Method for diagnosing hydraulic machine system fault through AI model
By employing AI models for fault diagnosis in hydraulic press systems, the problem of difficult fault location in hydraulic press systems has been solved, enabling real-time and accurate fault diagnosis and repair, and reducing operation and maintenance costs.
Patent Information
- Application Number
- CN202411556508.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-04
- Publication Date
- 2024-11-29
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing technologies lack effective online monitoring methods for fault diagnosis in hydraulic press systems, resulting in high costs for fault location and repair, and requiring highly skilled maintenance personnel.
AI models are used to diagnose faults in hydraulic press systems, including data acquisition, preprocessing, feature extraction, and model training. Data is collected using sensors, and real-time fault diagnosis is achieved through feature engineering and model training. A comprehensive evaluation is then performed using a knowledge base.
It enables real-time fault diagnosis of hydraulic press systems, reduces maintenance costs, and improves the accuracy and efficiency of fault location.
Smart Images

Figure CN119042195A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of hydraulic press system fault diagnosis, and in particular to a method for diagnosing hydraulic press system faults through an AI model. Background Art
[0002] In recent years, with the continuous improvement of the digitalization level of heavy hydraulic presses, their components and subsystems have become more and more complex. Damage to any component may change the working state of the entire hydraulic system. Before the failure occurs, due to the lack of timely monitoring of the system status, the factors that may cause the failure cannot be located and eliminated in a timely and accurate manner. However, the traditional fault, manual troubleshooting and full-site maintenance methods have extremely high requirements for operation and maintenance personnel, and users also invest high labor costs and supporting funds for maintenance and repair. In order to improve the fault diagnosis method and maintenance service quality, the AI model is used to diagnose the fault of the hydraulic press for online monitoring, so that maintenance and fault diagnosis can be carried out in time when the hydraulic system operation status is abnormal. Summary of the invention
[0003] The purpose of the present invention is to solve the problems existing in the prior art and to propose a method for diagnosing hydraulic press system faults through an AI model.
[0004] In order to achieve the above object, the present invention adopts the following technical solutions: A method for diagnosing hydraulic press system faults through an AI model, wherein the hydraulic press system includes a data acquisition module, a data preprocessing module, a feature extraction module, and an AI model training module, and the method includes the following steps: S1, data collection; S2, data preprocessing; S3, feature extraction stage; S4, model training stage; S5, model result processing and visualization.
[0005] Furthermore, in said S1, at least one or more types of sensors are arranged on the hydraulic press to collect data of different categories respectively; in said S2, data sets are collected, class labels are determined, and the collected label data are standardized, and the data is changed into a distribution with a mean of 0 and a standard deviation of 1 by subtracting the mean from the original data and then dividing it by the standard deviation, and the standardized output range is from negative infinity to positive infinity: ;in, represents a data point in the original data. represents the mean of the data set, Represents the standard deviation of the data set; the steps of data processing are: S21: import the data set, which includes a training set and a test set respectively; S22: view the feature columns and label columns of the training set and the test set respectively; S23: view the first five rows of data, clarify the project requirements, and determine the data to be predicted; S24: view the feature fields, clarify the meaning of each column of data, and the format of the data; S25: view the information of the column where the data is located; S26: check whether there are outliers in the data and the normal distribution transformation; The S3 includes standardizing the collected data and then performing feature engineering, including finding the mean to reflect the central trend, finding the variance to measure the degree of dispersion, and analyzing the trend pattern by time aggregation to determine the possible faults and fault types that may occur at a certain moment or within a certain period of time; determining the fault type by comparative analysis, combining domain knowledge and building a model, and its main steps include S31, checking missing values; S32, checking the data missing rate; S33, filling missing values; S34, feature process; S35, encoding the data; S36, checking feature importance.
[0006] Furthermore, the S4 includes: S41, weight initialization: initializing the weight to a very small random number to ensure that the processing units in the network have different weights at the beginning of training, can learn different features, provide a random starting point for subsequent calculations, ensure that the algorithm will not fall into a local optimal solution, and increase the algorithm exploration ability; S42, selecting input samples: randomly selecting a data point in the original data As input sample " ", the input sample " ” is compared and calculated with the nodes in the competition layer in subsequent steps.
