Method and system for predicting use state of lining plate of crusher based on TFT (Thin Film Transistor)

Abstract

The objective of the invention is to provide a TFT-based crusher lining plate use state prediction method and system in order to solve the problems of the existing crusher lining plate service life prediction technology. In the invention, the TFT is used as a core backbone structure of the model, and an improved module for explicit alignment and multi-task prediction is added on the basis. The method comprises the following steps: obtaining time sequence data and scalar data by obtaining and preprocessing data related to lining plate consumption; and constructing a life prediction model, and performing feature extraction and prediction on the obtained data to obtain a prediction result. According to the method, a lining plate sensor does not need to be adopted, the cost of the method for predicting the service life of the lining plate is reduced, the prediction accuracy is improved, the state and the use condition of the crusher lining plate can be analyzed in real time, the situation that equipment is shut down abnormally due to abrasion of the lining plate is greatly reduced, and then production is improved and stabilized.

Description

Technical Field

[0001] This invention belongs to the field of equipment life prediction technology in equipment management, and particularly relates to a method and system for predicting the service status of crusher liners based on TFT (Temporal Fusion Transformer). Background Technology

[0002] Crushers are widely used material crushing equipment in industrial production. During the crushing process, the liner, as a critical component of the crusher, bears enormous impact forces and wear. With increasing usage time, the liner gradually wears down, directly affecting ore processing efficiency and equipment operating safety. As the mining depth of mineral resources increases and the ore grade decreases, the amount of non-minerals in the ore increases, leading to increasingly prominent liner wear problems. Effectively predicting the remaining service life (RUL) of the liner is crucial for avoiding unplanned downtime and reducing maintenance costs.

[0003] The service life of liners is affected by a variety of factors, including ore conditions, equipment conditions, and production conditions.

[0004] Currently, the relevant literature on methods for predicting the service life of crusher liners is as follows:

[0005] In existing technologies, the status and remaining life of crusher liners are typically monitored and predicted online by installing sensors on the liner or equipment side. For example, a digital wireless sensor is installed in the concave liner of a gyratory crusher and coupled with a discrete element model to reconstruct the thickness distribution and wear trajectory (Ou et al., Modelling of gyratory crusherliner wear using a digital wireless sensor, 2023); or an industrial solution is to embed proprietary sensors into the liner casting to output the wear level in real time and guide replacement (WO2020210875A1, wear sensing liner); in addition, an online monitoring method is disclosed that uses an isokinetic wear probe inserted through the shell and uses ultrasonic flight time or vibration amplitude to jointly evaluate the liner wear and impact location (WO2022000072A1, Monitoring liner wear in industrial mills). The common feature of the above methods is that they all require the installation of sensors on the liner or equipment side. They face problems such as power supply and sealing difficulties in high impact, high abrasion and dust heat environments, high calibration and maintenance costs, high dependence on the accuracy and completeness of historical data, poor model generalization ability, and difficulty in cross-model migration. On the other hand, the solution based on downtime measurement cannot make continuous online predictions and is difficult to reflect the life evolution under complex operating conditions in a timely manner.

[0006] In summary, there is currently no effective solution in the technology that can effectively predict the complex factors affecting the service life of the liner without using a liner sensor. Summary of the Invention

[0007] The purpose of this invention is to address the problems existing in current crusher liner life prediction technologies by proposing a crusher liner usage status prediction method and system based on TFT (Temporal Fusion Transformer). The TFT model is a time series prediction model that combines a gated residual network with a multi-head attention mechanism, capable of simultaneously handling static and dynamic features and automatically capturing the dependencies between key time slices and important variables. In this invention, TFT is used as the core backbone structure of the model, and improved modules for explicit alignment and multi-task prediction are added on top of it. This method acquires and preprocesses data related to liner consumption to obtain time series data and scalar data; then, it constructs a life prediction model, extracts features from the acquired data, and makes predictions to obtain the prediction results. This method eliminates the need for liner sensors, reducing the cost of predicting liner life and improving prediction accuracy. Furthermore, it enables real-time analysis of the crusher liner's status and usage, significantly reducing equipment downtime due to liner wear, thereby improving production and stabilizing production.

[0008] One of the technical solutions of the present invention is a method for predicting the service status of crusher liners based on TFT, comprising the following steps:

[0009] S1. Obtain production data and liner technology data related to crusher liner consumption, including time series data composed of historical data and scalar data that does not change over time. Preprocess these data to obtain a dataset of time series data and a dataset of scalar data. Then divide the dataset of time series data and the dataset of scalar data into training dataset and test dataset, respectively.

