Method for predicting working state of hydraulic pump of die forging press based on convolutional neural network

Abstract

The invention provides a method for predicting the working state of a hydraulic pump of a die forging press based on a convolutional neural network. The method comprises the following steps: S1, monitoring the working state of the hydraulic pump in real time according to an equipment mechanism; s1.1, performing structured processing on the original data of the hydraulic pump sensor; s1.2, performing alarm rule judgment on the abnormal data; s1.3, carrying out filtering processing on the normal data; s2, performing prediction according to the working state of the hydraulic pump of the CNN; S2.1, performing a one-dimensional time sequence on data recorded by each sensor; s2.2, calculating an autocorrelation coefficient (ACF) to obtain data at the current moment; s2.3, calculating the maximum lag order of the current data; s2.4, taking the maximum lag order as the size of a convolution sliding window; s2.5, generating a training data label according to the data and the sliding window value; and S2.6, constructing a CNN model according to the training data. Through efficient acquisition, storage and write-back of sensor data, the purpose of real-time monitoring of a sensor measuring point value is achieved.

Description

Technical Field

[0001] The invention relates to the field of monitoring technology, and in particular to a method for predicting the working state of a hydraulic pump of a die forging press based on a convolutional neural network. Background Art

[0002] As one of the common processing methods, forging production plays an important role in the manufacturing industry. Forging can produce blanks or directly produce finished products. Parts manufactured by forging equipment have the advantages of high productivity, high performance and precision, and play a very critical and important role in the national economy, social development and scientific development.

[0003] As an indispensable and important equipment in the machinery manufacturing industry, the working status of large die forging presses will directly affect the quality of forgings and the safe operation of the production line. As the core component of the die forging press, the hydraulic pump plays a key role in the healthy operation of the equipment. Therefore, real-time monitoring and future prediction of the working status of the hydraulic pump are of great significance to ensure the safe and reliable operation of the equipment.

[0004] In current actual production, the real-time monitoring of the working status of the hydraulic pump of the die forging press is mainly carried out by installing sensors on the hydraulic pump and cooperating with on-site manual work based on the key measurement point values; the prediction of future status is mainly based on expert experience or by establishing a machine learning model based on sensor data. The above methods have the following problems:

[0005] (1) Manual monitoring has the characteristics of high monitoring cost and delayed results, which is not conducive to enterprises to increase costs and efficiency;

[0006] (2) Expert experience is highly subjective and the prediction results are unstable;

[0007] (3) In order for machine learning models to achieve good prediction results, they need to have sufficient business experience to extract features from the data. Summary of the invention

[0008] The technical problem solved by the present invention is to provide a method for predicting the working state of a hydraulic pump of a die forging press based on a convolutional neural network, which achieves the purpose of real-time monitoring of sensor measuring point values ​​through efficient collection, storage and writing back of sensor data.

[0009] The technical solution adopted by the present invention to solve its technical problem is:

[0010] The method for predicting the working state of the hydraulic pump of a die forging press based on a convolutional neural network comprises the following steps:

[0011] S1. Real-time monitoring of the hydraulic pump working status according to the equipment mechanism;

[0012] S1.1, structure the raw data of the hydraulic pump sensor and write it into the PostgreSQL database;

[0013] S1.2. Extract the original data and make alarm rule judgment on abnormal data;

[0014] S1.3. Normal data will be filtered and displayed on the front end;

[0015] S2. Predict the working status of the hydraulic pump based on CNN:

[0016] S2.1. According to the operation mechanism of the hydraulic pump, the data recorded by each sensor is processed into a one-dimensional time series. A CNN model is constructed for each sensor using one-dimensional convolution to process the raw data of the sensor into a one-dimensional vector.

[0017] S2.2. For the data of the current batch, calculate its autocorrelation coefficient (ACF) to obtain the data at the current moment;

[0018] S2.3, calculate the maximum lag order of the current data;

[0019] S2.4, the process of determining the size of the sliding window by taking the maximum lag order as the size of the convolution sliding window;

[0020] S2.5. Generate training data labels based on the data and sliding window values. The training data is the data in a sliding window, and the label is the measurement point value at the next moment.

[0021] S2.6. Build a CNN model based on the training data.

[0022] In step S1.1, each data in the raw data consists of a timestamp and a json dictionary, and the json dictionary includes a key and a value, the key is the sensor measurement point name, and the value is the measurement point value corresponding to the current moment.

[0023] In step S1.2, the abnormal data is written into the abnormal result table of the database and fed back to the front-end page for the crew to inspect and process.

[0024] In step 2.2, the calculation formula of the ACF is as follows:

[0025]

[0026] In formula (1), x t is the measured point value at time t; x t-k is the measured point value at time tk.

[0027] In step 2.3, the maximum lag order of the current data is calculated as follows:

[0028] 1) Calculate the maximum points in the current ACF coefficient and record the index position of each maximum point in the entire ACF array;

[0029] 2) Traverse the ACF coefficients corresponding to the index position and determine the index position of the global maximum ACF value, which is the maximum lag order of the current data.