[0007] Furthermore, the S4 also includes S43, determining the winning node: including the following two steps: (1) traversing each node in the competition layer and calculating the input sample " ” and the similarity between the nodes, the similarity calculation method is determined according to the specific problem and data type, and the calculation method includes the Euclidean distance method or the cosine similarity method; (2) select the node with the smallest distance as the winning node; S44, determine the winning neighborhood and update amplitude: determine the nodes to be included in the winning neighborhood according to the neighborhood radius; calculate the update amplitude of each of these nodes through the "neighborhood function", the closer to the winning node, the larger the update amplitude, and the farther away from the winning node, the smaller the update amplitude.
[0008] Furthermore, the S4 also includes S45, updating the weights of the nodes in the winning neighborhood, using the following formula: To update, " is a function that depends on the input , Output and time step The learning rate or adjustment factor of " represents the weight at time step s;" ” means at time step The updated weights are ” is another time-step dependent The factor of gradually decreases with the increase of the number of iterations to control the amplitude of the update; ” is the target value or desired output value.
[0009] Furthermore, S4 also includes S46 iteration. After completing one round of iteration, the number of iterations is increased by one, and then the process returns to S42 to continue to randomly select an input sample. "The next round of iterations is carried out until the set number of iterations is met. In each round of iteration, the winning node and the nodes in its neighborhood will be updated according to the input sample, and the weights will be gradually adjusted to better adapt to the input sample." "By repeating these steps, the algorithm can gradually learn the input sample" " features and patterns to achieve the input sample " "Clustering or classification task; set up a simple linear regression model, and its weight update formula is: ,in, is the weight vector at iteration time step s, is the iteration time step The updated weight vector is is the learning rate, is the loss function About weight The gradient of is the input sample, is the corresponding target value; The calculation process in each iteration includes the following: randomly select an input sample and target value ; Calculate the predicted value: ; Calculate the loss: ; Compute the gradient: ; Update weights: ; Iteration steps: Initialization: Setup , that is, starting from the first iteration, Iteration process: Repeat the above steps. Termination condition: Continue iterating until ,in Represents the total number of preset iterations.
[0010] Furthermore, the S5 includes S51: after the model training is completed, when subsequent real-time data is input, the model will automatically play a role in real-time diagnosis; S52: the model will analyze the real-time data according to the patterns and features learned during the training process to determine whether there is a fault and the possible type of fault; S53: the model will call the knowledge base, which stores a large amount of professional knowledge, historical fault cases and related feature information; S54: by comparing the fault category diagnosed by the model with the content in the knowledge base, a comprehensive assessment of the credibility of the fault category is made.
[0011] Furthermore, the S54 also includes S541: if the model diagnosis result is highly consistent with the typical fault characteristics in the knowledge base, then the credibility of the fault category is high; S542: if the two are very different, the credibility is low; S543: the model will also give the probability of the fault occurring based on the similarity between the real-time data and the existing fault data and the statistical characteristic factors of the data; S544: the probability can provide an important reference for decision makers so that they can take corresponding prevention and treatment measures in time.
[0012] Furthermore, in each iteration, the model parameters are updated according to the input samples to better adapt to the data distribution. In the round iteration, the parameters are The update is expressed as: ,in, Indicated in The parameters of the model after rounds of iterations, Indicated in The parameters of the model after rounds of iterations, is the learning rate, Represents the loss function In the current parameters The gradient at; Set the loss function ,in is the actual model value, is the prediction model value; set the prediction model value , the gradient is calculated as: In the first round of iteration, the gradient is calculated , parameter update, using the formula: Update the parameters.
[0013] Furthermore, the gradient of the loss function is calculated by the following formula: , represents the number of samples, Indicates that it is training samples, It represents the prediction function of the model. For a single sample, the gradient can be expressed as: ,in, Representation function Parameters The gradient is a vector that contains the function at point Partial derivatives in all parameter directions, Another way to write the gradient is It represents the derivative and the function At the point About parameters The partial derivative of .
