A development platform alarm data prediction method and system based on an LSTM neural network

Abstract

The application discloses a development platform alarm data prediction method and system based on an LSTM neural network, relates to the technical field of artificial intelligence, and comprises the following steps: a development platform monitoring system receives text data uploaded by a user, automatically performs preprocessing, and constructs an adaptive LSTM model; the adaptive LSTM model enters a model training stage and a model optimization stage, and the model performance is monitored and adjusted in real time; in the model training stage, the system monitors and displays performance indexes of a training process in real time, and when an index is abnormal, the system starts a feedback adjustment process; after adjustment, the system enters the model optimization stage, a sliding threshold method is applied to evaluate the model accuracy, and parameter adjustment improves the model prediction performance. The development platform alarm data prediction method based on the LSTM neural network improves the model training efficiency and prediction performance, reduces the complexity of model development and maintenance through the application of the sliding threshold method for model verification and decision-making, and provides clear model training feedback.

Description

TECHNICAL FIELD

[0001] The application relates to the field of artificial intelligence, and particularly to an alarm data prediction method and system based on an LSTM neural network. BACKGROUND

[0002] The long short-term memory (LSTM) neural network widely used in the field of artificial intelligence faces many challenges in achieving efficient and excellent model training. First, designing an LSTM model requires accurate selection of the number of hidden layers and the number of neurons, which usually depends on manual trial and error, which is time-consuming and cannot guarantee optimal results. Second, the selection of training parameters, such as learning rate, also has a huge impact on model training efficiency and performance. A too high learning rate may cause the model to miss the optimal solution, and a too low learning rate will result in a slow training process. Finally, the traditional model validation method usually uses a fixed threshold for evaluation, which may ignore the performance of the model at other thresholds and cannot fully evaluate the performance of the model. These problems greatly limit the application and development of LSTM models in various fields. The artificial intelligence development platform is mainly used to provide a unified model development and training platform for algorithm engineers, improving the efficiency of model development and training for algorithm engineers. The monitoring system, as an important functional module of the artificial intelligence development platform, realizes real-time monitoring of the running status and resource usage of nodes in the entire cluster, thereby effectively improving the utilization rate of computing resources. And when processing the monitoring data generated by large-scale clusters, the security of the monitoring data is guaranteed.

[0003] The traditional deep learning model training method usually includes data preprocessing, model design, model training and model validation. In this process, the training parameters, model structure and optimization algorithm usually need to be manually set, which requires experienced data scientists to manually adjust and optimize. This process is time-consuming and labor-intensive, and when dealing with large-scale data sets and complex models, performance is often limited by human resources and experience. In addition, the traditional model validation method usually uses a fixed threshold, which cannot well reflect the performance of the model at different thresholds.

[0004] Therefore, an alarm data prediction method based on an LSTM neural network is needed to automatically determine the number of hidden layers and the number of neurons of the model according to the characteristics of the input data, optimize the training process of the model through the method of adaptive learning rate, and improve the efficiency of model training and the prediction performance of the model. SUMMARY

[0005] This section is intended to summarize some aspects of the embodiments of the present application and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the abstract and title of the specification in order to avoid obscuring the purpose of this section, the abstract and the title of the specification, and such simplifications or omissions are not to be construed as limiting the scope of the present application.

[0006] In view of the above problems, the present application is proposed.

[0007] Therefore, the technical problem solved by the present application is that the existing artificial intelligence model training method is time-consuming and inefficient, and how to improve the training efficiency while ensuring the model performance and realize automatic optimization.

[0008] To solve the above technical problems, the present application provides the following technical solutions: a development platform alarm data prediction method based on an LSTM neural network, comprising:

[0009] The development platform monitoring system receives the text data uploaded by the user, automatically performs preprocessing, and constructs an adaptive LSTM model; the adaptive LSTM model enters a model training stage and a model optimization stage, and the model performance is monitored and adjusted in real time; in the model training stage, the system monitors and displays the performance indicators of the training process in real time, and when encountering abnormal indicators, the system starts a feedback adjustment process; after adjustment, it enters the model optimization stage, and the sliding threshold method is applied to evaluate the model accuracy, and the parameter adjustment improves the model prediction performance; the text data uploaded by the user is input into the optimized adaptive LSTM model for alarm prediction.

[0010] As a preferred scheme of the development platform alarm data prediction method based on the LSTM neural network, the automatic preprocessing includes data cleaning, missing value processing, text segmentation and conversion to word vectors, in the data cleaning stage, special characters and irrelevant words in the text are removed through regular expressions, and after preprocessing, the preprocessing results are displayed through the user interface.