[0010] The time series data includes ore hardness, particle size distribution, crusher power, silo level, feed rate, feed particle size, discharge particle size, and liner replacement cycle.

[0011] The scalar data includes the shape, structure, material, hardness, crusher model, and crushing stage of the liner;

[0012] The method for preprocessing data is as follows:

[0013] S11. Remove irrelevant and abnormal data;

[0014] S12. Integrate data from the same data source and then unify the timestamps to form a unified dataset;

[0015] Integrating data from the same data source includes extracting common attribute columns from different data sources, such as liner replacement cycle and ore feed rate; unifying timestamps by using explicit time alignment to synchronize and interpolate timestamps of multi-source data, ensuring that data with different sampling frequencies and time windows remain consistent in the time dimension, and ensuring that the same attribute columns from different event sources are aligned in time to form a unified dataset.

[0016] S13. Perform min-max normalization on the data in the dataset to scale the data to the [0,1] interval;

[0017] The aforementioned minimum-maximum normalization of the data can be performed using the following formula:

[0018]

[0019] Where x represents the original data points, and min(x) and max(x) are the minimum and maximum values ​​in the dataset, respectively. The data is after normalization;

[0020] S2. Construct a lifetime prediction model

[0021] The lifetime prediction model is implemented based on TFT as a framework. The TFT model utilizes gated residual structure and temporal attention mechanism to achieve multidimensional time dependency modeling. The lifetime prediction model includes a feature extraction module and a prediction module. The feature extraction module includes a temporal processing unit, a scalar processing unit, and a multi-head attention unit. The prediction module consists of an LSTM (Long Short-Term Memory) unit and a multi-task prediction unit including multiple parallel FC (fully connected) layers, wherein the multiple parallel FC layers are connected to the LSTM unit respectively.

[0022] S21, Feature Extraction of Data

[0023] The data features are extracted through the model’s temporal processing unit and scalar processing unit to obtain temporal features and scalar features respectively. The two are then fused together, and the key and information-rich parts of the features are automatically extracted through the multi-head attention unit to obtain the feature vector finally output by the feature extraction module.

[0024] S211, Feature Extraction of Time Series Data

[0025] The time series data is processed in the time series processing unit. First, it passes through two consecutive one-dimensional convolutional layers (i.e., the first convolutional layer and the second convolutional layer) to capture the local temporal features of the data. Then, it passes through a time attention layer to identify key information in the time series through a weighting mechanism. Finally, it passes through a global max pooling layer to obtain an output vector of fixed size k, which is the final time series feature.

[0026] S212, Feature Extraction of Scalar Data

[0027] Scalar data is processed in a scalar processing unit, and then passed through a fully connected layer to expand the dimension of each scalar into an array of fixed size k, so that the scalar data and the time series data are aligned in dimension, thus obtaining scalar features;

[0028] S213. Merge the time series features obtained in S211 and the scalar features obtained in S212 into a matrix in the k-dimensional direction to form a merged feature vector; then pass it through a multi-head attention layer to enhance the correlation between the time series data and the scalar data, and obtain the final feature vector of the feature extraction module.

[0029] S22. Input the final feature vector of the feature extraction module into the prediction module of the model. Use the multi-task learning framework of the multi-task prediction unit to predict three tasks at the same time: liner wear state, crusher operating state and liner remaining service life (RUL).

[0030] S221. Input the final feature vector into the LSTM unit of the prediction module to capture the long-term and short-term dependencies in the time series data.

[0031] The general features learned by the S222 and LSTM units are input into the three task-specific fully connected layers of the multi-task prediction unit for further processing and learning, so as to obtain the prediction results for three tasks: liner remaining service life (RUL), liner wear status, and crusher operating status.

[0032] S3. Train the lifespan prediction model

[0033] S31. Constructing the loss function

[0034] The remaining useful life (RUL) of the primary liner is constructed as a regression task, and its loss function is... The calculation formula is as follows:

[0035]

[0036] Where B is the number of samples in a single iteration. and These are the true value and the predicted value of the j-th sample, respectively;

[0037] The auxiliary tasks of liner wear status and crusher operating status are constructed as classification tasks, and their loss function is... The calculation formula is as follows:

[0038]

[0039] Where B is the number of samples in a single iteration, and C is the number of classes. and These are vectors representing the true label and predicted probability of the j-th sample, respectively.