[0014] Compared with the prior art, the present invention has the following advantages: The present invention provides a method for diagnosing hydraulic press system faults through an AI model. The method for diagnosing hydraulic press faults using an AI model performs online monitoring, and can perform maintenance and fault diagnosis in a timely manner when an abnormal situation occurs in the operating state of the hydraulic system. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] Figure 1 This is a flowchart of the framework of the AI model for diagnosing hydraulic press system faults in the present invention; Figure 2 It is a framework diagram of the data collection and data preprocessing process in the present invention; Figure 3 It is a framework diagram of the feature extraction process in the present invention; Figure 4 This is a flow chart of the model training phase in the present invention; Figure 5 This is a framework diagram of the model result processing flow in the present invention. DETAILED DESCRIPTION
[0016] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0017] Embodiment 1: In this embodiment, a method for diagnosing hydraulic press system failures through an AI model is shown in the accompanying drawings of this specification. Figures 1 to 5As shown, the hydraulic press system includes a data acquisition module, a data preprocessing module, a feature extraction module, and an AI model training module, and its diagnostic method includes the following steps: S1, data collection; S2, data preprocessing; S3, feature extraction stage; S4, model training stage; S5, model result processing and visualization.
[0018] Based on the above examples, please refer to Figure 2 In the figure, a flow chart of data collection and data preprocessing is shown. The embodiment process is as follows: In S1, at least one or more sensors are set on the hydraulic press to collect different types of data respectively. The sensors include at least temperature sensors, sound sensors, pressure sensors, and liquid sensors. In this embodiment, the temperature sensor is used to collect the temperature of the hydraulic press, such as the temperature when the hydraulic motor is overheated, the temperature when the oil temperature is too high, or the temperature when the hydraulic pump is overheated. The sound sensor is used to collect the noise when the hydraulic press is working. Traditionally, if the sound generated during the operation of the hydraulic press is too loud, it will become noise and affect the health of the operator. According to the national standard GB / T 26484-2011 "Noise Limits for Hydraulic Presses", the noise limits for hydraulic presses are clearly defined. For example, the sound power level of a hydraulic press during continuous idling should not exceed a certain number of decibels, and the specific value depends on the nominal force of the hydraulic press. For hydraulic presses with a nominal force not exceeding 1600kN, the sound power level should not exceed 90 decibels (A); for a nominal force between 1600kN and 4000kN, it should not exceed 94 decibels (A); for a nominal force between 4000kN and 10000kN, it should not exceed 102 decibels (A); for a nominal force between 10000kN and 20000kN, it should not exceed 110 decibels (A). In addition, the definition of noise usually refers to the sound exceeding 70 decibels as noise. Therefore, if the sound of the hydraulic press exceeds the limit value specified in the above national standards, or exceeds 70 decibels, it can be considered that it generates noise; the pressure sensor is used to collect the pressure during the working process of the hydraulic press; in S2, the data set is collected and the fault class label is determined. In this embodiment, the fault class label type is shown in the following table: Label Fault type 1 Oil leakage 2 No pressure on the system 3 Excessive noise 4 Hydraulic pump overheating 5 Oil temperature is too high 6 Servo pressure is too high 7 Hydraulic motor overheating The temperature and pressure data of the faulty tags are standardized. The original data is subtracted from the mean and then divided by the standard deviation to change the data into a distribution with a mean of 0 and a standard deviation of 1. Although this method does not necessarily result in a normal distribution, it makes full use of all data information in the dimensionless process. After standardization, the variables have not only the same mean but also the same standard deviation, eliminating the differences in the degree of variation of the variables, making the importance of the transformed variables in the cluster analysis equally important, facilitating the elimination of the dimension effect, and incorporating the data of each indicator into the comprehensive evaluation index. The standardized output range is from negative infinity to positive infinity: ; For data processing, the self-organizing map (SOM) is a distance-based algorithm, and then the type features in the input matrix X are One-Hot encoded and standardized. represents a data point in the original data, u represents the mean of the data set, Represents the standard deviation of the data set; the steps of data processing are: S21: import the data set, which includes a training set and a test set respectively; S22: view the feature columns and label columns of the training set and the test set respectively; S23: view the first five rows of data, clarify the project requirements, and determine the data to be predicted; S24: view the feature fields, clarify the meaning of each column of data, and the data format; S25: view the information of the column where the data is located, including the