[0011] As a preferred scheme of the development platform alarm data prediction method based on the LSTM neural network, the construction of the adaptive LSTM model includes constructing an adaptive LSTM model according to the distribution statistical information of the input data, the feature dimension size and the sample quantity;

[0012] The number of hidden layers is calculated, the number of neurons in each layer is automatically increased based on the experience formula of input feature dimension and sample quantity using a heuristic algorithm until the accuracy rate improvement value of the validation set in the model is less than the accuracy rate set value in the continuous three training periods;

[0013] The adaptive learning rate adjustment method is used to automatically select the learning rate, set the initial learning rate, and after each training period, if the accuracy of the model on the validation set is not improved, the learning rate is halved until the learning rate is less than the learning rate setting value, the activation function is automatically selected, and the activation function with the highest accuracy on the validation set is selected by traversing the ReLU, tanh and sigmoid activation functions.

[0014] As a preferred scheme of the development platform alarm data prediction method based on the LSTM neural network, the model training stage includes that, in the model training stage, the system uses the Adam optimizer with adaptive learning rate to train the model, and when the loss function value of the model in five continuous training periods is less than the loss function setting value, the system automatically reduces the learning rate to half of the current learning rate.

[0015] As a preferred scheme of the development platform alarm data prediction method based on the LSTM neural network, the model optimization stage includes that, in the model training process, the system collects the performance accuracy of the training process in real time, and when the verification accuracy in ten continuous training periods does not improve by more than 1%, the system automatically enters the parameter adjustment stage, automatically infers the problem reason according to the collected performance index in the training process, adjusts the corresponding model parameters according to the inference result, and restarts the training task using the new parameters.

[0016] If the verification accuracy in twenty continuous training periods does not improve by more than 1%, and at least one parameter adjustment has been performed during this period, the system automatically enters the model structure adjustment stage, adjusts the number of hidden layers, changes the activation function, changes the structure of the model, and restarts the training task using the new model structure.

[0017] If the verification accuracy in thirty continuous training periods does not improve by more than 1%, and at least one model structure adjustment has been performed during this period, the current model is considered to have reached the performance limit, a conservative strategy is adopted, the current model is temporarily saved, and the model is used as a benchmark to evaluate using the sliding threshold method.

[0018] As a preferred scheme of the development platform alarm data prediction method based on the LSTM neural network, the sliding threshold method includes that the system generates a model explanation report, decodes the output of the model on the validation set using the sliding threshold method, compares the decoding result with the true result, and calculates the accuracy ;

[0019] If the model on the validation set is evaluated as excellent, if is evaluated as good, if is evaluated as general, and if , the evaluation is poor.

[0020] As a preferred scheme of the LSTM neural network-based development platform alarm data prediction method, wherein: the model accuracy evaluation comprises evaluating the accuracy of the model on the test set;

[0021] If the accuracy of the model on the test set is excellent, the current training task is automatically stopped, and the model is deployed to the production environment;

[0022] If the accuracy evaluation is good, start the parameter fine-tuning stage, adjust the parameters of the model by cross-validation on the test set, and after fine-tuning, evaluate the accuracy of the model again to determine whether to deploy to the production environment;

[0023] If the accuracy evaluation is general, start the model structure adjustment stage, adjust the number of hidden layers, and change the activation function to change the structure of the model. After the structure adjustment is completed, the system will evaluate the accuracy of the model again and decide whether to enter the next stage;

[0024] If the accuracy evaluation is poor, abandon the current model, train a new model, use different parameters and structures from the abandoned model, and use the test set to verify the performance to determine the parameters and learning rate of the new model.

[0025] Another object of the present application is to provide an LSTM neural network-based development platform alarm data prediction method system, which can solve the problems of low model training efficiency, unstable model performance and lack of self-adjustment ability of existing artificial intelligence systems through automatic model design and optimization, and real-time model performance monitoring and adjustment.

[0026] An artificial intelligence prediction performance optimization system based on an LSTM neural network, characterized by comprising a data collection module, a model construction module, a model optimization module, and a model evaluation module; the data collection module is used to receive user-uploaded text data and automatically perform preprocessing by removing special characters and irrelevant words in the text through regular expressions; the model construction module is used to construct an adaptive LSTM model, train the model to improve the learning efficiency and performance of the model; the model optimization module is used to monitor and adjust the performance of the model in real time during the model training process, infer the problem causes and adjust the corresponding model parameters; the model evaluation module is used to evaluate the performance of the model on the validation set and automatically make decisions according to the evaluation results.