[0040] The three losses are summed according to their assigned weights to obtain the total loss:

[0041] +

[0042] in, The weights of the remaining useful life loss function for the main task liner. The weights of the loss function for the liner wear state in subtask one. The weights of the loss function for the crusher's operating state in subtask two;

[0043] Better , , They were set to 1, 0.1, and 0.1 respectively.

[0044] S32, Iterative Training

[0045] The training datasets of time series data and scalar data are input into the model built by S2, the loss between the predicted results and the true labels is calculated, the backpropagation algorithm is executed, and the model parameters are updated according to the loss.

[0046] S33, Model Testing

[0047] Input the test datasets of time series data and scalar data into the iteratively trained model for testing. If the test results are satisfactory, the model training ends; otherwise, proceed to steps S32 and S33 to continue training and testing until the test results are satisfactory, and the trained model is obtained.

[0048] S34. Use root mean square error and scoring function to evaluate the accuracy of model predictions;

[0049] The loss function is used during the training phase, and the scoring function is used during the testing phase.

[0050] The formula for calculating the root mean square error is as follows:

[0051]

[0052] Where N is the total number of evaluation processes (the number of samples in the test set). and These are the true value and the predicted value of the i-th sample, respectively.

[0053] The calculation formula for the scoring function is as follows:

[0054]

[0055] in, Let be the weight of the i-th prediction error, and N be the total number of evaluation processes. These are parameters used to adjust the score;

[0056] To reflect the importance of errors at different forecast periods, for example, if the accuracy of early forecasts is more important than the accuracy of late forecasts, this can be reflected by setting weights; Depending on the machine's duty cycle count, for example, if you are more concerned about avoiding premature parts replacement (avoiding unnecessary maintenance costs), you can choose a larger one. A smaller value can be chosen to increase the penalty for premature prediction; if the primary concern is avoiding the potential failure risk caused by late prediction, a smaller value can be selected. value;

[0057] S4. Collect time series and scalar data of crusher liners in actual production, and use the trained model to predict the liner usage status.

[0058] The second technical solution of the present invention is a TFT-based crusher liner life prediction system, the system comprising:

[0059] The data input module is used to acquire and preprocess data related to liner consumption to obtain time series data and scalar data.

[0060] The feature extraction module includes a time series processing unit, a scalar processing unit, and a multi-head attention unit. The time series processing unit processes the time series data to obtain a time series feature vector; the scalar processing unit processes the scalar data to obtain a scalar feature vector; and the multi-head attention unit obtains a total feature vector based on the time series feature vector and the scalar feature vector.

[0061] The prediction module is used to output prediction results based on the total feature vector. The prediction results include the liner wear status, the crusher operating status, and the remaining service life of the liner.

[0062] Furthermore, in the aforementioned system, the data input module includes preprocessing such as deleting irrelevant and abnormal data, data integration, and data normalization.

[0063] Furthermore, in the aforementioned system, the scalar data includes liner shape, structure, material, hardness, crusher model, and crushing stage; the time series data includes ore hardness, particle size distribution, crusher power, silo level, feed rate, feed particle size, discharge particle size, and liner replacement cycle.

[0064] Compared with the prior art, the advantages of the present invention are:

[0065] 1. This method eliminates the need for liner sensors, which not only reduces the cost of predicting liner lifespan but also improves the accuracy of the prediction.

[0066] 2. This method combines historical data and scalar data, adds an attention mechanism, and adds a multi-task learning mechanism, which improves the accuracy of life prediction. It can analyze the status and usage of crusher liners in real time and accurately predict the remaining service life.

[0067] 3. This method not only reduces prediction costs and dependence on on-site hardware, but also significantly improves the generalization ability across operating conditions and machine types, providing more stable and efficient operation and maintenance decision support for mine production.