number, mean, median, maximum value, minimum value, and data distribution, the proportion of each interval, and draw a picture for intuitive display; S26: check whether there are outliers in the data and the normal distribution transformation, delete the outliers in the data, and transform the data to obey the normal distribution; In this embodiment, please continue to refer to the accompanying drawings of the specification. Figure 3 As shown in, it is a feature extraction process framework diagram, which includes standardizing the collected data and then performing feature engineering, including finding the mean to reflect the central trend, finding the variance to measure the degree of dispersion, and analyzing the trend pattern by time aggregation to determine the possible faults and fault types that may occur at a certain moment or within a certain period of time; determining the fault type by comparative analysis, combining domain knowledge and building a model, and its main steps include S31, checking missing values; S32, checking the data missing rate; S33, filling missing values; S34, feature process; S35, encoding the data; S36, checking feature importance. In S31 of this embodiment, when checking missing values, in order to check the missing conditions of the training set and the test set respectively, when processing missing values, the training set and the test set can be spliced, merged, and subsequently split; in S32 of this embodiment, checking data missing When the missing rate is found, statistics are taken for each column of data to check the missing rate, and the feature columns with a high missing rate are deleted or filled; in S33 of the present embodiment, when filling the missing values, the mean, maximum value, minimum value, value at the previous moment, and value at the next moment can be used to fill; in S34 of the present embodiment, when the feature process is performed, the time is aggregated by minutes, hours, and days, and the average is taken to form a new feature column; in S35 of the present embodiment, when encoding the data, the data before encoding is checked, the string data is converted into digital data, and the discontinuous numbers or texts are converted into continuous numerical variables; in S36 of the present embodiment, when checking the importance of the features, some evaluation algorithms are used to judge the importance of the features. In the case of insufficient CPU resources and other situations, the feature items with lower importance can be reduced to reduce complexity and improve efficiency.
[0019] Example 2, please refer to the attached figure in the specification Figure 4 As shown, it is a flow chart of the model training phase in this embodiment, and the implementation process is as follows: S4 includes: S41, weight initialization: initializing the weight to a very small random number to ensure that each neuron has a different weight at the beginning of training, can learn different features, and provide a random starting point for subsequent calculations, ensuring that the algorithm will not fall into a local optimal solution, and increasing the algorithm exploration ability; In the above-mentioned implementation case, before the training of the neural network begins, the weights in the network need to be initialized. The weights are the edges connecting neurons, and they determine how the input signal affects the activation of neurons. The weights are initialized to very small random numbers, rather than zero or fixed values, in order to break the symmetry and prevent all neurons from learning the same features. This ensures that each neuron in this embodiment has a different weight at the beginning of training, so that different features can be learned; if all weights are initialized to the same value, then during the training process, the neurons may learn the same features, resulting in the network being unable to effectively learn complex patterns. The beneficial effect of random initialization is that the exploration ability of the network can be increased, helping it to jump out of the local optimal solution and find the global optimal solution. In addition, random initialization can increase the exploration ability of the algorithm because it provides a random starting point for the network, allowing the network to explore different paths during the training process, rather than falling into a fixed pattern at the beginning.
[0020] Embodiment 3, based on the above embodiment S4, further includes S42, selecting an input sample: randomly selecting an input sample " ", the input sample " "In subsequent steps, it is compared and calculated with the nodes in the competition layer; in this embodiment, when training a neural network, input data is required to update the weights. These data are usually divided into multiple batches, called "batches" or "mini-batches". In each training step, an input sample is randomly selected" ", which ensures that the network will not overfit to a specific sample, and also increases the randomness of training, which helps the network learn more general features; in a neural network, a competitive layer is a special network structure in which neurons compete with each other to decide which neurons should be activated. When selecting input samples" "After that, the sample is fed into the network and compared and calculated with the nodes in the competitive layer. This process involves calculating the similarity or difference between the input sample and the nodes in the competitive layer, and then updating the network's weights based on these calculations to improve the network's accuracy for the input sample." ” classification or prediction ability.