[0030] In step 2.5, the process of generating training data labels is as follows:

[0031]

[0032] In formula (2), X is a sequence of 100 time steps; X train Each subsequence in is a data sample of a window length (window), corresponding to Y label The training labels in .

[0033] The beneficial effects of the present invention are:

[0034] 1. Through the efficient collection, storage and writing back of sensor data, the purpose of real-time monitoring of sensor measurement point values ​​is achieved, and the key measurement point data of hydraulic pump sensors can be monitored in seconds; and built-in alarm rules are used to display abnormal real-time data, greatly reducing the cost of manual monitoring, and can quickly feedback sensor abnormal values ​​to solve the lag of abnormal results of manual monitoring.

[0035] 2. Establish a CNN network model through the original data of the sensor, and combine it with statistical methods to automatically set the sliding window length to predict the future state of the sensor measurement point value; the model can be built using the original data of the sensor, and there is no need for rich knowledge of the equipment mechanism to extract data features; combined with the autocorrelation coefficient (ACF), the sliding window size of the training data is automatically calculated to avoid the subjective influence of human experience; for time series data, the CNN model based on the sliding window can better preserve and extract the characteristics of time series data in a continuous time period, and the prediction results have higher robustness and smaller prediction errors than traditional machine learning algorithms. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a work flow chart of the present invention;

[0037] Figure 2 Flowchart for real-time monitoring of sensor data;

[0038] Figure 3 The original data is serialized into JSON format.

[0039] Figure 4 It is a table showing the abnormal results of rule judgment;

[0040] Figure 5 It is a display diagram of the original data of the terminal;

[0041] Figure 6 This is a graph showing the data after the filter denoises it;

[0042] Figure 7 This is the flow chart of CNN hydraulic pump working status prediction. DETAILED DESCRIPTION

[0043] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0044] Example 1

[0045] like Figure 1 As shown, the working status (including normal state and fault state) of the hydraulic pump group of the die forging press is monitored in real time, and the current equipment failure can be presented and the equipment failure in the next 3 minutes can be predicted.

[0046] First, the raw data collected by the sensor will pass through the mechanism-based working state monitoring system and CNN model at the same time. The working state monitoring system first returns the real-time fault information of the equipment, and then returns the fault warning information after 3 minutes based on the future prediction value of the sensor measurement point by the CNN model.

[0047] The method for predicting the working state of the hydraulic pump of a die forging press based on a convolutional neural network comprises the following steps:

[0048] S1, such as Figure 2 As shown, the working status of the hydraulic pump is monitored in real time according to the equipment mechanism;

[0049] S1.1. Structural processing of the raw data of the hydraulic pump sensor and writing it into the PostgreSQL database; due to the large amount of data collected by the equipment sensor, in order to ensure the reading and writing speed and storage space, we serialize all the measuring points into json. The specific process is as follows: Figure 3 As shown in the figure, each piece of data consists of a timestamp and a json dictionary. The key in the json dictionary is the sensor measurement point name, and the value is the measurement point value corresponding to the current moment. This method can achieve better reading and writing speed and smaller storage space.

[0050] S1.2. Extract the original data and judge the alarm rules for abnormal data; if the data of a certain measuring point is abnormal, it will be written into the abnormal result table of the database and fed back to the front-end page for the crew to inspect and handle. The abnormal result table is as follows: Figure 4As shown in Table 1, each piece of data contains a timestamp, a pump group effective operation time, and an alarm information JSON dictionary. We use the plpython language to integrate 6 abnormal alarm rules in the postgresql database, as shown in Table 1;

[0051]

[0052] Table 1 is the abnormal alarm rule table

[0053] S1.3, when the die forging press is working, due to the complex on-site environment, the final data collected will have certain noise interference; therefore, the normal data will be filtered to facilitate the display of normal data on the front-end data; Figures 5 to 6 As shown, the data is denoised using a Butterworth filter;

[0054] S2, such as Figure 7 As shown in the figure, a single-channel CNN model is used to build a CNN model for each sensor of the hydraulic pump, and the data trend of the sensor measurement point is predicted. The working status of the hydraulic pump is predicted based on CNN:

[0055] S2.1. According to the operation mechanism of the hydraulic pump, the data recorded by each sensor is processed into a one-dimensional time series. A CNN model is constructed for each sensor using one-dimensional convolution to process the raw data of the sensor into a one-dimensional vector.

[0056] S2.2. For the data of the current batch, calculate its autocorrelation coefficient (ACF) to obtain the data at the current moment;

[0057] S2.3, calculate the maximum lag order of the current data;

[0058] S2.4. The maximum lag order is used as the size of the convolution sliding window (for example, if the sliding window size is 100, it means that the current data is related to the data of the previous 100 time steps). The process of determining the sliding window size is also the feature selection process of the data. Using the ACF coefficient to automatically calculate the sliding window size can effectively avoid the interference of manual experience.

[0059] S2.5. Generate training data labels based on the data and sliding window values. The training data is the data in a sliding window, and the label is the measurement point value at the next moment.

