A device predictive maintenance method based on swarm intelligence algorithm

Abstract

The application belongs to the field of intelligent maintenance of equipment, and discloses a kind of equipment predictive maintenance method based on swarm intelligence algorithm, and is suitable for mechanical equipment doing periodic motion.It includes: determining the data amount of equipment operation time series data in a cycle in training data set;According to the data amount in a cycle, a time convolution network prediction model based on periodic signal is established;Take the training sample generated according to the equipment operation time series data in the training data set, input the training sample into the time convolution network prediction model based on periodic signal, and use the tuna algorithm to optimize and solve the hyperparameters of the time convolution network prediction model based on periodic signal, until the maximum iteration number is reached to fix the hyperparameters.According to the periodic characteristics of equipment time series data, a time convolution network prediction model based on periodic signal is established.The intelligent swarm optimization algorithm is used to optimize and solve the hyperparameters in the time convolution network prediction model based on periodic signal, so that the performance of the model prediction is improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent equipment maintenance, specifically relating to a predictive maintenance method for equipment based on a swarm intelligence algorithm. First, a temporal convolutional network prediction model is established based on the time-series periodic signals generated when certain equipment performs periodic movements. Then, the tuna algorithm is used to generate hyperparameters in this model. The generated parameters are substituted into the prediction model to calculate and solve for the target loss function. Based on the value of the loss function, the optimal individual is determined. The process is iterated based on the optimal individual until the optimal hyperparameters and their corresponding prediction results are found. Background Technology

[0002] Currently, the working environment of some mechanical equipment is becoming increasingly complex. Harsh working conditions such as high temperature, high pressure, and continuous operation accelerate the wear and aging of components, affecting the overall performance and health of the equipment. More seriously, they can cause huge economic losses and even casualties. Therefore, predictive maintenance technology to prevent equipment failure has emerged.

[0003] Neural networks are a commonly used method in predictive maintenance. They can automatically adjust weights and parameters through learning and training, handle nonlinear relationships, and automatically extract features, thus adapting to complex patterns and data structures.

[0004] While neural networks have shown relatively good results in handling some prediction problems, they also have some limitations. The most prominent is the long-term dependency problem with longer time-series data, where information gradually weakens or is lost as it is passed through the sequence. To address this, Long Short-Term Memory (LSTM) recurrent neural networks were developed. They possess a memory function for time-series data and solve the gradient decay problem caused by gradual reduction during gradient backpropagation. They use gating and memory cells to learn to discard past information and extract current input information. However, in practical applications, they also have some drawbacks and limitations, such as complex model structure and computationally expensive processing time.

[0005] To address the issues of long-term dependencies and complex model structures in time series data, some researchers have proposed temporal convolutional networks (TCNNs). These networks preserve the characteristics of time-series data while simplifying model parameters, demonstrating good performance in processing such data. However, the hyperparameters used during training, such as the learning rate, the number of hidden neurons, and the number of training iterations, are mostly determined through manual, empirical tuning. This method carries significant errors, preventing the predictive deep learning model from fully realizing its potential. Summary of the Invention

[0006] This invention addresses the characteristic of certain devices exhibiting periodic motion by establishing a temporal convolutional network prediction model based on periodic signals. An intelligent swarm optimization algorithm is employed to optimize the hyperparameters in this model, thereby improving its prediction performance. The intelligent swarm optimization algorithm used is the Tuna algorithm.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] A predictive maintenance method for equipment based on swarm intelligence algorithms, applicable to mechanical equipment undergoing periodic motion, comprising:

[0009] Determine the amount of data in one cycle of the device operation time series data in the training dataset;

[0010] Based on the amount of data within a period, establish a temporal convolutional network prediction model based on periodic signals;

[0011] Training samples generated from the device operation time series data in the training dataset are taken sequentially. The training samples are input into the temporal convolutional network prediction model based on periodic signals. The tuna algorithm is used to optimize and solve the hyperparameters in the temporal convolutional network prediction model based on periodic signals until the maximum number of iterations is reached. The hyperparameters are then fixed and used for the training and application of the temporal convolutional network prediction model based on periodic signals.