[0021] Embodiment 4, based on the above embodiment S4, further includes step S43, determining the winning node, and determining the winning node includes the following two steps: (1) traversing each node in the competition layer, calculating the input sample " ” and the similarity between the nodes, the similarity calculation method is determined according to the specific problem and data type, and the calculation method includes the Euclidean distance method or the cosine similarity method; (2) the node with the smallest distance is selected as the winning node, called the "winner node" or "BMU (best matching unit)"; S44, determine the winning neighborhood and the update amplitude, and determine the nodes to be included in the winning neighborhood according to the neighborhood radius "σ (sigma)"; calculate the update amplitude of each of these nodes through the "neighborhood function", the closer to the winning node, the larger the update amplitude, and the farther away from the winning node, the smaller the update amplitude.
[0022] In embodiment 5, the S4 further includes S45, updating the weight "Weight" of the node in the winning neighborhood, using the formula: To update, " is a function that depends on the input , Output and time step The learning rate or adjustment factor of " represents the weight at time step S;" ” means at time step The updated weights are ” is another time-step dependent The factor of gradually decreases with the increase of the number of iterations to control the amplitude of the update; ” is the target value or desired output value.
[0023] Based on the above embodiments, a simple embodiment is listed, assuming that is the initial weight, is the learning rate, is a constant factor, is the target value, calculated according to the parameter values of the above embodiment ,
[0024]
[0025]
[0026]
[0027] The results of the above implementation can be: Initial weight
[0028] Learning Rate
[0029] factor
[0030] Target value
[0031] Updated weights The above values can be adjusted according to different application scenarios in the embodiment.
[0032] Embodiment 6, step S4 also includes S46 iteration. After completing one round of iteration, the number of iterations is increased by one, and then the second step is returned to continue to randomly select an input sample for the next round of iteration until the set number of iterations is met. In each round of iteration, the winning node and the nodes in its neighborhood are updated according to the input sample, and the weights are gradually adjusted to better adapt to the distribution of the input data. By continuously repeating these steps, the algorithm can gradually learn the characteristics and patterns of the input data and realize the clustering or classification task of the data. In this embodiment, we can set a simple linear regression model, and its weight update formula is: ,in, is the iteration time step The weight vector when is the iteration time step The updated weight vector is is the learning rate, is the loss function About weight The gradient of is the input sample, is the corresponding target value, and the embodiment data is shown in the following process: 1. Initialization: Weight
[0033] Learning Rate
[0034] Iterations
[0035] 2. Each round of iteration: Randomly select an input sample and target value , Calculate the predicted value: ; Calculate the loss: ; Compute the gradient: ; Update weights: ; 3. Iteration process: Iteration steps: Initialization: Setup , that is, starting from the first iteration, Iteration process: Repeat the above steps. Termination condition: Continue iterating until ,in Represents the total number of preset iterations. The example calculation is as follows: Assume that in the first iteration, Input Sample , target value , initial weight , Let’s calculate the predicted value: ; Calculate the loss: ; Compute the gradient: ; Update weights: ; In this way, each round of iteration will update the weights according to the input samples until the set number of iterations is reached. This process can help the model learn the characteristics and patterns of the input data and achieve data classification or clustering tasks.
[0036] Example 7, please refer to the attached drawings of the specification Figure 5 As shown in , it represents the framework diagram of the model result processing flow in this embodiment, and the implementation process is as follows: S5 includes S51. After the model training is completed, when there is subsequent real-time data input, the model will automatically play a role in real-time diagnosis; S52. The model will analyze the real-time data according to the patterns and features learned during the training process to determine whether there is a fault and the possible fault type; S53. The model will call the knowledge base, which stores a large amount of professional knowledge, historical fault cases and related feature information; S54. By comparing the fault category diagnosed by the model with the content in the knowledge base, a comprehensive assessment of the credibility of the fault category is made; S54 also includes S541. If the model diagnosis result is highly consistent with the typical fault feature in the knowledge base, then the credibility of the fault category is high; S542. If the difference between the two is large, the credibility is low; S543. The model will also give the probability of fault occurrence based on the similarity between the real-time data and the existing fault data and the statistical characteristic factors of the data; S544. The probability can provide an important reference for decision makers so that corresponding prevention and treatment measures can be taken in time.
