its output close to resistor R14 is low. + and T - The resistance between them decreases until U3A outputs a low level, OPT2 is cut off, and its output close to resistor R14 is high. The program then determines that temperature protection is needed. Figure 5 shown.

[0201] When a relatively minor fault is diagnosed in the motor, only an alarm (such as a buzzer circuit) is required. When a more serious fault is diagnosed, a trip circuit is used for protection. The trip circuit is different from the alarm circuit. It is a high-voltage circuit and requires the relay to be disconnected. Its weak-current part can be connected to the pin of the controller, while its high-current part uses the relay to achieve the tripping action. The tripping process is as follows: when the motor fails and needs to trip, the controller sends a tripping signal. After the signal is isolated by the optical coupler, it finally disconnects the circuit through the intermediate relay. The trip circuit is as follows: Figure 6 shown.

[0202] The software modules running on the computer include:

[0203] The knowledge graph construction module is used to collect knowledge data covering electric motors, remove duplicate and redundant parts of the knowledge data and classify them. The classified knowledge data includes multi-level node attributes, and each level of node attributes includes node labels, node types and node descriptions. The resource description framework based on the semantic web is used to formally represent the organized knowledge data, define various entities, relationships and attributes, and construct a conceptual model that reflects the knowledge system in the field of electric motors. The knowledge graph construction tool Neo4j is used to create a complete knowledge graph for electric motors, and the graph database is used to store and manage the knowledge graph.

[0204] A data set construction module is used to collect fault type data and operating factor data during normal operation and fault operation of the motor as a data set, preprocess the data of the data set, perform multi-level classification on the preprocessed data set, and generate a label for each fault type; the protection measures corresponding to the fault type include but are not limited to start timeout protection, underload and overload protection, stall protection, short circuit protection, residual current protection, three-phase current imbalance protection, overvoltage and undervoltage protection, and temperature protection;

[0205] The state assessment and prediction model building module is used to build and train a state assessment and prediction model based on multi-classification neural network fusion based on the multi-classification data set;

[0206] The fault probability prediction module is used to collect the operating factor data corresponding to each fault category to be predicted in real time. Each classification neural network outputs its predicted probability for each fault category respectively. The probability values ​​of the same fault category output by each classification neural network are multiplied by the corresponding weight coefficient and then added to obtain the fused fault category probability value. The fault category with the largest probability value is selected as the prediction result.

[0207] An analysis and diagnosis module is used to fuse the prediction results with the knowledge in the knowledge graph, use the knowledge stored in the knowledge graph to verify the output of the neural network, and feed back the knowledge in the knowledge graph as prior information to the neural network for neural network optimization, and continuously cycle the verification and optimization steps until the fusion analysis and diagnosis results are obtained through the knowledge graph verification and the prediction are consistent;

[0208] The control instruction generation module is used to generate adaptive control strategy instructions based on the fusion analysis and diagnosis results, combined with the real-time operating parameters of the motor and the operation and maintenance strategies and control rules stored in the knowledge graph;

[0209] The instruction control module is used to apply the generated adaptive control strategy instructions to the motor and its related accessories through the corresponding actuators to achieve actual control of the motor operation; at the same time, it monitors the effect of the control instructions in real time, collects the motor's operating factor data again through sensors, and feeds it back to the fault probability prediction module, and then repeats the subsequent process.

[0210] The operation of the above-mentioned software modules is referred to in Example 1 and will not be described in detail here.

[0211] In summary, the present invention constructs a motor fault knowledge base through a knowledge graph to provide data reference for the prediction and optimization of the neural network model. Multiple types of neural networks are constructed to evaluate and predict different types of motor faults, and the probability values ​​of the same fault category output by each classification neural network are multiplied by the corresponding weight coefficient and added to obtain the fused fault category probability value. The fault category with the largest probability value is selected as the prediction result, making full use of the advantages of various neural networks in processing different types of data.

[0212] By fusing the prediction results with the knowledge in the knowledge graph, the knowledge stored in the knowledge graph is used to verify the output of the neural network, and the knowledge in the knowledge graph is fed back to the neural network as prior information for neural network optimization, until the fusion analysis and diagnosis results are consistent with the prediction through knowledge graph verification. The accuracy of the comprehensive knowledge graph database and the ability of the neural network to process high-dimensional complex data are combined to ensure the comprehensiveness and accuracy of motor fault diagnosis and protection.