[0027] A computer device comprising a memory and a processor, wherein the memory stores a computer program, and wherein the processor implements the steps of the method described above when executing the computer program.

[0028] A computer readable storage medium, having stored thereon a computer program, wherein the computer program is executed by a processor to implement the steps of the method.

[0029] The application provides the LSTM neural network-based development platform alarm data prediction method, which improves the model training efficiency and prediction performance, effectively avoids the problems of model overfitting and underfitting, enhances the generalization ability of the model, automatically performs model verification and decision making through the application of the sliding threshold method, greatly reduces the complexity of model development and maintenance, provides clear model training feedback, and enables users to understand the real-time progress and performance of model training. BRIEF DESCRIPTION OF DRAWINGS

[0030] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings needed to be used in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor. Among them:

[0031] Figure 1 The application provides an LSTM neural network-based development platform alarm data prediction method for an embodiment of the application.

[0032] Figure 2 The application provides an LSTM neural network-based artificial intelligence efficiency improvement system for a second embodiment of the application. DETAILED DESCRIPTION

[0033] In order to make the above-mentioned purposes, features and advantages of the application more obvious and easy to understand, the specific embodiments of the application will be described in detail below with reference to the drawings. Obviously, the described embodiments are part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor should be within the protection scope of the application.

[0034] In the following description, many specific details are set forth in order to provide a thorough understanding of the application, but the application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the concept of the application, therefore the application is not limited to the specific embodiments disclosed below.

[0035] Secondly, the "one embodiment" or "an embodiment" referred to herein means a specific feature, structure, or characteristic under discussion. The appearance of "in one embodiment" at various places in the specification does not necessarily all refer to the same embodiment, nor is it necessary that every embodiment includes the same features or characteristics.

[0036] The application is described in detail in conjunction with the schematic drawings, and in the detailed description of the embodiments of the application, the sectional view of the device structure is partially enlarged without the general proportion for the convenience of illustration, and the schematic drawings are only examples, which should not limit the scope of protection of the application herein. In addition, three-dimensional spatial dimensions of length, width and depth should be included in actual production.

[0037] Meanwhile, in the description of the application, it should be noted that the terms "upper, lower, inner and outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the application. In addition, the terms "first, second or third" are only for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0038] Unless otherwise specifically defined and limited, the terms "mounting, connecting, connection" in the application should be broadly understood, for example: it can be fixed connection, detachable connection or integral connection; it can also be mechanical connection, electrical connection or direct connection, it can also be indirectly connected through an intermediate medium, or it can be the communication between two elements. For those skilled in the art, the specific meaning of the above terms in the application can be understood according to the specific circumstances.

[0039] Embodiment 1

[0040] Reference Figure 1 For an embodiment of the application, a development platform alarm data prediction method based on an LSTM neural network is provided, comprising:

[0041] The development platform monitoring system receives the text data uploaded by the user, automatically performs preprocessing, and constructs an adaptive LSTM model;

[0042] The automatic preprocessing includes: first, removing special characters and irrelevant words in the text through regular expressions, designing specific regular expressions to match and delete non-alphanumeric characters, removing common irrelevant words, simplifying the text content, and facilitating subsequent processing, using a deep learning method to identify and filter out junk information or invalid information in the preprocessing stage, deleting records containing missing values, using average, median statistical indicators to fill in missing values, using a deep learning-based autoencoder to predict missing values, and performing word segmentation based on the maximum matching method.

[0043] According to the distribution statistical information of the input data, the feature dimension size and the sample quantity, an adaptive LSTM model is constructed. First, statistical analysis is performed on the input data to obtain the distribution information of the data, such as mean, standard deviation, maximum value, minimum value, etc. Then, according to the feature dimension and the sample quantity, the required model complexity is estimated to avoid overfitting and underfitting.

[0044] The number of hidden layers is calculated using a heuristic algorithm based on an empirical formula of the input feature dimension and the sample quantity. In the model design stage, the depth of the model (i.e. the number of hidden layers) is determined. According to the input feature dimension and the sample quantity, an initial hidden layer number initial value is preset. As the model training proceeds, the number of hidden layers is dynamically adjusted according to the performance of the model on the validation set.

[0045] Adjusting the number of hidden layers includes increasing the number of hidden layers to increase the complexity of the model, so that it can fit more complex functions; and reducing the number of hidden layers to reduce the risk of overfitting.