[0037] In this implementation, data of an embodiment of hydraulic press operation can be generated according to the above fault training steps: 1. Training data: Set up normal operation data: 1000 samples, each sample includes temperature, pressure and vibration readings; Fault data: 200 samples, each sample includes temperature, pressure, vibration readings and fault class labels; 2. Real-time data The current operating data of the hydraulic press: pressure = 150bar, flow = 50L / min, temperature = 75°C; 3. Knowledge Base Content Fault type A: pressure>145bar, flow<45L / min, temperature>80°C; Fault type B: pressure <130bar, flow rate >55L / min, temperature <70°C; 4. Model diagnosis According to the model analysis of real-time data, the fault type A is diagnosed; The model diagnosis results are highly consistent with the characteristics of fault type A in the knowledge base, so the credibility is high and the probability of fault occurrence is 90%. Based on the 90% probability of occurrence, the decision maker decides to check the hydraulic press immediately and take preventive measures. Through this process, the fault diagnosis of the hydraulic press can analyze the operating data of the equipment in real time, identify potential dangers, identify potential failures, and provide decision support to ensure the stable operation of the equipment.
[0038] Example 8: In each round of iteration, the model parameters are updated according to the input samples to better adapt to the data distribution. In the k-round iteration, the update of the parameter θ is expressed as: ,in, Indicated in The parameters of the model after rounds of iterations, Indicated in The parameters of the model after rounds of iterations, is the learning rate, Represents the loss function In the current parameters The gradient at .
[0039] Example process: Step 1: Initialize parameters and set the initial parameters for the hydraulic press operation model ; Step 2: Select a learning rate and determine an appropriate learning rate , used to control the step size of parameter update; Step 3: Define the loss function and choose a loss function , used to evaluate the difference between the model prediction and the actual value; Step 4: Calculate the gradient. In the kth iteration, calculate the loss function About parameters Gradient ; Step 5: Update parameters using formula Update parameters; Step 6: Repeat the iterations, repeating steps 4 and 5 until the number of iterations is met or the model converges.
[0040] The example data are as follows: Setting initial parameters , Setting the learning rate , Setting the loss function ,in is the actual value, is the predicted value, Set the data samples to 10, each sample contains the input variables of the hydraulic press and the target variable , assuming the prediction model , The gradient is calculated as follows: In the first iteration, the gradient is calculated , parameter update, using the formula: Update parameters; Example calculation: Assume that in the first round of iteration, the following data is set: Input variables ; Target variable
[0041] 1. Prediction model: , 2. Loss function: , 3. Gradient calculation: , 4. Parameter update: .
[0042] In Example 9, the gradient of the loss function is calculated by the following formula: ,in, represents the number of samples, Indicates that it is training samples, represents the prediction function of the model, The embodiment process in the above embodiment 9 is as follows: Step 1: Define the model prediction function; Define the prediction function of the model , which is based on the input , target value and current parameters To predict the output; Step 2: Initialize parameters; Setting initial parameters ; Step 3: Calculate the gradient; For each training sample i, calculate the prediction function About parameters Gradient ; Step 4: Calculate the average gradient; Calculate the average value of the gradient of all training samples and get ; Step 5: Update parameters; Update the parameters using gradient descent: ; Step 6: Repeat the iteration; Repeat steps 3 to 5 above until the number of iterations is met or the model converges.
[0043] Example data: Suppose we have 3 training samples, input and target value As shown in the following table: , Model prediction function: set up , Initial parameters: set up , Gradient calculation: For each sample ,calculate The gradient is: , Parameter update: Compute the average gradient: , Update parameters: , Example calculation: Setting the learning rate , 1. Gradient calculation For sample 1: , For sample 2: , For sample 3: , 2. Average gradient, 3. Parameter update ,
[0044] Through this process, the model parameters are updated according to the input samples to better adapt to the data distribution. This process is repeated until the model converges, which can improve the model's prediction accuracy for the hydraulic press operation data.
[0045] Continuing with the above example, for a single sample, the gradient can be expressed as: , in this embodiment, Represents the model prediction function About parameters The gradient of the function at point Partial derivatives in all parameter directions, Another way to write the gradient is It represents the derivative and the function At the point About parameters The partial derivative of training samples.