[0012] Several alternative methods are provided below, but they are not intended as additional limitations on the overall solution above. They are merely further additions or optimizations. Provided there are no technical or logical contradictions, each alternative method can be combined individually with respect to the overall solution above, or multiple alternative methods can be combined with each other.

[0013] Preferably, determining the amount of data within one period of the device operation time-series data in the training dataset includes:

[0014] Obtain device operation time-series data to form a training dataset of size X. (The sentence fragment about the training dataset appears to be incomplete and lacks context. It has been left as is.) Randomly select a data point x1 from the time-series data of each device operation, and then search for data points x1 in the training dataset that are equal to data point x1, starting from data point x1. n 1 <n≤X;

[0015] Compare m data points with a random step size less than n starting from data point x1 with data point x1 and data point x1 starting from data point x1. n If the values ​​of m data points with the same random step size starting from the same point are all equal, then take the values ​​from data point x1 to data point x2. n-1Given data within a period, determine the data quantity T = n-1 within that period; if they are not equal, continue searching sequentially for data x that is equal to data x1. n .

[0016] Preferably, the temporal convolutional network prediction model based on periodic signals includes an input layer, a hidden layer, and an output layer. The hidden layer uses the most recently input training sample as the origin, and determines the data belonging to the most recently input period by working backwards along the input time according to the amount of data within one period. The remaining historical training samples are considered as data outside the most recently input period. When the hidden layer performs convolution operations on the data belonging to the most recently input period, the kernel size is kernelsize=2. When performing convolution operations on the data outside the most recently input period, the kernel size is kernelsize=T, where T is the amount of data within one period. The output layer performs convolution operations on all the data output by the hidden layer to obtain the final prediction result.

[0017] Ideally, the input and output times of all hidden layers should be the same.

[0018] Preferably, the hidden layer and the output layer are performed using dilated causal convolution.

[0019] Preferably, the generation of the training samples includes: using a sliding window to process the device operation time-series data in the training dataset and outputting the training samples.

[0020] Preferably, the step of using the tuna algorithm to optimize and solve the hyperparameters in the temporal convolutional network prediction model based on periodic signals, and then fixing the hyperparameters after reaching the maximum number of iterations, includes:

[0021] Initialize a tuna population, wherein each individual in the tuna population represents a combination of hyperparameters of a temporal convolutional network prediction model based on periodic signals;

[0022] Each individual is fed into a temporal convolutional network prediction model based on periodic signals, and the corresponding predicted value is obtained based on the current input training samples.

[0023] Calculate the loss function between the predicted value and the actual value corresponding to the training sample to determine the optimal individual in the tuna population;

[0024] Determine if the maximum number of iterations has been reached. If so, output the optimal individual as the hyperparameter of the temporal convolutional network prediction model based on periodic signals. Otherwise, update the position of individuals in the tuna population according to the iteration rule, and substitute each individual with the updated position back into the temporal convolutional network prediction model based on periodic signals to perform prediction iterations for the new training samples.

[0025] Preferably, the initialization of the tuna population includes:

[0026]

[0027] In the formula, i represents the i-th individual, and NU represents the total number of individuals in the tuna population. Let represent the initial position of the i-th individual, ub represent the upper bound of the search space, and lb represent the lower bound of the search space.

[0028] Preferably, the loss function is a mean squared error function, and the loss function for calculating the difference between the predicted value and the actual value corresponding to the training sample includes:

[0029]

[0030] In the formula, MSE is the mean squared error, T is the amount of data in one period, and y k For the k-th true value, This is the k-th predicted value.

[0031] Preferably, the iteration rule is as follows:

[0032] If the current iteration number t is less than or equal to Random iteration is used, and its iteration expression is as follows:

[0033]

[0034] In the formula, This represents the position of the i-th individual generated in the (t+1)-th iteration, where T is the amount of data in one period. This represents the reference position randomly generated in the t-th iteration. This represents the position of the first individual generated in the t-th iteration, and NU represents the total number of individuals in the tuna population. This represents the position of the i-th individual generated in the t-th iteration. Let represent the position of the (i-1)th individual generated in the t-th iteration, and β be the step size, calculated as follows:

[0035]

[0036] In the formula, b is a random number uniformly distributed between 0 and 1;

[0037] The current iteration number t is greater than The optimal value iteration is used, and its iteration expression is as follows:

[0038]

[0039] In the formula, It is the optimal individual in the t-th iteration of the tuna population.