[0068] 4. This method improves the structure of conventional TFTs to address the characteristics of multi-source data in industrial applications:

[0069] 1) An explicit time alignment method was added to the input stage to align the differences in sampling frequency and time window of time-series features from different sources (such as power, material level, feed rate, etc.); 2) A "multi-task prediction unit" was added to the output stage to achieve joint prediction of multiple related indicators such as liner wear status, crusher operating status and liner remaining service life. Attached Figure Description

[0070] Figure 1 This is a schematic flowchart of the method of the present invention;

[0071] Figure 2 This is a schematic diagram of the lifespan prediction model architecture of the present invention;

[0072] Figure 3 This is a schematic diagram of the time-series and scalar data types that need to be collected in this invention;

[0073] Figure 4 This is a schematic diagram illustrating the composition of time series data from historical data of a certain crusher in an embodiment of the present invention. Detailed Implementation

[0074] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0075] To address the shortcomings of existing technologies for predicting the service life of crusher liners, this solution not only considers time-series data composed of historical data but also incorporates scalar data, such as liner shape, liner structure, and manufacturing date, enabling analysis of factors affecting liner service life from multiple dimensions. Furthermore, by introducing an attention mechanism, this solution allows the model to identify and focus on analyzing key time periods in the time series and important features in the scalar data, improving the model's ability to identify wear patterns and reducing reliance on the completeness and accuracy of historical data. Finally, this solution employs a multi-task learning framework, enabling the model to simultaneously predict multiple related tasks, such as liner wear status, crusher operating status, and remaining service life, thereby improving the model's generalization ability and prediction accuracy.

[0076] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will use data from a primary crusher at an iron ore beneficiation plant from 2022 to 2024 as an example, and describe the embodiments of the present invention in further detail with reference to the accompanying drawings.

[0077] In a specific embodiment, the main process of this solution is to collect various data related to liner consumption, including time-series data composed of historical data and scalar data that does not change over time, such as liner technical condition and ore conditions. The collected data is preprocessed by data cleaning, fusion, and normalization. The data processing module of the model is used to extract data features, fuse time-series features with scalar features, and automatically extract the key and information-rich parts of the features through an attention mechanism. The feature vector is input into the prediction module of the model, and a multi-task learning framework is used to predict three tasks simultaneously: liner wear status, crusher operating status, and liner remaining service life (RUL).

[0078] Based on the above, combined with Figure 1 As shown, the execution flow of this embodiment preferably includes the following steps:

[0079] S1. Collect various production data and liner technology data related to crusher liner consumption, including time-series data composed of historical data and scalar data that does not change over time, such as liner technical condition, ore condition, equipment condition, and production condition. Perform data cleaning, data fusion, normalization, and other preprocessing on the collected data; obtain the time-series data dataset and the scalar data dataset; and divide the time-series data dataset and the scalar data dataset into training dataset and test dataset, respectively.

[0080] Specifically, in this embodiment, the implementation process of S1 is as follows:

[0081] S11. The various data related to liner consumption to be collected specifically include: data related to the technical condition of the liner, such as shape, structure, material, and hardness; data related to ore conditions, such as ore hardness and particle size distribution; data related to equipment conditions, such as crusher model, crusher power, crushing stage number, and silo level; and data related to production conditions, such as feed particle size, discharge particle size, feed rate, and replacement cycle. Among these, liner model and specifications, crusher model, and other indicators do not change over time and are scalar data, as shown in Table 1. Ore hardness, particle size distribution, crusher power, silo level, and feed rate, etc., change over time, and historical data are collected to form time series data, as shown in Table 1. Figure 4 As shown.

[0082] Table 1. Scalar Data of Coarse Crusher

[0083]

[0084] S12 preprocesses the collected time series and scalar data, specifically including: removing irrelevant and outlier data; integrating data from the same data source and then unifying the timestamps to form a unified dataset for easier analysis; and performing min-max normalization on the data to scale it to the [0,1] interval.

[0085] Preferably, integrating data from the same data source includes extracting common attribute columns from different data sources; unifying timestamps, and using an explicit time alignment module to synchronize and interpolate timestamps of multi-source data, so that data with different sampling frequencies and time windows remain consistent in the time dimension;

[0086] Preferably, the data undergoing minimum-maximum normalization can be converted using the following formula:

[0087]

[0088] Where x represents the original data points, and min(x) and max(x) are the minimum and maximum values ​​in the dataset, respectively. This is the normalized data.

[0089] S2. Constructing a lifetime prediction model. In this embodiment, the model is based on TFT, which serves as the core backbone structure of the model. An improved multi-task prediction component is added on top of this. The TFT model is a time series prediction model that combines a gated residual network with a multi-head attention mechanism. It can simultaneously handle static and dynamic features and automatically capture the dependencies between key time slices and important variables. Figure 2As shown, in this embodiment, the lifetime prediction model mainly includes a feature extraction module and a prediction module. The feature extraction module is connected to the prediction module. The feature extraction module includes a time series processing unit, a scalar processing unit, and a multi-head attention unit. The time series processing unit is mainly used to process time series data, and the scalar processing unit is used to process scalar data. The time series processing unit and the scalar processing unit send the processed data to the multi-head attention unit. The time series processing unit consists of a first convolutional layer (i.e., ...) connected in sequence. Figure 2 conv1 in the middle), the second convolutional layer (i.e. Figure 2 The first convolutional layer consists of a conv2 layer, a temporal attention layer, and a global max pooling layer. Time-series data is input into the first convolutional layer. The scalar processing unit consists of fully connected (FC) layers. The prediction module consists of an LSTM unit and a multi-task prediction unit comprising multiple parallel FC layers, wherein the multiple parallel FC layers are connected to the LSTM unit.