[0046] The implementation process in this embodiment is as follows: Step 1: Define the model prediction function: Define the prediction function of the model , which is based on the input , target value and current parameters To predict the output; Step 2: Initialization parameters: Set initial parameters ; Step 3: Calculate the gradient of a single sample: training book, calculate the prediction function About parameters The gradient of Step 4: Update parameters: Update parameters using gradient descent method: .
[0047] The example data are as follows: Suppose we have 1 training sample, input and target value As shown in the following table: , Model prediction function: Assumptions , Initial parameters: Assumptions , Gradient calculation: For sample 1, calculate The gradient is: , Parameter update: Compute the gradient: , Update parameters: , Example calculation: Assuming the learning rate , gradient calculation: , Parameter update: , Through this process, the model parameters are updated based on a single input sample to better fit the data distribution. Repeating this process until the model converges can improve the accuracy of the model's prediction of the data.
[0048] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.
Claims
1. A method for diagnosing hydraulic press system faults through an AI model, wherein the hydraulic press system includes a data acquisition module, a data preprocessing module, a feature extraction module, an AI model training module, and a model processing result module. It is characterized in that The method includes the following steps: S1, data collection; S2, data preprocessing; S3, feature extraction stage; S4, model training stage; S5, model result processing and visualization.
2. A method for diagnosing hydraulic press system faults through an AI model according to claim 1, Features: In the S1, at least one type or more of sensors are arranged on the hydraulic press to collect data of different categories respectively; in the S2, data sets are collected, class labels are determined, and the collected label data are standardized. The data is changed into a distribution with a mean of 0 and a standard deviation of 1 by subtracting the mean from the original data and then dividing it by the standard deviation. The standardized output range is from negative infinity to positive infinity: ;in, represents a data point in the original data. represents the mean of the data set, Represents the standard deviation of the data set; the steps of data processing are: S21: import the data set, which includes a training set and a test set respectively; S22: view the feature columns and label columns of the training set and the test set respectively; S23: view the first five rows of data, clarify the project requirements, and determine the data to be predicted; S24: view the feature fields, clarify the meaning of each column of data, and the format of the data; S25: view the information of the column where the data is located; S26: check whether there are outliers in the data and the normal distribution transformation; The S3 includes standardizing the collected data and then performing feature engineering, including finding the mean to reflect the central trend, finding the variance to measure the degree of dispersion, and analyzing the trend pattern by time aggregation to determine the possible faults and fault types that may occur at a certain moment or within a certain period of time; determining the fault type by comparative analysis, combining domain knowledge and building a model, and its main steps include S31, checking missing values; S32, checking the data missing rate; S33, filling missing values; S34, feature process; S35, encoding the data; S36, checking feature importance.
3. A method for diagnosing hydraulic press system faults through an AI model according to claim 1, It is characterized in that The S4 includes: S41, weight initialization: initializing the weight to a very small random number to ensure that the processing units in the network have different weights at the beginning of training, can learn different features, provide a random starting point for subsequent calculations, ensure that the algorithm will not fall into a local optimal solution, and increase the algorithm's exploration ability; S42, selecting input samples: randomly selecting a data point in the original data As input sample" ", the input sample" ” is compared and calculated with the nodes in the competition layer in subsequent steps.
4. According to the method for diagnosing hydraulic press system faults by using an AI model as described in claim 3, It is characterized in that The S4 also includes S43, determining the winning node: including the following two steps: (1) traversing each node in the competition layer and calculating the input sample" ” and the similarity between the nodes, the similarity calculation method is determined according to the specific problem and data type, and the calculation method includes the Euclidean distance method or the cosine similarity method; (2) select the node with the smallest distance as the winning node; S44, determine the winning neighborhood and update amplitude: determine the nodes to be included in the winning neighborhood according to the neighborhood radius; calculate the update amplitude of each of these nodes through the "neighborhood function", the closer to the winning node, the larger the update amplitude, and the farther away from the winning node, the smaller the update amplitude.
