Biomass granulator fault prediction method based on deep learning

Abstract

The invention relates to the technical field of intelligent manufacturing and fault prediction, and discloses a biomass granulator fault prediction method based on deep learning. According to the method, operation data are collected in real time through multiple sensors installed on the biomass granulator, and data preprocessing is carried out. And then, constructing a GDN model by using a graph neural network, and inputting the preprocessed data into the model for fault prediction. The GDN model comprises sensor embedding, graph structure learning, prediction based on graph attention, graph deviation scoring and other parts. And according to the prediction result, corresponding early warning strategies are adopted, including sending of early warning signals of different levels, maintenance reminding, maintenance notification, shutdown notification and the like. According to the method, the relevance between the sensors is fully considered, the prediction accuracy and reliability are improved, faults are found and processed in time, and damage to production efficiency and equipment caused by the faults is effectively avoided.

Description

Technical Field

[0001] The present invention relates to the field of intelligent manufacturing and fault prediction technology, and in particular to a biomass pellet machine fault prediction method based on deep learning. Background Art

[0002] During the production process of biomass pellet machines, various faults such as wear, blockage, overheating, etc. often occur due to long-term operation and the influence of material characteristics. These faults not only affect production efficiency and product quality, but may also cause serious damage to the equipment. Traditional fault detection methods mostly rely on manual inspections and empirical judgments, and there are problems such as untimely detection and low accuracy. Therefore, how to achieve accurate prediction and early warning of biomass pellet machine failures has become a problem that needs to be solved urgently. At present, although some fault prediction methods based on machine learning have been proposed, most of these methods only consider the data of a single sensor and ignore the correlation between sensors, resulting in poor prediction results. Summary of the invention

[0003] In view of the above-mentioned technical deficiencies, the purpose of the present invention is to provide a biomass pellet machine fault prediction method based on deep learning to solve the problem of low fault prediction accuracy in the prior art.

[0004] In order to solve the above technical problems, the present invention adopts the following technical solutions: In a first aspect, the present invention provides a biomass pellet machine fault prediction method based on deep learning.

[0005] Preferably, in a possible implementation manner of the first aspect, the method includes: Step S100: obtaining the operation data of the biomass pellet machine in real time through the sensor installed on the biomass pellet machine; Step S200: preprocessing the real-time biomass pellet machine operation data; Step S300: constructing a GDN model using a graph neural network, and using historical biomass pellet machine operation data and historical fault data for training; Step S400: input the pre-processed real-time biomass pellet machine operation data into the trained GDN model, and output the fault prediction result; Step S500: adopt corresponding early warning strategies according to the fault prediction results.

[0006] Preferably, in a possible implementation manner of the first aspect, the sensor includes a temperature sensor, a pressure sensor, a vibration sensor, a flow sensor and a rotation speed sensor.

[0007] Preferably, in a possible implementation manner of the first aspect, the biomass pellet machine operation data includes temperature data, pressure data, vibration frequency data, material flow data and motor speed data.

[0008] Preferably, in a possible implementation manner of the first aspect, the data preprocessing includes data cleaning, missing value filling, outlier processing and data normalization.

[0009] Preferably, in a possible implementation manner of the first aspect, the GDN model includes: Sensor embedding, using embedding vectors to capture the behavioral characteristics of each sensor on the biomass pellet machine; Graph structure learning, using a directed graph to represent sensors and their dependencies, with nodes being sensors and edges being dependencies, and calculating the similarity between node embedding vectors and nodes; Attention-based graph model, introduces a graph attention-based feature extractor, uses a sliding window to capture the relationship between sensors, fuses the information of a node with its neighboring nodes, and calculates the aggregate representation; Sensor prediction, obtains the representation of all nodes through the feature extractor, multiplies it with the time series embedding, inputs it into the fully connected layer, and outputs the value of the sensor at the predicted time; Figure 1. Deviation score. According to the deviation between the sensor predicted output and the sensor observed data, the sensor anomaly value is calculated. The maximum anomaly value of all sensors is taken as the anomaly score. If it exceeds the anomaly threshold, it is marked as abnormal. The fault condition of the biomass pellet machine is judged according to the anomaly score, the sensor to which it belongs, and related sensors.

[0010] Preferably, in a possible implementation manner of the first aspect, the historical biomass pellet machine operation data includes historical temperature data, historical pressure data, historical vibration frequency data, historical material flow data, historical motor speed data and a timestamp.