[0213] Based on the fusion analysis and diagnosis results, combined with the real-time operating parameters of the motor and the operation and maintenance strategies and control rules stored in the knowledge graph, adaptive control strategy instructions are generated. The generated adaptive control strategy instructions are applied to the motor and its related accessories through the corresponding actuators to achieve actual control of the motor operation. At the same time, the effect of the control instructions after execution is monitored in real time to ensure the flexibility and effectiveness of the execution strategy.

[0214] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A control method for a motor protector, characterized in that: The following steps are involved: S1. Knowledge graph construction, collecting knowledge data covering the basic principles, electrical characteristics, mechanical structure and characteristics, environmental influencing factors, and causes, development processes, common operation and maintenance and processing strategies of various types of faults of electric motors, removing duplicate and redundant parts of the knowledge data and classifying them. The knowledge data after classification includes multi-level node attributes, and the node attributes of each level include node labels, node types and node descriptions; S2. Use the resource description framework based on the semantic web to formalize the organized knowledge data, define various entities, relationships and attributes, and construct a conceptual model that reflects the knowledge system in the field of electric motors; S3. Create a complete electric motor knowledge graph with the help of Neo4j, a knowledge graph construction tool, and use the graph database to store and manage the knowledge graph; S4, collecting fault type data and operation factor data of the motor during normal operation and fault operation as a data set, preprocessing the data of the data set, and performing multi-level classification on the preprocessed data set, and generating a label for each fault type; the protection measures corresponding to the fault type include but are not limited to start timeout protection, underload and overload protection, stall protection, short circuit protection, residual current protection, three-phase current imbalance protection, overvoltage and undervoltage protection, and temperature protection; S5. Based on the multi-classified data set, a state assessment and prediction model based on multi-classification neural network fusion is constructed and trained; S6. Collect the operating factor data corresponding to each fault category to be predicted in real time, and each classification neural network outputs its prediction probability for each fault category respectively. Multiply the probability value of the same fault category output by each classification neural network by the corresponding weight coefficient and add them together to obtain the fused fault category probability value, and select the fault category with the largest probability value as the prediction result; S7, integrating the prediction result with the knowledge in the knowledge graph, using the knowledge stored in the knowledge graph to verify the output of the neural network, and at the same time feeding back the knowledge in the knowledge graph as prior information to the neural network for neural network optimization, and continuously looping the verification and optimization steps until the fusion analysis diagnosis result is obtained through the knowledge graph verification and the prediction is consistent; S8, based on the fusion analysis and diagnosis results, combined with the real-time operating parameters of the motor and the operation and maintenance strategies and control rules stored in the knowledge graph, generate adaptive control strategy instructions; S9. The generated adaptive control strategy instructions are applied to the motor and its related accessories through the corresponding actuators to realize the actual control of the motor operation. At the same time, the effect of the control instructions after execution is monitored in real time. After the motor operation factor data is collected again by the sensor and fed back to step S6, the process of S6-S9 is repeated.

2. The control method of the motor protector according to claim 1, characterized in that: The state assessment and prediction model based on multi-classification neural network fusion is constructed and trained based on the multi-classification data, specifically: Divide the preprocessed data set into n training sets and n test sets, where the training set and the test set are in one-to-one correspondence, and n is an integer greater than 2; For each training set, the corresponding neural network model is used for training to obtain a trained neural network model; The fault factor data in each test set are respectively input into the corresponding neural network model, and m output result sets are obtained by classification through the Softmax function, where m is a positive integer; According to the historical prediction performance evaluation indicators of each classifier on the training set and the validation set, a corresponding weight coefficient is assigned to each classifier, and the classifiers are weightedly fused in combination with the weight coefficient to obtain a state evaluation and prediction model of multi-classification neural network fusion; wherein the historical prediction performance evaluation indicators include accuracy and F1 value.