5. A method for diagnosing hydraulic press system faults through an AI model according to claim 4, It is characterized in that The S4 also includes S45, updating the weights of the nodes in the winning neighborhood, using the following formula: Update, where " " is a function that depends on the input , Output and time step The learning rate or adjustment factor of ; " represents the weight at time step s;" ” means at time step The updated weights; " is another factor that depends on the time step s and gradually decreases with the number of iterations to control the magnitude of the update;" ” is the target value or desired output value.
6. A method for diagnosing hydraulic press system faults through an AI model according to claim 5, It is characterized in that The S4 also includes S46 iteration. After completing one round of iteration, the number of iterations is increased by one, and then the process returns to S42 to continue to randomly select an input sample "X i "The next round of iterations is carried out until the set number of iterations is met. In each round of iteration, the winning node and the nodes in its neighborhood will be updated according to the input sample, and the weights will be gradually adjusted to better adapt to the input sample"X i "By repeating these steps, the algorithm can gradually learn the input sample "X i "The characteristics and patterns of the input samples are realized" "Clustering or classification task; set up a simple linear regression model, and its weight update formula is: ,in, is the weight vector at iteration time step s, is the iteration time step The updated weight vector is is the learning rate, is the loss function About weight The gradient of is the input sample, is the corresponding target value; The calculation process in each iteration includes the following: randomly select an input sample and target value ; Calculate the predicted value: ; Calculate the loss: ; Compute the gradient: ; Update weights: ; Iteration steps: Initialization: Setup , that is, starting from the first iteration, Iteration process: Repeat the above steps. Termination condition: Continue iterating until ,in Represents the total number of preset iterations.
7. A method for diagnosing hydraulic press system faults through an AI model according to claim 1, It is characterized in that The S5 includes S51: after the model training is completed, when there is subsequent real-time data input, the model will automatically play a role in real-time diagnosis; S52: the model will analyze the real-time data according to the patterns and features learned during the training process to determine whether there is a fault and the possible type of fault; S53: the model will call the knowledge base, which stores a large amount of professional knowledge, historical fault cases and related feature information; S54: by comparing the fault category diagnosed by the model with the content in the knowledge base, a comprehensive assessment of the credibility of the fault category is made.
8. A method for diagnosing hydraulic press system faults through an AI model according to claim 7, It is characterized in that The S54 also includes S541. If the model diagnosis result is highly consistent with the typical fault characteristics in the knowledge base, then the credibility of the fault category is high; S542. If the two are very different, the credibility is low; S543. The model will also give the probability of fault occurrence based on the similarity between the real-time data and the existing fault data and the statistical characteristic factors of the data; S544. The probability can provide an important reference for decision makers so that they can take corresponding prevention and treatment measures in time.
9. A method for diagnosing hydraulic press system faults through an AI model according to claim 6, Features: In each round of iteration, the model parameters are updated according to the input samples to better adapt to the data distribution. In the k-th round of iteration, the update of the parameter θ is expressed as: ,in, Indicated in The parameters of the model after rounds of iterations, Indicated in The parameters of the model after rounds of iterations, is the learning rate, Represents the loss function In the current parameters The gradient at; Set the loss function ,in is the actual model value, is the prediction model value; set the prediction model value , the gradient is calculated as: In the first round of iteration, the gradient is calculated , parameter update, using the formula: Update the parameters.
10. A method for diagnosing hydraulic press system faults through an AI model according to claim 9, Features: The gradient of the loss function is calculated by the following formula: , represents the number of samples, Indicates that it is training samples, It represents the prediction function of the model. For a single sample, the gradient can be expressed as: ,in, Representation function Parameters The gradient is a vector that contains the function at point Partial derivatives in all parameter directions, Another way to write the gradient is It represents the derivative and the function At the point About parameters The partial derivative of .
Citation Information
Patent Citations
Photovoltaic array fault diagnosis and early warning method
CN106961249A
A fault diagnosis method for refrigeration system
CN109902741A
Remote fault diagnosis method and device for hydraulic machine and storage medium
CN111078966A
SOM network clustering electromechanical equipment bearing fault analysis method based on transfer learning and manifold distance
CN113496061A
Hydraulic machine fault diagnosis method based on ES-MLSTM
CN113530921A
Cited By
Multi-axis leveling electro-hydraulic system fault reconstruction method based on iterative learning and sliding mode observation
CN121364621A