[0011] Preferably, in a possible implementation manner of the first aspect, step S400 specifically includes: Input the preprocessed real-time biomass pellet machine operation data into the trained GDN model; The GDN model outputs an anomaly score. When the anomaly score exceeds the anomaly threshold, it is marked as an abnormal state. According to the abnormal score, the belonging sensor and the related sensor are judged. The belonging sensor is the sensor directly affected when the biomass pellet machine fails, and the related sensor is the sensor indirectly affected when the biomass pellet machine fails. According to the abnormal score, the information of the corresponding sensor and related sensors, the fault condition of the biomass pellet machine is judged, including the fault type, fault occurrence time and fault severity.

[0012] Preferably, in a possible implementation manner of the first aspect, the fault severity is divided into three levels: slight, medium and severe according to the deviation between the predicted output of the biomass pellet machine sensor and the sensor observation data.

[0013] Preferably, in a possible implementation of the first aspect, the early warning strategy includes: Send different levels of early warning signals according to the fault type and fault severity; For minor faults, the system sends maintenance reminders to remind you to regularly check and maintain the biomass pellet machine; For medium faults, the system sends a maintenance notification and takes maintenance measures on the biomass pellet machine immediately; For serious faults, the system sends a shutdown notification and immediately shuts down the biomass pellet machine for inspection.

[0014] The beneficial effects of the present invention are: by collecting operating data in real time through a variety of sensors installed on the biomass pellet machine, and using the graph neural network to build a GDN model for fault prediction, the correlation between sensors is fully considered, and the accuracy and reliability of the prediction are improved. At the same time, the method also adopts corresponding early warning strategies according to different fault types and severity, realizes timely detection and processing of faults, and effectively avoids the damage caused by faults to production efficiency and equipment. In summary, the method has the advantages of accurate prediction, timely early warning, and easy management, which is of great significance for improving the operating efficiency and reliability of the biomass pellet machine. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0016] Figure 1 A flow chart of a biomass pellet machine fault prediction method based on deep learning is provided for this application. DETAILED DESCRIPTION

[0017] 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.

[0018] Embodiment 1: Figure 1 As shown, the present invention provides a biomass pellet machine fault prediction method based on deep learning, the method comprising: Step S100: collecting the operation data of the biomass pellet machine in real time through the sensors installed on the biomass pellet machine.

[0019] Specifically, the sensors include temperature sensors, pressure sensors, vibration sensors, flow sensors and speed sensors. The biomass pellet machine operation data includes temperature data, pressure data, vibration frequency data, material flow data and motor speed data.

[0020] In this embodiment, the temperature sensor uses the ETS386-3-150-000 temperature sensor; the pressure sensor uses the 2699 S18UUA pressure sensor; the vibration sensor LV-S61 laser displacement sensor indirectly monitors the vibration frequency by measuring the displacement change caused by vibration; the flow sensor uses the ZYLC oval gear flowmeter; the speed sensor uses the HCNJ-101 speed torque sensor. All sensors transmit data to the data center via the CAN bus.

[0021] Each sensor encodes the collected data and packages the data into CAN frames. Each CAN frame contains a data identifier (ID), a data length code (DLC) and a data field. The sensor sends the data frame to the CAN bus through the CAN bus interface. The CAN bus is a multi-master-slave serial communication network. The control center receives the data frame from the sensor through the CAN bus interface, identifies the source and type of the data according to the data identifier (ID), and then parses the data field to extract temperature data, pressure data, vibration frequency data, material flow data and motor speed data.

[0022] Step S200: preprocessing the collected biomass pellet machine operation data.

[0023] Specifically, data preprocessing includes data cleaning, missing value filling, outlier processing, and data normalization.

[0024] In this embodiment, the control center obtains the biomass pellet machine operation data collected by the sensor and then performs data preprocessing. The biomass pellet machine operation data is stored as a CSV file named 'biomass_pellet_machine_data.csv', which has multiple columns, including 'timestamp', 'temperature', 'pressure', 'vibration_frequency', 'material_flow', and 'motor_speed'. Next, the read_csv function in the Pandas library is used to read the CSV file, and the read data is stored in a DataFrame object named data.