[0040] Hyperparameter settings play a crucial role in the accuracy of deep learning prediction models. Most deep learning models rely on manual, empirical hyperparameter tuning, which is prone to significant errors and prevents the model from fully realizing its predictive potential. Furthermore, the tedious tuning process is a major headache for practitioners. This invention proposes using an intelligent swarm optimization algorithm to optimize deep learning hyperparameters, thereby maximizing the performance of the deep learning model while simultaneously solving the problems associated with tedious manual tuning. Attached Figure Description

[0041] Figure 1 This is a flowchart of a predictive maintenance method for equipment based on swarm intelligence algorithms according to the present invention;

[0042] Figure 2 This is a schematic diagram of the temporal convolutional network prediction model based on periodic signals according to the present invention;

[0043] Figure 3 This is a schematic diagram of the data sliding window principle of the present invention;

[0044] Figure 4 This is a flowchart of the tuna optimization algorithm of the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention.

[0047] Taking a loom as an example, the movement of a loom can be broken down into shedding motion, weft insertion motion, beating-up motion, warp feed motion, and take-up motion. The working process of a loom is also composed of these five basic motion cycles repeating cyclically. Therefore, the time-series data generated during the operation of a loom has a periodic characteristic. Establishing a model based on the characteristics of the loom's periodic motion can simplify the solution process and enable predictive maintenance of the equipment.

[0048] To overcome the shortcomings of hyperparameter adjustment in predictive maintenance models for looms, this embodiment proposes a predictive maintenance method for equipment based on swarm intelligence algorithms, such as... Figure 1 As shown in the figure, it specifically includes the following steps:

[0049] Step 1: Prepare the historical operation information of the loom, including but not limited to time-series data such as tension, encoder angle, and main shaft speed. In this embodiment, there is no limit to the collected historical operation information. For data within one cycle and data of more than one cycle, the model uses different convolution kernels for calculation. There is also no limit to the output result.

[0050] Step 2: Perform preprocessing of data normalization to make the features of the data more obvious.

[0051] For numerical data such as tension and main shaft speed, the normalization processing method used is max-min, and the calculation formula is as follows:

[0052]

[0053] where x′ is the data after normalization of data x, x is the value in the original data, x min is the minimum value of the original data sample, x max is the maximum value of the original data sample.

[0054] For values in angular measure (360° system) such as encoder angle, the processing method is: convert it to the radian system to obtain a radian value, and use the sine value or cosine value of this radian value as the input data.

[0055] Step 3: Divide the data set. The data set is divided into a training data set and a test data set in a ratio of 7:3. First, use the data in the training data set to train the model, and then use the data in the test data set to test the model until the preset conditions are met.

[0056] Step 4: Determine the cycle of the loom operation time-series data in the training data set and clarify the amount of data T within one cycle.

[0057] Assume that the amount of data in the training data set is X. To avoid the influence brought by data fluctuations, randomly select a data x1 from the first loom operation time-series data in the training data set, and sequentially search for the data x equal to the data x1 in the training data set n , 1 < n ≤ X. To ensure that the data from x1 to the data x n-1 is one cycle of data, compare whether the values of m data with a random step size less than n starting from the data x1 and m data with the same random step size starting from the data x n are equal one by one. If they are not equal, continue to search for the next data equal to the data x1. If they are equal, take the data from x1 to the data x n-1Given data within a period, determine the amount of data within that period as T = n-1.

[0058] To improve accuracy, the above operation can be repeated multiple times to obtain multiple selected data volumes T for a single period. The T with the most frequent values ​​is then used as the final data volume for a single period.

[0059] Step 5: Based on the amount of data within a period, establish a temporal convolutional network prediction model based on the periodic signal.