[0046] After modifying the number of hidden layers, the model needs to be retrained. After retraining the model, it needs to be evaluated again to determine whether the performance of the model has improved after adjusting the number of hidden layers. By repeatedly adjusting the number of hidden layers and evaluating the effect, the optimal number of hidden layers can be found.

[0047] Automatically increasing the number of neurons in each layer, the number of neurons in each hidden layer is determined, starting from 10 and increasing by 10 each time until the accuracy rate of the validation set in the model improves by less than the accuracy rate set value in three consecutive training periods. In this embodiment, the accuracy rate set value is set to 0.1%. This method is a greedy strategy that increases the number of neurons to improve the complexity of the model and thus improve the performance of the model, but also avoids overcomplicating the model leading to overfitting.

[0048] The number of neurons determines the complexity and capacity of the model, i.e. the complexity and quantity of functions or patterns that the model can learn. If there are too few neurons, the model cannot fully learn the information in the data, leading to underfitting. If there are too many neurons, the model will overfit the training data, leading to overfitting and poor generalization to unseen data.

[0049] Therefore, automatically increasing the number of neurons in each layer is a preferred strategy, by starting with a small number of neurons and gradually increasing the number of neurons, observing how the performance of the model changes on the validation set, if increasing the number of neurons improves the accuracy of the validation set, continue to increase; if the accuracy improvement value is less than 0.1% in the last three training cycles, stop increasing, which is close to the optimal complexity of the model, and further increasing will lead to overfitting.

[0050] An adaptive learning rate adjustment method is used to automatically select the learning rate. The learning rate is an important hyperparameter that determines the magnitude of parameter updates during training. In this embodiment, the initial learning rate is set to 0.01. After each training cycle, if the accuracy of the model on the validation set does not improve, the learning rate is halved until the learning rate is less than the learning rate setting value. This strategy helps to quickly converge in the early stage, while ensuring the stability of the model in the later stage. In a preferred embodiment, the learning rate setting value is set to 0.0001.

[0051] The learning rate is a key hyperparameter that determines the step size of weight updates during training. If the learning rate is too high, the model will oscillate around the optimal solution and cannot converge. If the learning rate is too small, the convergence speed of the model will be slow and more training time will be needed, falling into a poor local optimum.

[0052] Therefore, automatically selecting the learning rate is another preferred strategy. The initial learning rate is set to 0.01, which is a preferred initial value. Then, after each training cycle, if the accuracy of the model on the validation set does not improve, the learning rate is halved. If the accuracy does not improve, the learning rate is too high, the model oscillates around the optimal solution and cannot converge. By reducing the learning rate, the model's training is more stable and can find a better solution. This process will continue until the learning rate is less than 0.0001, which is to prevent the learning rate from being reduced to almost zero, causing the model to almost stop learning.

[0053] Automatically selecting the activation function, the activation function is an important component of the neural network, which determines the output of the neuron. By iterating through the ReLU, tanh, and sigmoid activation functions, the activation function with the highest accuracy on the validation set is selected to find the most suitable activation function for the current task.

[0054] The adaptive LSTM model enters the model training stage and the model optimization stage, and the model performance is monitored and adjusted in real time; in the model training stage, the system monitors and displays the performance indicators of the training process in real time, and when the indicators are abnormal, the system starts the feedback adjustment process; the model training stage includes, in the model training stage, the system uses the optimizer Adam with adaptive learning rate to train the model, when the loss function value of the model decreases by less than the loss function set value for five consecutive training periods, the system automatically reduces the learning rate to half of the current value, in this embodiment, the loss function set value is set to 0.001.

[0055] The feedback adjustment process includes, in the process of model training and optimization, the system automatically monitors the performance indicators of the model, automatically adjusts the model parameters, learning rate or structure according to the changes of the performance indicators. The feedback adjustment process automatically infers the problem causes according to the performance of the model on the training and validation data, and makes corresponding adjustments according to the inference results, including but not limited to changing the learning rate, adjusting the model parameters, changing the model structure, etc., to improve the performance and accuracy of the model.

[0056] When the model enters the training stage, the optimizer Adam is first used to start training, during the training process, an initial learning rate is set, and the performance of the model on the validation set is evaluated at the end of each training period, the change of the loss function value of the model is observed, if the loss function value of the model decreases by less than 0.001 in five consecutive training periods, the system will automatically reduce the learning rate to half of the current value, if the performance of the model stops improving, it is judged that because the learning rate is too high, the model oscillates in the optimization process and cannot converge to a stable, optimal solution.