3. The control method of the motor protector according to claim 2, characterized in that: The neural network model includes a support vector machine classifier, a random forest classifier, and an LSTM neural network classifier: The test set includes a first test set, a second test set and a third test set, the first test set corresponds to the support vector machine classifier, the second test set corresponds to the random forest classifier, and the third test set corresponds to the LSTM neural network classifier; The fault factor data in each test set are respectively input into the corresponding neural network model to obtain m output result sets, where m is a positive integer, including: Using the first test set as input of the support vector machine classifier, and extracting a first feature in a discriminator of the support vector machine classifier; Using the second test set as input of the random forest classifier, and extracting a second feature in a hidden layer of the random forest classifier; Using the third test set as input of the LSTM neural network classifier, and extracting a third feature in the convolutional layer of the LSTM neural network classifier; The first feature, the second feature and the third feature are respectively input into a pre-trained multi-task learning neural network model to obtain m output result sets, where m is a positive integer, wherein the multi-task learning neural network model is used to implement learning tasks, and different learning tasks correspond to different output layers, and each output layer has a different classification target, and the multi-task learning neural network model includes a hidden layer shared with the support vector machine classifier, the random forest classifier, and the LSTM neural network classifier.

4. The control method of the motor protector according to claim 3, characterized in that: The training method of the multi-task learning neural network model includes: Determine a network architecture of the multi-task learning neural network model, wherein the network architecture includes a hidden layer and m output layers, each of the output layers being used to output a corresponding learning task; Creating a loss function, the loss function is used to integrate the loss of each perspective of the multi-task learning neural network model; Constructing the multi-task learning neural network model according to the network architecture, and training the multi-task learning neural network model using a target training set; When training the multi-task learning neural network model, optimizing the loss function in real time so that the learning performance of the multi-task learning neural network model in all the perspectives is optimal; Wherein, the loss function includes: Where L is the sum of the losses of each perspective of the multi-task learning neural network model, L i is the loss of the i-th learning task, ω i is the weight corresponding to the loss of the i-th learning task.

5. The control method of the motor protector according to claim 1, characterized in that: The use of the knowledge stored in the knowledge graph to verify the output of the neural network is specifically as follows: The knowledge graph stores the fault phenomenon data and fault factor data knowledge set of the motor during normal operation and fault operation, represented by nodes and edges, and is defined as follows: x , R, E y ) triples, where E x is the head entity, R is the relationship, E y is the tail entity; The neural network defines a category label y∈{1, 2, ..., C}, and a fault factor data sample x corresponding to a category label is given. i , the predicted probability of a fault belonging to category c is P(y=c|x) If the fault factor data sample x i and the head entity E in the knowledge graph x Correspondingly, the Pearson correlation coefficient is used to determine the relationship between category c and tail entity E. y The correlation coefficient r is used, and the predicted probability is the product of P(y=c|x) and the correlation coefficient as the final probability.

6. The control method of the motor protector according to claim 1, characterized in that: The method of feeding back the knowledge in the knowledge graph as prior information to the neural network for neural network optimization is specifically as follows: Convert entities and relationships in the knowledge graph into low-dimensional vector representations through a specific embedding algorithm; Integrate the generated entity vectors and relationship vectors into the input layer or hidden layer of the neural network, so that the neural network can use the knowledge carried by these vectors during training and reasoning; Logical rules are extracted from the knowledge graph and converted into constraints in the neural network training process.

7. The control method of the motor protector according to claim 1, characterized in that: The classification of the knowledge data of the knowledge graph and the multi-level classification of the preprocessed data set both use the decision tree algorithm: The CART algorithm is selected for motor fault type classification. The Gini index is used as the feature selection standard when building the decision tree. The calculation formula of the Gini index is: Where D represents the data set, K is the number of fault type categories, and p k is the proportion of samples belonging to the kth fault type in the data set to the total samples; Calculate the Gini index of all available features in the data set, select the feature with the smallest Gini index as the basis for dividing the root node, and for feature A, divide the data set D into D1, D2, ..., Dn subsets according to its different values. Then the Gini index of feature A is For each subset divided from the root node, repeat the above process of calculating the Gini index and selecting the minimum Gini index feature to construct child nodes, continuously branching and continuously dividing the data set; Set the stopping condition. If all samples in a node belong to the same fault type, that is, the classification purity reaches the highest, then the node will be used as a leaf node and will not be divided further.