[0060] Modeling time series data essentially involves establishing a mapping relationship. This involves taking the input time series data x1, x2…x… n Mapped to the predicted outputs y1, y2…y n In the past, when discussing time series data, the neural network models that came to mind were recurrent neural networks (RNNs) and their variants, such as long short-term memory recurrent recurrent neural networks (LSRNNs) and gated recurrent units (GRUs). However, much work has shown that it is difficult to find new models within the RNN framework that can outperform LSRNNs in many tasks. But by stepping outside the RNN framework, temporal convolutional networks derived from convolutional neural networks (CNNs) can easily achieve performance exceeding that of LSRNNs and GRUs in many tasks.

[0061] In fact, temporal convolutional networks do not refer to a specific model, but rather to a framework similar to recurrent neural networks. The design of temporal convolutional networks is ingenious. Unlike convolutional long short-term memory recurrent neural networks (LSTM recurrent neural networks), which use convolution operations to process image information (LSTM recurrent neural networks operate on the input image at only one time step), temporal convolutional networks directly utilize the powerful properties of convolution to extract features across time steps.

[0062] Temporal convolutional networks process and compute time-series data through convolution. When the kernel size is greater than 1, the output size is smaller than the input size, thus obtaining a new sequence that retains the features of the original time series while being shorter. Repeating the convolution operation yields a precise prediction result. It's important to note that in traditional temporal convolutional networks, the kernel size of each hidden layer is the same. It has the same processing power for all data in the sequence, and all data have the same weight in the result.

[0063] This embodiment improves upon the traditional temporal convolutional network to address the characteristics of time-series periodic data. For example... Figure 2As shown, in the hidden layer, the most recently input training sample is used as the origin. Data belonging to the most recently input period is determined by tracing back along the input time direction according to the amount of data within one period. The remaining historical training samples are considered data outside the most recently input period. When performing convolution operations on data belonging to the most recently input period, the kernel size is 2. If the amount of data T within one period is odd, zeros are padded to the end of the data within the most recently input period. When performing convolution operations on data outside the most recently input period, the kernel size is T. If the data at the beginning of the data outside the most recently input period is less than one period, zeros are padded to the beginning. For the first time step, there is no historical information; the temporal convolutional network considers its historical data to be all zeros.

[0064] This embodiment accelerates convergence while preserving data characteristics as much as possible by using different convolution kernels. In the output layer, convolution operations are performed on all output data from the hidden layer to obtain the final output result. Since the hidden layer result is obtained by convolving the input data, historical information is preserved. Simultaneously, a greater weight is allocated to data within the cycle, increasing the importance of recent loom operation information to the prediction results and thus improving the reliability of the prediction results to some extent.

[0065] To improve data utilization and reduce the impact of extreme data on the model when inputting data, this embodiment uses a sliding window to process the data. The schematic diagram of the sliding window is shown below. Figure 3 As shown. For determining the data belonging to the latest period by taking the latest input training sample as the origin and proceeding backwards along the input time according to the amount of data within a period, when the input training sample is data from a single moment, then the historical data from T moments prior to the current moment is taken as the data for the latest period; when the input training sample is a matrix data containing multiple moments output by a sliding window, then the matrix data from the current moment is taken as the origin, and the matrix data from T moments prior to the current moment is taken as the data for the latest period.

[0066] To achieve similar functionality to recurrent neural networks, temporal convolutional networks need to address two issues: (1) How can temporal convolutional networks, like recurrent neural networks, output time steps of the same length as the input time step, or in other words, how can each input time step have a corresponding output? (2) How can we ensure that historical data is not missed?

[0067] To address the first issue, this embodiment uses zero-padding to ensure that the input and output time steps of each hidden layer in the temporal convolutional network are identical, maintaining the same number of time steps. Specifically, regardless of the kernel size and dilation, if the input of the first hidden layer has n time steps, its output will also have n time steps. Similarly, the input and output time step lengths of the second, third, and subsequent hidden layers are all n.