[0057] A lower limit value is set, when the learning rate reaches this lower limit, it will not be reduced any more, to prevent the learning rate from being reduced to a level where the model parameters cannot be effectively updated, to ensure that the training process can continue, if the performance of the model still cannot be improved significantly after five training periods, a stop mechanism is set to prevent overtraining of the model.

[0058] In this process, a graph of the loss function value and the training period is drawn, and the effect of different learning rates on the performance of the model is recorded, the performance of the model on the validation set at the end of each training period is recorded, and after the training is completed, the model with the best performance is selected for subsequent use.

[0059] After adjustment, enter the model optimization stage, apply the sliding threshold method to evaluate the accuracy of the model, adjust the parameters to improve the prediction performance of the model, and then input the text data uploaded by the user into the optimized adaptive LSTM model for alarm prediction.

[0060] The model optimization stage includes that the system collects the performance accuracy of the training process in real time during the model training process. If the verification accuracy does not improve by more than 1% for ten consecutive training cycles, the parameter adjustment stage is automatically entered. According to the performance index collected during the training process, the problem reason is automatically inferred, the corresponding model parameters are adjusted according to the inference result, and the training task is restarted using the new parameters.

[0061] If the verification accuracy does not improve by more than 1% for twenty consecutive training cycles, and at least one parameter adjustment has been performed during this period, the model structure adjustment stage is automatically entered. The number of hidden layers is adjusted, the activation function is changed, the structure of the model is changed, and the training task is restarted using the new model structure.

[0062] If the verification accuracy does not improve by more than 1% for thirty consecutive training cycles, and at least one model structure adjustment has been performed during this period, the current model is considered to have reached the performance limit, a conservative strategy is adopted, the current model is temporarily saved, and the model is used as a benchmark for evaluation using a sliding threshold method.

[0063] If the verification accuracy does not improve by more than 1% for ten consecutive training cycles, the parameter adjustment stage is automatically entered: during the process of training a neural network, if the model's verification accuracy does not improve by more than 1% for ten consecutive training cycles, it means that the model has been unable to effectively obtain more information from the data under the current parameters, and has fallen into a local optimum. The system will automatically enter the parameter adjustment stage by adjusting the learning rate, regularization parameter, etc. to try to find a new parameter setting to enable the model to continue to improve its performance. The main purpose of this stage is to adjust the parameters of the model without changing the structure of the model.

[0064] If the verification accuracy does not improve by more than 1% for twenty consecutive training cycles, and at least one parameter adjustment has been performed during this period, the model structure adjustment stage is automatically entered: if the model's performance has not been significantly improved after the parameter adjustment in the previous stage, it means that only adjusting the parameters cannot solve the problem, and the structure of the model is not suitable for the current data or task. The system will automatically enter the model structure adjustment stage to try to change the structure of the model by adjusting the number of hidden layers, changing the activation function, etc. in order to expect the model to obtain better performance under the new structure. The purpose of this stage is to change the complexity of the model to find a model structure that is more suitable for the current data and task.

[0065] If the validation accuracy does not improve by more than 1% in the last 30 consecutive training cycles and at least one model structure adjustment has been made during this period, the current model is considered to have reached its performance limit, and the model that has been optimized to a satisfactory level should not be abandoned blindly. In order to avoid wasting more time and computing resources on a model that cannot meet the requirements, the system will consider all factors, including data preprocessing, model structure selection, parameter setting, etc., in order to find a model that better meets the current task requirements. During the retraining process, the system will return to the starting point of training and go through the parameter adjustment and model structure adjustment stages again until a satisfactory model is found.

[0066] In order to balance the improvement speed of the model and the efficiency of the training, the threshold value is set to not improve by more than 1%. If the threshold value is set higher than the threshold value, the model needs to be trained for a large number of training cycles to reach this improvement goal, which will greatly increase the training time and computing resource consumption. If the threshold value is set lower than the threshold value, although it can enter the next optimization stage faster, it will lead to frequent invalid optimization of the model, reducing the efficiency of the training. Therefore, the threshold value of 1% is a balanced choice between the two extremes.

[0067] The sliding threshold method includes the system generating a model explanation report, applying the sliding threshold method to decode the output of the model on the validation set, comparing the decoding result with the true result, and calculating the accuracy ;

[0068] If the model's performance on the validation set is , it is evaluated as excellent, if , it is evaluated as good, if , it is evaluated as average, and if , it is evaluated as poor.