[0025] In the data cleaning stage, the drop_duplicates function was used to remove duplicate data rows in data to ensure the uniqueness of the data. Then, missing values ​​were filled. For each column in data except the 'timestamp' column, the average value was calculated and used to fill all missing values ​​in the column. In the outlier processing stage, the IQR (interquartile range) method was used to detect and process outliers. For each column in data except the 'timestamp' column, the first quartile (Q1) and the third quartile (Q3) of the column were first calculated, and then the IQR (i.e., Q3-Q1) was calculated. Then, the upper and lower bounds of the outliers were set (Q1-1.5IQR and Q3+1.5IQR, respectively), and values ​​less than the lower bound or greater than the upper bound were replaced with the average value of the column. Finally, data normalization was performed, and all columns except the 'timestamp' column were selected as feature columns that need to be normalized, and the data of these feature columns were scaled to the range of [0,1] using MinMaxScaler.

[0026] Step S300: Construct a GDN model using a graph neural network.

[0027] Specifically, the GDN model is trained using historical biomass pellet machine operation data and historical fault information, which are obtained from the control center and divided into training set and validation set according to 8:2. The GDN model includes sensor embedding, which uses embedding vectors to capture the characteristics of each sensor; graph structure learning, which learns the relationship between sensor pairs and encodes them as edges in the graph; graph attention-based prediction, which predicts the future behavior of sensors based on the attention function of adjacent sensors in the graph; and graph deviation score, which is used to evaluate the deviation between the score and the actual situation.

[0028] In this embodiment, the sensor is embedded first. For the biomass pellet machine operation data, there are similarities between the data. For example, the temperature data inside the biomass pellet machine will also affect the pressure data. Therefore, a flexible way is used to represent each sensor. Each sensor is represented as , d is the d features extracted from the sensor, and the similarity between these embeddings It indicates the similarity of behaviors, so sensors with similar embedding values ​​have a high correlation.

[0029] Then we conduct graph structure learning, using a directed graph, where nodes represent sensors and edges represent dependencies between sensors. The edge from one sensor to another indicates that the first sensor is used to simulate the behavior of the second sensor. The adjacency matrix A is used to represent this directed graph. The prior information is represented as a set of candidate relationships for each sensor i. , compute nodes The embedding vector of its candidate node The similarity between:

[0030]

[0031] That is, first calculate the embedding vector of sensor i Relationship with Candidate Then select the normalized dot product between values, Represents the normalized dot product The node index of the value.

[0032] Then define the attention-based graph model, in time When the size is The sliding window defines the model input for historical time series data , Represents the historical sequence time window, judging the value at the current moment :

[0033] In order to capture the relationship between sensors, a feature extractor based on graph attention is introduced. Based on the learned graph structure, the information of a node is fused with its neighboring nodes. Different from the existing graph attention mechanism, the feature extractor contains the sensor embedding vector , compute nodes Aggregate representation of :

[0034] in For Node The input features of is the neighbor set of node i obtained by learning the adjacency matrix A, is a trainable weight matrix, performing a shared linear transformation on each node, paying attention to the coefficients Calculated as

[0035]

[0036]

[0037] Indicates series connection, Embed the sensor And the corresponding transformed features Connect is the vector of learning coefficients of the attention mechanism. LeakyReLU is used as the nonlinear activation to calculate the attention coefficients, and the softmax function is used to normalize the attention coefficients.

[0038] Through the above feature extractor, we get the representation of all N nodes, namely For each , embedding it with the corresponding time series Element-wise multiplication (expressed as ), and the stacked results calculated by all nodes are used as the input of the fully connected layer. The output dimension is , thus predicting the time Vector sensor values , while the sensor observation data for:

[0039] Using sensors to predict output and sensor observation data The mean squared error between as a function to minimize the loss:

[0040] Finally, the graph deviation score is performed. Time sensor The outliers are:

[0041] Since different sensors have different characteristics, their deviation values ​​also have different scales. In order to prevent the deviation of any one sensor from exceeding that of other sensors, the error value of each sensor is normalized:

[0042] in and Represents the median and quartiles, taking the largest anomaly value of all sensors as the anomaly score:

[0043] if If it exceeds the threshold, it will be marked with the abnormal score, and the fault of the biomass pellet machine will be judged according to the abnormal score, the sensor to which it belongs, and the related sensors. The threshold is obtained by iterative fitting of the model on the validation set.

[0044] Step S400: input the pre-processed real-time biomass pellet machine operation data into the trained GDN model, and output the fault prediction result.