[0068] To address the second problem, conventional temporal convolutional networks utilize causal convolution, where causality means that the output data y at time t is consistent with the output data at time t. t Its input can only be time t and times before t, i.e., x0, x1, ..., xt. t However, this presents a problem: when the time series is very long, the convolution kernel size is limited, and the network depth becomes unacceptable when the output needs to consider the values ​​of the entire sequence. To address this issue, this embodiment introduces dilated causal convolution, which still results in causality, but with the introduction of a dilation factor. After calculating a certain value, its output is returned to the network as input, effectively solving the problem of excessive network depth.

[0069] It is important to note that this invention only describes the improvements to the temporal convolutional network. In practical applications, convolutional layers, fully connected layers, pooling layers, etc., can be added to obtain better data characteristics. Specifically, in addition to the above operations, in a concrete implementation, after the temporal convolutional sequence, two convolutional layers are used to extract features in the spindle speed and tension dimensions, respectively. A ReLU activation function is used after each convolutional layer to increase the network's non-linear fitting ability. Residual connections are used between convolutional layers to enhance the network's robustness. Finally, a fully connected layer is used to output the prediction result.

[0070] Residual connections can be added to the network to aid gradient propagation and avoid the vanishing gradient problem. After convolutional operations, activation functions can be used to introduce non-linear characteristics. Pooling operations can also be used to reduce feature dimensionality and further capture key patterns. Users can add these features according to their needs and data characteristics.

[0071] Step 6: Use the tuna algorithm to optimize and solve the hyperparameters in the temporal convolutional network prediction model based on periodic signals.

[0072] Hyperparameters are the framework parameters in machine learning models. The setting of hyperparameters plays a crucial role in the accuracy of deep learning prediction models. Most deep learning models use manual, empirical hyperparameter tuning, which has significant errors and prevents the predictive deep learning model from fully realizing its potential. This invention proposes a periodic temporal convolutional network model based on the tuna optimization algorithm to predict short-term potential failures. The algorithm flowchart is shown below. Figure 4As shown.

[0073] Step 6.1: Initialize the tuna population.

[0074] First, the parameters of the tuna optimization algorithm are set, where the tuna population size is NU = 20. Assume the optimization hyperparameters are: number of hidden layer neurons nh ∈ [1, 50], learning rate Lr ∈ [0.01, 0.1], and dropout rate rd ∈ [0, 1]. Based on the optimization hyperparameters, the dimension of the tuna algorithm is determined to be d = 3. Assuming the data volume in one cycle is T = 100, the maximum number of iterations is set to k = T = 100.

[0075] Initialize a tuna population, where each individual represents a different combination of hyperparameters. To avoid over-concentration of individuals, ensure they are evenly distributed throughout the search space. The population generation expression is as follows:

[0076] In the formula, i represents the i-th individual, and NU represents the total number of individuals in the tuna population. Let represent the initial position of the i-th individual, ub represent the upper bound of the search space, and lb represent the lower bound of the search space.

[0077] Step 6.2: Substitute into the model to solve for the optimal individual.

[0078] Each individual corresponds to a specific hyperparameter value. These values ​​are substituted into a temporal convolutional network prediction model based on periodic signals to calculate the predicted value for the current input training sample. The loss function for each individual is then calculated separately. This embodiment uses... The mean squared error between the predicted and actual values ​​is used as the loss function to evaluate the performance of each parameter. The general expression is:

[0079]

[0080] In the formula, N * y is the number of training samples. i For the true value, This is the predicted value. Since the training samples used in this invention are periodic time-series data, and the data differences between each week are not significant, data from one week is used to replace the entire number of training samples, reducing training time. The modified expression for calculating the average error in this embodiment is as follows:

[0081]

[0082] The individual with the smallest average error among all individuals is the optimal individual.

[0083] Step 6.3: Determine if the maximum number of iterations has been reached. If so, output the optimal individual as the hyperparameter of the temporal convolutional network prediction model based on periodic signals and its corresponding predicted value; otherwise, update the position of individuals in the tuna population according to the iteration rule, and substitute each individual with the updated position into the temporal convolutional network prediction model based on periodic signals again to perform prediction iteration for the new training samples, repeating the process of solving, calculating the loss function, and judging.