[0069] Each level represents a performance state of the model. Excellent represents that the model has reached a high level of accuracy and can be directly put into use. Good represents that the performance of the model is close to excellent and can be further improved through fine-tuning. Average represents that the performance of the model still has room for improvement and may need to be adjusted to a greater extent. Poor indicates that the performance of the model is not good and may need to be abandoned and replaced with a different parameter and structure for training. According to the current performance of the model, determine which optimization strategy should be taken, both to improve the efficiency of optimization and to effectively avoid overfitting and underfitting.

[0070] The model accuracy evaluation includes evaluating the accuracy of the model on the test set; if the accuracy of the model on the test set is excellent, the current training task is automatically stopped, and the model is deployed to the production environment; if the accuracy evaluation is good, the parameter fine-tuning stage is started, the parameters of the model are adjusted through cross-validation on the test set, after fine-tuning, the accuracy of the model is evaluated again to determine whether to deploy to the production environment; if the accuracy evaluation is general, the model structure adjustment stage is started, the number of hidden layers is adjusted, and the activation function is changed to change the structure of the model; after the structure adjustment is completed, the system will evaluate the accuracy of the model again and decide whether to enter the next stage; if the accuracy evaluation is poor, the current model is abandoned, a new model is trained, and different parameters and structures from the abandoned model are used, and the test set is used for performance verification to determine the parameters and learning rate of the new model.

[0071] To achieve the most effective resource utilization and the highest training efficiency, when the model performance is excellent, additional resources are not needed to optimize, and it can be directly used; when the model performance is good, the fine-tuning method is used to avoid overfitting of the model in the optimization process and improve the accuracy of the model; when the model performance is general, the model structure is changed to find a more suitable model structure for the data and improve the performance of the model; when the model performance is poor, the current model is abandoned, and new parameters and structures are tried to avoid wasting resources on inefficient models. The dynamic evaluation and classification of model performance and the different processing according to the classification results improve the efficiency of model optimization and fully utilize the computing resources.

[0072] Embodiment 2

[0073] Reference Figure 2 For an embodiment of the present application, an artificial intelligence prediction performance optimization system based on an LSTM neural network is provided, comprising:

[0074] A data collection module, a model construction module, a model optimization module, and a model evaluation module.

[0075] The data collection module is used to receive user-uploaded text data and automatically perform preprocessing by removing special characters and irrelevant words in the text through regular expressions.

[0076] The model construction module is used to construct an adaptive LSTM model to train the model and improve the learning efficiency and performance of the model.

[0077] The model optimization module is used to monitor and adjust the performance of the model in real time during the model training process, infer the problem causes and adjust the corresponding model parameters.

[0078] The model evaluation module is used to evaluate the performance of the model on the validation set and automatically make decisions based on the evaluation results.

[0079] Example 3

[0080] One embodiment of the present application, which is different from the first two embodiments, is as follows:

[0081] The functions described can be implemented in software, firmware, hardware, or any combination thereof. If implemented in software and as an independent application, it can be stored in a one or more of computer-readable storage media described above. The functions described can be stored as one or more instructions and / or data on non-transitory computer-readable storage medium and executed by one or more processors of the computer device. The general structure, logic, and / or circuitry of the computer device can be implemented in hardware, software, firmware, or any combination thereof. The computer device can be a personal computer, server, network device, or any other computing device capable of executing instructions.

[0082] Logic and / or steps represented in flow diagrams or otherwise described herein, for example, can be embodied in computer-readable instructions, orders / writes of operations, or in any other training data, configuration data, or any other data that can be used by or in conjunction with an instruction execution system, apparatus, or device, such as a computer-based system, processor, or other system that can fetch instructions from the instruction execution system, apparatus, or device and execute instructions. For purposes of this specification, a "computer-readable medium" can be any apparatus that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device.

[0083] More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection (electronic) having one or more wires, a portable computer diskette (magnetic), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber (optical), and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can also be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, for example, via an optical scanner, then compiled, interpreted, or otherwise processed, and stored in a computer memory in order to be executed.

[0084] It should be understood that portions of the application can be implemented in hardware, software, firmware, or combinations thereof. In the embodiments described above, the various steps or methods can be implemented, in part, or in whole, by software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, and in another embodiment, any of the following techniques, which are well known in the art, can be used to implement the application: a hybrid of the techniques mentioned above; a combination of one or more of the techniques mentioned above; a combination of one or more of the techniques mentioned above with one or more other techniques not mentioned above; and / or any other techniques for implementing the application.