[0090] In this embodiment, the model's temporal processing unit and scalar processing unit are used to extract data features, the temporal features and scalar features are fused, and the key and information-rich parts of the features are automatically extracted by the multi-head attention unit and input into the subsequent prediction module for prediction.

[0091] Specifically, in this embodiment, the above S2 implementation process is as follows:

[0092] S21. The feature extraction process for time series data is as follows: First, the data passes through two consecutive one-dimensional convolutional layers (i.e., the first convolutional layer and the second convolutional layer) to capture the local temporal features of the data; then, it passes through a time attention layer to identify key information in the time series through a weighting mechanism; finally, it passes through a global max pooling layer to transform the multi-dimensional feature map into an output vector of fixed size k, which is the output of the time series part of the feature extraction module, i.e., the time series features.

[0093] S22. The feature extraction process for scalar data is as follows: The scalar data passes through a fully connected layer, which expands the dimension of each scalar into an array of fixed size k, so that the scalar data can be aligned with the time series data in terms of dimension, which is convenient for subsequent merging processing. At this time, the output of the scalar part is obtained, that is, the scalar features.

[0094] S23. After obtaining the outputs of the time series and scalar parts respectively, these two parts are merged into a single matrix. Since the outputs in both the time series processing unit and the scalar processing unit have been converted into fixed-length vectors, they can be directly concatenated in the k-dimensional direction to form the merged feature vector. This is then passed through a multi-head attention layer to enhance the correlation between the time series data and the scalar data, resulting in the final feature vector output by the feature extraction module.

[0095] S3. Input the feature vectors output by the feature extraction module into the prediction module of the model. Employ a multi-task learning framework to simultaneously predict three tasks: liner wear state, crusher operating state, and remaining liner service life (RUL). Figure 2 As shown;

[0096] Specifically, in this embodiment, the above-mentioned S3 implementation process is as follows:

[0097] S31. Input the feature vector into the LSTM unit to capture the long-term and short-term dependencies in the time series data;

[0098] S32. After the LSTM unit, add a layer consisting of the LSTM unit and three parallel fully connected layers (i.e., Figure 2 The multi-task prediction unit (FC layer) in the LSTM unit is responsible for handling a specific prediction task in each fully connected layer. The LSTM unit outputs the learned general features, which are then input into three task-specific fully connected layers for further processing and learning, and to generate the prediction results for the specific task. It should be noted that this embodiment takes three tasks as examples: liner wear status, crusher operating status, and liner remaining service life (RUL). Therefore, three parallel fully connected layers are set up accordingly. If the number of tasks increases or decreases, the number of fully connected layers can be increased or decreased accordingly.

[0099] In a more preferred embodiment, three different fully connected layers are set for the three tasks listed in this embodiment: For subtask one, the liner wear state, the output dimension is 10, the activation function is softmax, and the loss function is set to the cross-entropy loss function. The liner wear state is divided into 10 categories, with wear states increasing in severity from 1 to 10. For subtask two, the crusher operating state, the output dimension is 8, the activation function is softmax, and the loss function is the cross-entropy loss function. The crusher operating state is divided into 8 categories: normal operation, excessively high temperature, abnormal vibration, large current fluctuation, excessive noise, overload operation, blockage, and shutdown.

[0100] Specifically, the wear condition of the liner is graded based on historical replacement records and liner thickness measurement data. Specifically, the wear degree is divided into 10 intervals according to the attenuation ratio of the liner thickness relative to the initial thickness. Each interval represents a wear ratio change of approximately 10%, with the severity increasing from "1" (brand new condition) to "10" (critical replacement condition).

[0101] The crusher's operating status is classified based on the temperature, vibration, current, power, and noise signals collected by the monitoring system, combined with the operation log annotations. A total of 8 states are set: normal operation, excessively high temperature, abnormal vibration, large current fluctuation, excessive noise, overload operation, blockage, and shutdown. The settings are based on a combination of equipment operating experience and signal thresholds.