8. The control method of the motor protector according to claim 1, characterized in that: The method generates adaptive control strategy instructions based on the fusion analysis diagnosis results, combined with the real-time operating parameters of the motor and the operation and maintenance strategies and control rules stored in the knowledge graph, specifically: According to the current running state of the motor and the fault diagnosis result, a comprehensive judgment is made on the current running state of the motor, wherein the running state includes the normal running state, the fault warning state, and the fault response state; Search the knowledge graph for relevant knowledge that matches the current fault type and motor operating status, and perform operation and maintenance strategy matching and control rule matching. The operation and maintenance strategy matching includes generating operation and maintenance suggestions for specific faults and operation and maintenance adjustments based on the operating status. The control rule matching includes adjusting the control threshold under the fault state and generating control logic for different combinations of faults and operating states. Generate adaptive control strategy instructions for different fault types, where the different fault types include electrical faults, mechanical faults, and other faults.

9. The control method of the motor protector according to claim 8, characterized in that: The adaptive control strategy instruction is specifically: Normal operating state strategy: When the evaluation model determines that the motor is in a normal operating state, the control method monitors the changes of various operating parameters in real time, and dynamically adjusts the threshold range of each monitoring parameter through the fuzzy logic optimization algorithm based on historical operating data and current working conditions; Fault warning status strategy: If the evaluation model determines that the motor is in a minor fault state, a fault warning signal is issued and the motor's operating parameters are automatically adjusted; Fault response status strategy: When the motor is judged to be in a moderate or severe fault state, the corresponding protection action is triggered immediately, and the key information of various operating parameters and fault types when the fault occurs is recorded.

10. A motor protector corresponding to the control method according to any one of claims 1 to 9, characterized in that: include: A controller, an interface circuit module, a signal acquisition module, a signal processing module, a human-machine interface module, an execution module, and a power module. The controller is electrically connected to the interface circuit module, the interface circuit module, the human-machine interface module, the execution module, and the power module. The input end of the signal processing module is electrically connected to the signal acquisition module. The power module is also electrically connected to the signal processing module and the execution module. The interface circuit module is connected to the computer for communication. The software modules running on the computer include: The knowledge graph construction module is used to collect knowledge data covering electric motors, remove duplicate and redundant parts of the knowledge data and classify them. The knowledge data after classification includes multi-level node attributes, and the node attributes of each level include node labels, node types and node descriptions; the sorted knowledge data is formally represented by using a resource description framework based on the semantic web, and various entities, relationships and attributes are defined to construct a conceptual model reflecting the knowledge system in the field of electric motors; a complete knowledge graph of electric motors is created with the help of the knowledge graph construction tool Neo4j, and the knowledge graph is stored and managed using a graph database; A data set construction module is used to collect fault type data and operating factor data during normal operation and fault operation of the motor as a data set, preprocess the data of the data set, perform multi-level classification on the preprocessed data set, and generate a label for each fault type; the protection measures corresponding to the fault type include but are not limited to start timeout protection, underload and overload protection, stall protection, short circuit protection, residual current protection, three-phase current imbalance protection, overvoltage and undervoltage protection, and temperature protection; The state assessment and prediction model building module is used to build and train a state assessment and prediction model based on multi-classification neural network fusion based on the multi-classification data set; The fault probability prediction module is used to collect the operating factor data corresponding to each fault category to be predicted in real time. Each classification neural network outputs its predicted probability for each fault category respectively. The probability values ​​of the same fault category output by each classification neural network are multiplied by the corresponding weight coefficient and then added to obtain the fused fault category probability value. The fault category with the largest probability value is selected as the prediction result. An analysis and diagnosis module is used to fuse the prediction results with the knowledge in the knowledge graph, use the knowledge stored in the knowledge graph to verify the output of the neural network, and feed back the knowledge in the knowledge graph as prior information to the neural network for neural network optimization, and continuously cycle the verification and optimization steps until the fusion analysis and diagnosis results are obtained through the knowledge graph verification and the prediction are consistent; The control instruction generation module is used to generate adaptive control strategy instructions based on the fusion analysis and diagnosis results, combined with the real-time operating parameters of the motor and the operation and maintenance strategies and control rules stored in the knowledge graph; The instruction control module is used to apply the generated adaptive control strategy instructions to the motor and its related accessories through the corresponding actuators to achieve actual control of the motor operation; at the same time, it monitors the effect of the control instructions in real time, collects the motor's operating factor data again through sensors, and feeds it back to the fault probability prediction module, and then repeats the subsequent process.

Images