[0085] Example 4

[0086] For an embodiment of the application, a LSTM neural network-based development platform alarm data prediction method is provided. In order to verify the beneficial effects of the application, economic benefit calculation and simulation experiments are used for scientific demonstration.

[0087] MATLAB and CloudSim are used to evaluate the algorithm. The simulation has been run in an environment with an Intel processor and 12 GB RAM. The operating system used is 64-bit Windows 11 Ultimate. The point system is simulated using the MATLAB programming language, connecting records, and building data distribution.

[0088] Two parallel experiments are conducted: one uses the application's optimization method, and the other uses the traditional LSTM neural network training method. Both experiments use the same batch of preprocessed text data set for model training, which contains a certain amount of training data and validation data. The evaluation indicators of the two methods are tracked and compared respectively.

[0089] The evaluation indicators include: training time: the time required from the start of model training to reaching a certain accuracy threshold; validation accuracy: the accuracy of the model on the validation data set; training accuracy: the accuracy of the model on the training data set; model adjustment times: the number of times the model parameters or structure need to be adjusted in the process of reaching a certain accuracy threshold. The experimental results are shown in Tables 1 and 2.

[0090] Table 1 Comparison of experimental results

[0091] Training time (hours) Validation accuracy (%) Training accuracy (%) Model adjustment times Our invention method 8 85 90 3 Traditional method 12 80 85 5

[0092] Table 2 Comparison of validation accuracy under different thresholds

[0093] Average training period Validation accuracy Adjustment times 0.1% 10 88.70% 25 0.5% 12 90.20% 20 1% 15 94.50% 15 2% 18 91.20% 10 5% 20 89.00% 5

[0094] As shown in Table 1, the application's method is superior to the traditional LSTM neural network training method in all evaluation indicators.

[0095] The training of the model can be completed in 8 hours by the optimization method of the application, while the traditional method needs 12 hours. The optimization method has a significant improvement in training efficiency, and the accuracy on the validation dataset reaches 85%, while the accuracy of the traditional method is 80%. The prediction ability of the application is improved, and the accuracy on the training dataset reaches 90%, while the accuracy of the traditional method is 85%. The fitting degree of the application on the training data is better. The application only needs to be adjusted 3 times, while the traditional method needs to be adjusted 5 times. The application has more advantages in model parameter selection and structure design.

[0096] Through the above comparison, it can be clearly seen that the optimization method of the application has obvious advantages in improving the model training efficiency and prediction performance, and at the same time reduces the manual intervention demand in the model training process, and realizes a higher degree of automation.

[0097] As shown in Table 2, for a threshold value lower than 1%, although the training period is shorter, in deep learning, the model cannot fully learn and extract information in the data in each training period. If the threshold value is set too low, the model is adjusted or the structure is adjusted frequently before the model is fully learned, and too frequent adjustment will make the training process more complex and require more computing resources. Although the training period is shorter, the whole training process may become longer due to frequent adjustment, and the verification accuracy may not be as good as when the threshold value is set to 1%.

[0098] For a threshold value higher than 1%, although the number of adjustments is less, it will make the model enter the next stage too early, resulting in the model not fully learning and extracting information in the data. If the threshold value is set too high, the model will enter the next stage too early before it has fully learned the data, hindering the learning process of the model and affecting the final performance. In the above table, although the threshold value of 5% has the least number of adjustments, its verification accuracy is the lowest.

[0099] In summary, the threshold value of 1% finds a balance between ensuring that the model fully learns and maintaining training efficiency, so the application finally determines that the threshold value is 1%.

[0100] It should be noted that the above embodiments are only used to illustrate the technical solutions of the application and not to limit it. Although the application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the application can be modified or replaced by equivalents without departing from the spirit and scope of the application, and they should be included in the scope of the claims of the application.