[0102] For the primary task, the remaining service life (RUL) of the liner, the output dimension is 1, and the loss function is the root mean square error (RMSE). Since both auxiliary tasks are closely related to the liner's service life, the model learns general features related to these states during the learning process, helping it better understand the context and enhance its predictive ability. Simultaneously, the total loss function is a weighted sum of the loss functions for the three tasks, ensuring that the auxiliary tasks do not negatively impact the primary task.

[0103] S33. Use root mean square error and scoring functions to evaluate the accuracy of model predictions;

[0104] The loss function is used during the training phase, and the scoring function is used during the testing phase. The RMSE loss function is used during the training phase to guide model parameter optimization and convergence judgment. The scoring function... Evaluation metrics belonging to the testing phase are used to comprehensively evaluate the prediction results on the test set after the model training is completed.

[0105] The formula for calculating the root mean square error is as follows:

[0106] Where N is the total number of evaluation processes (the number of samples in the test set). and These are the true value and the predicted value of the i-th sample, respectively.

[0107] The calculation formula for the scoring function is as follows:

[0108]

[0109] in, Let be the weight of the i-th prediction error, and N be the total number of evaluation processes. These are parameters used to adjust the score. The importance of different forecasts in terms of their error can be reflected by setting weights; for example, if the accuracy of earlier forecasts is more important than that of later forecasts, this can be reflected by setting weights. Depending on the machine's duty cycle count, for example, if you are more concerned about avoiding premature parts replacement (avoiding unnecessary maintenance costs), you can choose a larger one. A smaller value can be chosen to increase the penalty for premature prediction; if the primary concern is avoiding the potential failure risk caused by late prediction, a smaller value can be selected. value;

[0110] Specifically, the training process of the crusher liner service life prediction model based on CLLPNet is as follows:

[0111] S4. Train the lifespan prediction model

[0112] S41. Constructing the loss function

[0113] The remaining useful life (RUL) of the primary liner is constructed as a regression task, and its loss function is calculated as follows:

[0114]

[0115] Where B is the number of samples in a single iteration. and These are the true value and the predicted value of the j-th sample, respectively;

[0116] The auxiliary tasks of liner wear status and crusher operating status are constructed as classification tasks, and their loss function is calculated as follows:

[0117]

[0118] Where B is the number of samples in a single iteration, and C is the number of classes. and These are vectors representing the true label and predicted probability of the j-th sample, respectively.

[0119] The three losses are summed according to their assigned weights to obtain the total loss:

[0120] +

[0121] in, , , They were set to 1, 0.1, and 0.1 respectively. The weights of the remaining useful life loss function for the main task liner. The weights of the loss function for the liner wear state in subtask one. The weights of the loss function for the crusher's operating state in subtask two;

[0122] S42, Iterative Training

[0123] The training datasets of time series data and scalar data are input into the data processing module of S2. The processed data is then input into the prediction module for feedforward propagation. The loss between the prediction result and the true label is calculated, the backpropagation algorithm is executed, and the model parameters are updated based on the loss.

[0124] S43, Model Testing

[0125] The model is tested by inputting the test sets of the time series data dataset and the scalar data dataset. If the test result is qualified, the model training ends; otherwise, proceed to steps S32 and S33 until the test result is qualified, and the trained model is obtained.

[0126] S5. Collect time series and scalar data of crusher liners in actual production, and use the trained model to predict the liner usage status.

[0127] In summary, this example provides a TFT-based method for predicting the service life of crusher liners. The method involves collecting various data related to liner consumption, including time-series data composed of historical data and scalar data that does not change over time, such as liner technical condition and ore conditions. The collected data undergoes preprocessing such as data cleaning, data fusion, and normalization. The model's data processing module extracts data features, fusing time-series and scalar features, and automatically extracts key and information-rich parts of the features using an attention mechanism. The feature vectors are input into the model's prediction module, employing a multi-task learning framework to simultaneously predict three tasks: liner wear state, crusher operating state, and remaining liner service life (RUL). In practical applications, this can improve maintenance efficiency, reduce maintenance costs, enhance equipment management, and improve production safety.

[0128] In another implementation, this solution can also be achieved through a system, namely a TFT-based crusher liner lifespan prediction system, which can be configured as follows:

[0129] The data input module is used to acquire and preprocess data related to liner consumption to obtain time series data and scalar data.