[0060] S2.6. Construct a CNN model based on the training data. The CNN model architecture is shown in Table 2.

[0061]

[0062] Table 2 shows the CNN model architecture

[0063] In step S1.1, each data in the raw data consists of a timestamp and a json dictionary, and the json dictionary includes a key and a value, the key is the sensor measurement point name, and the value is the measurement point value corresponding to the current moment.

[0064] In step S1.2, the abnormal data is written into the abnormal result table of the database and fed back to the front-end page for the crew to inspect and process.

[0065] In step 2.2, if a set of measured point values ​​in a certain period of time is described as a sequence X t , the calculation formula of the ACF is as follows:

[0066]

[0067] In formula (1), x t is the measured point value at time t; x t - k is the measured point value at time tk.

[0068] In step 2.3, the maximum lag order of the current data is calculated as follows:

[0069] 3) Calculate the local maximum points in the current ACF coefficient and record the index position of each maximum point in the entire ACF array;

[0070] 4) Traverse the ACF coefficients corresponding to the index position and determine the index position of the global maximum ACF value, which is the maximum lag order of the current data.

[0071] In step 2.5, the process of generating training data labels is as follows:

[0072]

[0073] In formula (2), X is a sequence of 100 time steps; X train Each subsequence in is a data sample of a window length (window), corresponding to Y label The training labels in .

[0074] By efficiently collecting, storing and writing back sensor data, the purpose of real-time monitoring of sensor measurement point values ​​is achieved, and the key measurement point data of hydraulic pump sensors can be monitored in seconds. It also has built-in alarm rules to display alarms for abnormal real-time data, greatly reducing the cost of manual monitoring, and can quickly feedback sensor abnormal values ​​to solve the lag of abnormal results in manual monitoring.

[0075] The CNN network model is established through the raw data of the sensor, and combined with statistical methods, the sliding window length is automatically set to predict the future state of the sensor measurement point value; the model can be built using the raw data of the sensor, and there is no need for rich knowledge of the equipment mechanism to extract data features; combined with the autocorrelation coefficient (ACF), the sliding window size of the training data is automatically calculated to avoid the subjective influence of human experience; for time series data, the CNN model based on the sliding window can better preserve and extract the characteristics of time series data in a continuous time period, and the prediction results have higher robustness and smaller prediction errors than traditional machine learning algorithms.

[0076] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various changes and variations. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting the working state of a hydraulic pump of a die forging press based on a convolutional neural network, characterized in that: The following steps are involved: S1. Real-time monitoring of the hydraulic pump working status according to the equipment mechanism; S1.1, structure the raw data of the hydraulic pump sensor and write it into the PostgreSQL database; S1.

2. Extract the original data and make alarm rule judgment on abnormal data; S1.

3. Normal data will be filtered and displayed on the front end; S2. Predict the working status of the hydraulic pump based on CNN: S2.

1. According to the operation mechanism of the hydraulic pump, the data recorded by each sensor is processed into a one-dimensional time series. A CNN model is constructed for each sensor using one-dimensional convolution to process the raw data of the sensor into a one-dimensional vector. S2.

2. For the data of the current batch, calculate its autocorrelation coefficient (ACF) to obtain the data at the current moment; S2.3, calculate the maximum lag order of the current data; S2.4, the process of determining the size of the sliding window by taking the maximum lag order as the size of the convolution sliding window; S2.

5. Generate training data labels based on the data and sliding window values. The training data is the data in a sliding window, and the label is the measurement point value at the next moment. S2.

6. Build a CNN model based on the training data.

2. The method for predicting the working state of a hydraulic pump of a die forging press based on a convolutional neural network according to claim 1, characterized in that: In step S1.1, each data in the raw data consists of a timestamp and a json dictionary, and the json dictionary includes a key and a value, the key is the sensor measurement point name, and the value is the measurement point value corresponding to the current moment.

3. The method for predicting the working state of a hydraulic pump of a die forging press based on a convolutional neural network according to claim 1, characterized in that: In step S1.2, the abnormal data is written into the abnormal result table of the database and fed back to the front-end page for the crew to inspect and process.

4. The method for predicting the working state of a die forging press hydraulic pump based on a convolutional neural network according to claim 1, characterized in that: In step 2.2, the calculation formula of the ACF is as follows: In formula (1), x t is the measured point value at time t; x t - k is the measured point value at time tk.

5. The method for predicting the working state of a die forging press hydraulic pump based on a convolutional neural network according to claim 1, characterized in that: In step 2.3, the maximum lag order of the current data is calculated as follows: 1) Calculate the maximum points in the current ACF coefficient and record the index position of each maximum point in the entire ACF array; 2) Traverse the ACF coefficients corresponding to the index position and determine the index position of the global maximum ACF value, which is the maximum lag order of the current data.

6. The method for predicting the working state of a hydraulic pump of a die forging press based on a convolutional neural network according to claim 1, characterized in that: In step 2.5, the process of generating training data labels is as follows: In formula (2), X is a sequence of 100 time steps; X train Each subsequence in is a data sample of a window length (window), corresponding to Y label The training labels in .

Images