[0045] Specifically, the preprocessed operating data is input into the sensor embedding part of the GDN model to obtain the embedding vector of each sensor; the relationship between sensors is learned and encoded through the graph structure learning part; the future behavior of the sensor is predicted based on the relationship between the embedding vector and the sensor using the graph attention-based prediction part; the graph deviation score is calculated to evaluate the deviation between the prediction result and the actual situation, and the fault prediction result is output.

[0046] In this embodiment, first, the preprocessed real-time biomass pellet machine operation data is fed into the sensor embedding part of the GDN model, and each sensor is represented as an embedding vector that captures the key features of the sensor, and the similarities between these features reflect the similarities between the sensor behaviors.

[0047] The data then passes through the graph structure learning part, where the model learns and encodes the relationship between sensors and constructs a directed graph in which nodes represent sensors and edges represent dependencies between sensors. The dependencies are determined by calculating the similarity between sensor embedding vectors. Sensors with high similarity will have stronger connections.

[0048] Next, the data enters the prediction part based on graph attention. The model uses the learned graph structure and sensor embedding vector to predict the future behavior of the sensor. The model introduces a feature extractor based on graph attention to fuse the information of the node with the information of its neighboring nodes to capture the relationship between the sensors. After calculating the representations of all nodes, these representations are used as input to the fully connected layer to output the prediction results, which include the predicted values ​​of the sensors at a certain point in the future. The model then calculates the graph deviation score to evaluate the deviation between the prediction results and the actual situation. The graph deviation score is calculated by comparing the mean square error between the sensor prediction output and the sensor observation data.

[0049] Finally, based on the graph deviation score and the threshold obtained through training, the model outputs the fault prediction results, which include the fault type, fault occurrence time, and fault severity. The fault type is determined based on the deviation pattern between the predicted result and the actual situation and the corresponding sensor. The fault occurrence time is determined based on the time when the abnormal value appears in the predicted result. The fault severity is evaluated based on the size of the deviation and is divided into three levels: minor, moderate, and severe.

[0050] Fault types include abnormal temperature, abnormal pressure, abnormal vibration, abnormal flow and abnormal speed. Taking abnormal temperature as an example, it is due to the deviation of the temperature sensor, resulting in abnormal temperature value detected, which may appear as overheating or overcooling. For this type of fault, the temperature sensor is the main responsible sensor, while the pressure sensor and vibration sensor may also be affected by temperature changes and become related sensors.

[0051] The average value of the deviation between the sensor observation value and the predicted value is set to X, the standard deviation is set to Y, and the minor fault threshold is set to X+0.5Y. When the deviation value is less than or equal to X+0.5Y, the system considers that the fault impact is small; similarly, the moderate fault threshold is set to X+Y. When the deviation value is greater than the minor fault threshold and less than or equal to X+Y, the system considers that the fault impact is moderate; the severe fault threshold is set to any value greater than the moderate fault threshold, that is, when the deviation value is greater than X+Y, the system considers that the fault impact is severe.

[0052] Step S500: adopt corresponding early warning strategies according to the fault prediction results.

[0053] Specifically, the fault prediction results include the fault type, fault occurrence time and fault severity, which are divided into three levels: minor, medium and severe. The early warning strategy includes: sending different levels of early warning signals according to the fault type and fault severity; for minor faults, sending maintenance reminders, regular inspection and maintenance; for medium faults, sending maintenance notifications, and taking maintenance measures immediately; for severe faults, sending shutdown notifications, and immediate shutdown inspection.

[0054] In this embodiment, the fault prediction result provides detailed information about possible faults of the biomass pellet machine, including the fault type, fault occurrence time, and fault severity. The fault types include abnormal temperature faults, abnormal pressure faults, abnormal vibration faults, abnormal flow faults, and abnormal speed faults, which are directly related to the key operating parameters of the biomass pellet machine. The fault occurrence time clarifies the time point when the abnormal value appears, providing a time basis for taking timely measures. The fault severity is divided into three levels: minor, medium, and severe. Each level corresponds to a different early warning strategy to ensure that the fault can be handled in a timely and effective manner.

[0055] For minor faults, the system sends a maintenance reminder signal. Although this type of fault will not cause an immediate serious impact on the operation of the biomass pellet machine, if it is not handled in time, it may gradually worsen and lead to more serious faults. Therefore, the system sends maintenance reminders to remind operators to regularly check and maintain related equipment to prevent the occurrence of faults.