[0130] The feature extraction module includes a time series processing unit, a scalar processing unit, and a multi-head attention unit. The time series processing unit processes the time series data to obtain a time series feature vector; the scalar processing unit processes the scalar data to obtain a scalar feature vector; and the multi-head attention unit obtains a total feature vector based on the time series feature vector and the scalar feature vector.

[0131] The prediction module is used to output prediction results based on the total feature vector. The prediction results include the liner wear status, the crusher operating status, and the remaining service life of the liner.

[0132] Preferably, in the data input module, preprocessing includes deleting irrelevant and abnormal data, data integration and extraction, and data normalization.

[0133] The scalar data includes liner model, crusher model, crushing stage number and material; the time series data includes ore hardness, particle size distribution, crusher power, silo level, feed rate, replacement cycle and crushing volume.

[0134] When the system is running, it can execute the TFT-based crusher liner usage status prediction method as detailed in the above embodiments.

[0135] In another embodiment, this solution can be implemented using a device, which may include corresponding modules that perform one or more steps in the various embodiments described above. Therefore, each or more steps in the various embodiments can be performed by a corresponding module, and the electronic device may include one or more of these modules. A module may be one or more hardware modules specifically configured to perform a corresponding step, or implemented by a processor configured to perform a corresponding step, or stored in a computer-readable medium for implementation by a processor, or implemented through some combination thereof.

[0136] This device can be implemented using a bus architecture. A bus architecture can include any number of interconnect buses and bridges, depending on the specific application of the hardware and overall design constraints. The bus connects various circuits, including one or more processors, memory, and / or hardware modules. The bus can also connect various other circuits such as peripherals, voltage regulators, power management circuitry, external antennas, etc.

[0137] The logic and / or steps represented in the flowchart or otherwise described herein may be specifically implemented in any readable storage medium for use by, or in conjunction with, an instruction execution system, apparatus or device (such as a computer-based system, a processor-included system or other system that can fetch and execute instructions from, an instruction execution system, apparatus or device).

[0138] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting the service status of crusher liners based on TFT, characterized in that, Includes the following steps: S1. Obtain production data and liner technology data related to crusher liner consumption, including time series data composed of historical data and scalar data that does not change over time. Preprocess these data to obtain a dataset of time series data and a dataset of scalar data. Then divide the dataset of time series data and the dataset of scalar data into training dataset and test dataset, respectively. S2. Construct a lifetime prediction model; The lifetime prediction model is based on TFT as a framework and includes a feature extraction module and a prediction module. The feature extraction module includes a timing processing unit, a scalar processing unit, and a multi-head attention unit. The prediction module consists of an LSTM unit and a multi-task prediction unit including multiple parallel FC layers, wherein the multiple parallel FC layers are respectively connected to the LSTM unit. S21. Feature extraction of data; The data features are extracted by the model’s temporal processing unit and scalar processing unit to obtain temporal features and scalar features respectively. The two are then fused and the feature vector is finally output by the feature extraction module through the multi-head attention unit. S22. Input the final feature vector of the feature extraction module into the prediction module of the model. Use the multi-task learning framework of the multi-task prediction unit to predict three tasks at the same time: liner wear state, crusher operating state and liner remaining service life (RUL). S3. Train the lifespan prediction model; S4. Collect time series and scalar data of crusher liners in actual production, and use the trained model to predict the liner usage status.

2. The TFT-based crusher liner usage status prediction method according to claim 1, characterized in that, In S1, the time series data includes ore hardness, particle size distribution, crusher power, silo level, feed rate, feed particle size, discharge particle size, and liner replacement cycle. The scalar data includes the shape, structure, material, hardness of the liner, crusher model, and crushing stage.

3. The TFT-based crusher liner usage status prediction method according to claim 1, characterized in that, In step S1, the method for preprocessing the data is as follows: S11. Remove irrelevant and abnormal data; S12. Integrate data from the same data source and then unify the timestamps to form a unified dataset; Integrating data from the same data source includes extracting common attribute columns from different data sources; Unified timestamps are used to synchronize and interpolate timestamps of multi-source data through explicit time alignment, so that data with different sampling frequencies and time windows remain consistent in the time dimension. S13. Perform min-max normalization on the data in the dataset to scale the data to the [0,1] interval; Use the following formula for conversion: ； Where x represents the original data points, and min(x) and max(x) are the minimum and maximum values ​​in the dataset, respectively. This is the normalized data.