[0084] The iterative process of tuna schooling is essentially a foraging process. The optimal solution found in the previous step is the tuna individual with the most abundant food. Other tuna will then move towards this food-rich individual in search of more food. Tuna exhibit specific school characteristics in their movement; they chase prey by forming tight spirals. Besides chasing prey, tuna in a school also exchange information. Each tuna follows the one in front, thus sharing information among adjacent tuna. Based on these principles, the mathematical formula for a conventional foraging strategy is as follows:

[0085]

[0086] In the formula, This represents the position of the i-th individual generated in the (t+1)-th iteration. Let α1 and α2 be the optimal individual in the t-th iteration of the tuna population, and β be the weighting coefficients that control the tendency of individuals to move towards the optimal individual and the previous individual.

[0087] The above foraging strategy has a fatal flaw: blindly following the best individual in the population to forage is detrimental to the group's foraging ability when even the best individual cannot find food. Therefore, a random coordinate is typically generated in the search space as a reference point. This allows each individual to explore a wider area and gives the tuna algorithm global exploration capabilities. The specific mathematical model is described below:

[0088]

[0089] in, This is a reference point randomly generated during the t-th iteration to guide the fish school. The expressions for calculating α1, α2, and β are as follows:

[0090]

[0091]

[0092] β=e b·l ·cos(2πb)

[0093] Where 'a' is a constant used to determine the extent to which the tuna follows the best individual and the previous individual in the initial stage; here, we take 'a' as the constant. b is a random number uniformly distributed between 0 and 1. Process parameters t represents the current iteration number, t max This represents the maximum number of iterations. To make the number of training iterations easier to control, we take t here. max =T. A randomly generated rand reference point allows the fish to move randomly within the optimization space in search of food.

[0094] To improve the effectiveness of individual location updates, a broad global exploration is typically performed first, followed by a gradual transition to precise local exploration. Therefore, as the number of iterations increases, the reference point for tuna foraging changes from random individuals to the optimal individual. In summary, the iterative rule used in this embodiment is as follows:

[0095] If the current iteration number t is less than or equal to Random iteration is used, and its iteration expression is as follows:

[0096]

[0097]

[0098] In the formula, This represents the position of the i-th individual generated in the (t+1)-th iteration, where T is the amount of data in one period. This represents the reference position randomly generated in the t-th iteration. This represents the position of the first individual generated in the t-th iteration, and NU represents the total number of individuals in the tuna population. This represents the position of the i-th individual generated in the t-th iteration. Let represent the position of the (i-1)th individual generated in the t-th iteration, and β be the step size, calculated as follows:

[0099]

[0100] In the formula, b is a random number uniformly distributed between 0 and 1.

[0101] The current iteration number t is greater than The optimal value iteration is used, and its iteration expression is as follows:

[0102]

[0103] In the formula, It is the optimal individual in the t-th iteration of the tuna population.

[0104] Using this iterative approach effectively leverages the characteristics of deep learning, utilizing the multiple features of the Tuna algorithm to perform large-scale searches within the search space. When the number of iterations is small, using random iteration effectively avoids getting trapped in suboptimal local searches. As the number of iterations increases, iterating based on the best individual result effectively seeks the globally optimal solution.

[0105] After iteration, new training samples are taken, and individuals from the tuna population are re-substituted into the previously established model to solve and calculate the loss function. This process of judgment-iteration update-solving-loss function calculation is repeated until the output conditions are met. After the iteration ends, hyperparameter training is performed on the current training dataset. These hyperparameters can be used for subsequent retraining or deployment of the temporal convolutional network prediction model based on periodic signals.

[0106] This embodiment is mainly used to maintain the hyperparameters of the temporal convolutional network prediction model based on periodic signals, eliminating the need for manual setting of hyperparameters. The trained hyperparameters can be used for prediction, early warning and monitoring during the loom operation process, improving the accuracy of early warning and monitoring.