[0056] For medium faults, the system sends a maintenance notification signal to take maintenance measures immediately. Medium faults have already had a certain impact on the operation of the biomass pellet machine. If not handled in time, it may cause equipment shutdown or damage. Therefore, the system sends a maintenance notification to require the operator to take immediate measures to make necessary repairs or replacements on the equipment to ensure the normal operation of the equipment.

[0057] For serious faults, the system sends a shutdown notification signal and immediately stops the machine for inspection. Serious faults have already had a serious impact on the operation of the biomass pellet machine. If it is not immediately stopped for inspection, it may cause serious damage to the equipment or even cause a safety accident. Therefore, the system sends a shutdown notification and conducts a comprehensive inspection and maintenance of the equipment to ensure the safe and stable operation of the equipment.

[0058] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A biomass pellet machine fault prediction method based on deep learning, characterized in that: The method comprises: Step S100: obtaining the operation data of the biomass pellet machine in real time through the sensor installed on the biomass pellet machine; Step S200: preprocessing the real-time biomass pellet machine operation data; Step S300: constructing a GDN model using a graph neural network, and using historical biomass pellet machine operation data and historical fault data for training; Step S400: input the pre-processed real-time biomass pellet machine operation data into the trained GDN model, and output the fault prediction result; Step S500: adopt corresponding early warning strategies according to the fault prediction results.

2. The biomass pellet machine fault prediction method according to claim 1, characterized in that: The sensors include a temperature sensor, a pressure sensor, a vibration sensor, a flow sensor and a rotation speed sensor.

3. The biomass pellet machine fault prediction method according to claim 2, characterized in that: The biomass pellet machine operation data includes temperature data, pressure data, vibration frequency data, material flow data and motor speed data.

4. The biomass pellet machine fault prediction method according to claim 1, characterized in that: The data preprocessing includes data cleaning, missing value filling, outlier processing and data normalization.

5. The biomass pellet machine fault prediction method according to claim 1, characterized in that: The GDN model includes: Sensor embedding, using embedding vectors to capture the behavioral characteristics of each sensor on the biomass pellet machine; Graph structure learning, using a directed graph to represent sensors and their dependencies, with nodes being sensors and edges being dependencies, and calculating the similarity between node embedding vectors and nodes; Attention-based graph model, introduces a graph attention-based feature extractor, uses a sliding window to capture the relationship between sensors, fuses the information of a node with its neighboring nodes, and calculates the aggregate representation; Sensor prediction, obtains the representation of all nodes through the feature extractor, multiplies it with the time series embedding, inputs it into the fully connected layer, and outputs the value of the sensor at the predicted time; Figure 1. Deviation score. According to the deviation between the sensor predicted output and the sensor observed data, the sensor anomaly value is calculated. The maximum anomaly value of all sensors is taken as the anomaly score. If it exceeds the anomaly threshold, it is marked as abnormal. The fault condition of the biomass pellet machine is judged according to the anomaly score, the sensor to which it belongs, and related sensors.

6. The biomass pellet machine fault prediction method according to claim 5, characterized in that: The historical biomass pellet machine operation data includes historical temperature data, historical pressure data, historical vibration frequency data, historical material flow data, historical motor speed data and a timestamp.

7. The biomass pellet machine fault prediction method according to claim 5, characterized in that: The step S400 specifically includes: Input the preprocessed real-time biomass pellet machine operation data into the trained GDN model; The GDN model outputs an anomaly score. When the anomaly score exceeds the anomaly threshold, it is marked as an abnormal state. According to the abnormal score, the belonging sensor and the related sensor are judged. The belonging sensor is the sensor directly affected when the biomass pellet machine fails, and the related sensor is the sensor indirectly affected when the biomass pellet machine fails. According to the abnormal score, the information of the corresponding sensor and related sensors, the fault condition of the biomass pellet machine is judged, including the fault type, fault occurrence time and fault severity.

8. The biomass pellet machine fault prediction method according to claim 7, characterized in that: The fault severity is divided into three levels: slight, medium and severe according to the deviation between the biomass pellet machine sensor predicted output and the sensor observed data.

9. The biomass pellet machine fault prediction method according to claim 8, characterized in that: The early warning strategy includes: Send different levels of early warning signals according to the fault type and fault severity; For minor faults, the system sends maintenance reminders to remind you to regularly check and maintain the biomass pellet machine; For medium faults, the system sends a maintenance notification and takes maintenance measures on the biomass pellet machine immediately; For serious faults, the system sends a shutdown notification and immediately shuts down the biomass pellet machine for inspection.

Images