Claims

1. A method for predicting alarm data on a development platform based on an LSTM neural network, characterized in that, include: The development platform monitoring system receives text data uploaded by users, automatically performs preprocessing, and builds an adaptive LSTM model; The adaptive LSTM model enters the model training and model optimization phases, and the model performance is monitored and adjusted in real time. During the model training phase, the system monitors and displays the performance metrics of the training process in real time. If any metrics are abnormal, the system initiates a feedback and adjustment process. After adjustment, the model is optimized by applying the sliding threshold method to evaluate the model accuracy and adjusting parameters to improve the model's predictive performance. User-uploaded text data is input into an optimized adaptive LSTM model for alarm prediction; The construction of the adaptive LSTM model includes constructing an adaptive LSTM model based on the distribution statistics of the input data, the feature dimension size, and the number of samples. The number of hidden layers is calculated using a heuristic algorithm based on empirical formulas for the input feature dimension and the number of samples. The number of neurons in each layer is automatically increased until the accuracy improvement of the model on the validation set is less than the accuracy set value over three consecutive training epochs. An adaptive learning rate adjustment method is used to automatically select the learning rate. An initial learning rate is set. After each training cycle, if the accuracy of the model on the validation set does not improve, the learning rate is halved until the learning rate is less than the set value. The activation function is automatically selected by iterating through ReLU, tanh, and sigmoid activation functions and finally selecting the activation function with the highest accuracy of the model on the validation set. The model training phase includes the following: during the model training phase, the system uses the Adam optimizer with an adaptive learning rate to train the model. When the loss function value of the model decreases by less than the set value of the loss function for five consecutive training cycles, the system automatically reduces the learning rate to half of the current value. The model optimization phase includes the following: during model training, the system will collect the performance accuracy of the training process in real time. If the verification accuracy does not improve by more than 1% for ten consecutive training cycles, it will automatically enter the parameter adjustment phase. Based on the performance indicators collected during training, the system will automatically infer the cause of the problem, adjust the corresponding model parameters according to the inference results, and restart the training task with the new parameters. If the validation accuracy does not improve by more than 1% within 20 consecutive training cycles, and at least one parameter adjustment has been made during this period, the system will automatically enter the model structure adjustment phase, which will adjust the number of hidden layers, change the activation function, change the model structure, and restart the training task with the new model structure. If the validation accuracy does not improve by more than 1% within 30 consecutive training cycles, and at least one model structure adjustment has been made during this period, the current model is considered to have reached its performance limit. A conservative strategy is adopted, the current model is temporarily saved, and this model is used as a benchmark for evaluation using the sliding threshold method.

2. The alarm data prediction method for a development platform based on LSTM neural networks as described in claim 1, characterized in that: The automatic preprocessing includes data cleaning, handling missing values, text segmentation, and conversion to word vectors. In the data cleaning stage, special characters and irrelevant words in the text are removed using regular expressions. After the preprocessing is completed, the preprocessing results are displayed through the user interface.

3. The alarm data prediction method for a development platform based on LSTM neural networks as described in claim 2, characterized in that: The sliding threshold method includes: generating a model interpretation report; applying the sliding threshold method to decode the model's output on the validation set; comparing the decoded result with the true result; and calculating the accuracy. ; If the model is on the validation set If rated as excellent, The evaluation is good. The evaluation is average. The rating was poor.

4. The alarm data prediction method for a development platform based on LSTM neural networks as described in claim 3, characterized in that: The evaluation of model accuracy includes evaluating the model's accuracy on the test set; If the model achieves excellent accuracy on the test set, automatically stop the current training task and deploy the model to the production environment; If the accuracy is rated as good, the parameter fine-tuning phase is initiated. The model parameters are adjusted by cross-validation on the test set. After fine-tuning, the model's accuracy is evaluated again to determine whether to deploy it to the production environment. If the accuracy is rated as average, the model structure adjustment phase will be initiated. The number of hidden layers will be adjusted, and the activation function will be changed to alter the model's structure. After the structure adjustment is completed, the system will re-evaluate the model's accuracy and decide whether to proceed to the next phase. If the accuracy evaluation is poor, abandon the current model, train a new model with different parameters and structure than the abandoned model, use a test set to verify the performance, and determine the parameters and learning rate of the new model.

5. A system employing the alarm data prediction method based on an LSTM neural network for a development platform as described in any one of claims 1 to 4, characterized in that, include: Data collection module, model building module, model optimization module, model evaluation module; The data collection module is used to receive text data uploaded by users and automatically perform preprocessing, removing special characters and irrelevant words from the text using regular expressions; The model building module is used to build an adaptive LSTM model and train the model to improve its learning efficiency and performance. The model optimization module is used to monitor and adjust model performance in real time during model training, infer the cause of the problem, and adjust the corresponding model parameters. The model evaluation module is used to evaluate the model's performance on the validation set and make automatic decisions based on the evaluation results.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the alarm data prediction method for the development platform based on the LSTM neural network as described in any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the alarm data prediction method for the development platform based on the LSTM neural network as described in any one of claims 1 to 4.

Images