[0107] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0108] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A predictive maintenance method for equipment based on swarm intelligence algorithms, applicable to mechanical equipment undergoing periodic motion, characterized in that, The predictive maintenance method for equipment based on swarm intelligence algorithms includes: Determine the amount of data in one cycle of the device operation time series data in the training dataset; Based on the amount of data within a period, a temporal convolutional network prediction model based on periodic signals is established. This model includes an input layer, a hidden layer, and an output layer. The hidden layer uses the most recently input training sample as the origin and determines the data belonging to the latest period by working backwards along the input time direction according to the amount of data within a period. The remaining historical training samples are considered as data outside the latest period. When the hidden layer performs convolution operations on the data belonging to the latest period, the kernel size... When performing convolution operations on data outside the latest cycle, the kernel size... , The data volume is within one period; the output layer performs a convolution operation on all the data output by the hidden layer to obtain the final prediction result. The input and output time lengths of all hidden layers are the same. The hidden layer and the output layer use dilated causal convolution to perform the convolution operation. Training samples generated from the device operation time series data in the training dataset are taken sequentially. The training samples are input into the temporal convolutional network prediction model based on periodic signals. The tuna algorithm is used to optimize and solve the hyperparameters in the temporal convolutional network prediction model based on periodic signals until the maximum number of iterations is reached. The hyperparameters are then fixed and used for the training and application of the temporal convolutional network prediction model based on periodic signals. The step of using the tuna algorithm to optimize and solve the hyperparameters in the temporal convolutional network prediction model based on periodic signals, and then fixing the hyperparameters after reaching the maximum number of iterations, includes: Initialize the tuna population, including: In the formula, Indicates the first Individual, This indicates the total number of individuals in a tuna population. Indicates the first The initial position of each individual. This represents the upper bound of the search space. The lower bound of the search space is indicated by the fact that each individual in the tuna population represents a combination of hyperparameters of a temporal convolutional network prediction model based on periodic signals. Each individual is fed into a temporal convolutional network prediction model based on periodic signals, and the corresponding predicted value is obtained based on the current input training samples. Calculate the loss function between the predicted value and the actual value corresponding to the training sample to determine the optimal individual in the tuna population; Determine if the maximum number of iterations has been reached. If so, output the optimal individual as the hyperparameter of the temporal convolutional network prediction model based on periodic signals. Otherwise, update the position of individuals in the tuna population according to the iteration rule, and substitute each individual with the updated position back into the temporal convolutional network prediction model based on periodic signals to perform prediction iterations for the new training samples.

2. The predictive maintenance method for equipment based on swarm intelligence algorithm according to claim 1, characterized in that, Determining the amount of data within one period of the device operation time-series data in the training dataset includes: The amount of data obtained from acquiring equipment operation timing data is The training dataset, in the training dataset Randomly select one data point from the device operation timing data. And there is data in the training dataset. Start by searching for data in sequence Equal data , ; Comparison with data The random step size starting from is less than of Data and data The same random step size starting from the same point If the values ​​of each data point are equal, then the data is selected. To data Given data within a period, determine the amount of data within that period. If they are not equal, continue searching for data in sequence. Equal data .

3. The predictive maintenance method for equipment based on swarm intelligence algorithm according to claim 1, characterized in that, The generation of the training samples includes: using a sliding window to process the device operation time-series data in the training dataset and outputting the training samples.

4. The predictive maintenance method for equipment based on swarm intelligence algorithm according to claim 1, characterized in that, The loss function is the mean squared error function, and the loss function for calculating the difference between the predicted value and the actual value corresponding to the training sample includes: In the formula, This is the mean square error value. The amount of data within one period. For the first A true value, For the first One predicted value.

5. The predictive maintenance method for equipment based on swarm intelligence algorithm according to claim 1, characterized in that, The iteration rule is as follows: At the current iteration number Less than or equal to Random iteration is used, and its iteration expression is as follows: In the formula, Indicates the first The generation during the nth iteration The position of each individual The amount of data within one period. Indicates the first The reference position is randomly generated in the next iteration. Indicates the first The position of the first individual generated in the next iteration. This indicates the total number of individuals in a tuna population. Indicates the first The generation during the nth iteration The position of each individual Indicates the first The generation during the nth iteration The position of each individual Let the step size be , and the calculation is as follows: In the formula, These are random numbers that are uniformly distributed between 0 and 1. At the current iteration number Greater than The optimal value iteration is used, and its iteration expression is as follows: In the formula, In the tuna population, the first The optimal individual in the next iteration.

Images