4. The TFT-based crusher liner usage status prediction method according to claim 1, characterized in that, The specific steps of S21 are as follows: S211, Feature extraction from time series data; The time series data is processed in the time series processing unit. First, it goes through two consecutive one-dimensional convolutional layers, then through a time attention layer, and finally through a global max pooling layer to obtain an output vector of fixed size k, which ultimately yields the time series features. S212, Feature extraction of scalar data; Scalar data is processed in a scalar processing unit, and then passed through a fully connected layer to expand the dimension of each scalar into an array of fixed size k, so that the scalar data and the time series data are aligned in dimension, thus obtaining scalar features; S213. Merge the time series features obtained in S211 and the scalar features obtained in S212 into a matrix in the k-dimensional direction to form a merged feature vector. Then, pass through a multi-head attention layer to enhance the correlation between the time series data and the scalar data, and obtain the final feature vector of the feature extraction module.

5. The TFT-based crusher liner usage status prediction method according to claim 1, characterized in that, The specific steps of S22 are as follows: S221. Input the final feature vector into the LSTM unit of the prediction module to capture the long-term and short-term dependencies in the time series data. The general features learned by the S222 and LSTM units are input into the three task-specific fully connected layers of the multi-task prediction unit for further processing and learning, resulting in prediction results for three tasks: liner remaining service life (RUL), liner wear status, and crusher operating status.

6. The TFT-based method for predicting the service status of crusher liners according to claim 1, characterized in that, The specific steps of S3 are as follows: S31. Constructing the loss function The remaining useful life (RUL) of the primary liner is constructed as a regression task, and its loss function is... The calculation formula is as follows: ； Where B is the number of samples in a single iteration. and These are the true value and the predicted value of the j-th sample, respectively; The auxiliary tasks of liner wear status and crusher operating status are constructed as classification tasks, and their loss function is... The calculation formula is as follows: ； Where B is the number of samples in a single iteration, and C is the number of classes. and These are vectors representing the true label and predicted probability of the j-th sample, respectively. The three losses are summed according to their assigned weights to obtain the total loss: + ； in, The weights of the remaining useful life loss function for the main task liner. The weights of the loss function for the liner wear state in subtask one. The weights of the loss function for the crusher's operating state in subtask two; S32, Iterative Training The training datasets of time series data and scalar data are input into the model built by S2, the loss between the predicted results and the true labels is calculated, the backpropagation algorithm is executed, and the model parameters are updated according to the loss. S33, Model Testing The test datasets of time series data and scalar data are input into the iteratively trained model for testing. If the test results are satisfactory, the model training ends; otherwise, proceed to steps S32 and S33 to continue training and testing until the test results are satisfactory, and the trained model is obtained.

7. The TFT-based crusher liner usage status prediction method according to claim 1, characterized in that, S3 also includes using root mean square error and a scoring function to evaluate the accuracy of model predictions; The loss function is used during the training phase, and the scoring function is used during the testing phase. The formula for calculating the root mean square error is as follows: ； Where N is the total number of evaluation processes, and These are the true value and the predicted value of the i-th sample, respectively; The calculation formula for the scoring function is as follows: ； in, Let be the weight of the i-th prediction error, and N be the total number of evaluation processes. These are parameters used to adjust the score.

8. A TFT-based crusher liner service life prediction system, used to implement the TFT-based crusher liner service status prediction method of claim 1, characterized in that, The system includes: The data input module is used to acquire and preprocess data related to liner consumption to obtain time series data and scalar data. The feature extraction module includes a time series processing unit, a scalar processing unit, and a multi-head attention unit. The time series processing unit processes the time series data to obtain a time series feature vector; the scalar processing unit processes the scalar data to obtain a scalar feature vector; and the multi-head attention unit obtains a total feature vector based on the time series feature vector and the scalar feature vector. The prediction module is used to output prediction results based on the total feature vector. The prediction results include the liner wear status, the crusher operating status, and the remaining service life of the liner.

9. The TFT-based crusher liner service life prediction system according to claim 8, characterized in that, In the data input module, preprocessing includes deleting irrelevant and abnormal data, data integration, and data normalization.

10. The TFT-based crusher liner service life prediction system according to claim 8, characterized in that, The scalar data includes liner shape, structure, material, hardness, crusher model, and crushing stage; the time series data includes ore hardness, particle size distribution, crusher power, silo level, feed rate, feed particle size, discharge particle size, and liner replacement cycle.

Images