Industrial equipment state monitoring and early warning system and method based on machine learning

Abstract

The invention relates to industrial equipment monitoring, in particular to an industrial equipment state monitoring and early warning system and method based on machine learning, and the method comprises the steps: carrying out the model training of a convolutional neural network, and obtaining a trained convolutional neural network; constructing a pruning strategy generation model based on the trained convolutional neural network, and performing multi-objective optimization solution on the pruning strategy generation model to obtain an optimized pruning strategy; pruning the trained convolutional neural network according to the optimized pruning strategy to obtain a pruned convolutional neural network; collecting a vibration signal of the industrial equipment to be monitored, converting the vibration signal into a spatial frequency image, and calculating the central component energy of the spatial frequency in the spatial frequency image; calculating an energy value according to the central component energy of the spatial frequency in the spatial frequency image; inputting the energy value into a pruning convolutional neural network for state monitoring; according to the technical scheme provided by the invention, the defect that the state of the industrial equipment is difficult to accurately monitor and early warn in real time can be effectively overcome.

Description

Technical Field

[0001] The present invention relates to industrial equipment monitoring, and in particular to an industrial equipment status monitoring and early warning system and method based on machine learning. Background Art

[0002] The industrial equipment status monitoring and early warning system is an intelligent system that integrates a variety of advanced technologies. It aims to monitor the operating status of industrial equipment in real time, detect and warn of potential faults in a timely manner, and ensure the safety and stability of industrial production.

[0003] The existing industrial equipment condition monitoring method combines traditional signal analysis methods with machine learning, including two stages: signal feature extraction and feature recognition and classification. First, the vibration signal collected by the sensor is preprocessed, the signal is reduced in dimension, and then the relevant features of the signal are extracted using traditional signal analysis methods. Then, according to different application scenarios, a suitable shallow neural network classifier is used for fault recognition and classification, so that fault diagnosis information can be obtained through feature recognition by the classifier.

[0004] However, most traditional classifiers are shallow learning models. This simple shallow architecture limits the learning of complex nonlinear relationships in industrial equipment condition monitoring, resulting in low accuracy of condition monitoring warning results. At the same time, the traditional neural network structure is relatively complex. On the one hand, it makes the model large and inconvenient to deploy. On the other hand, the amount of calculation is also large when executing tasks, which is inefficient and has poor real-time performance. Summary of the invention

[0005] 1. Technical issues to be resolved In view of the above-mentioned shortcomings of the prior art, the present invention provides an industrial equipment status monitoring and early warning system and method based on machine learning, which can effectively overcome the defects of the prior art that it is difficult to accurately and real-time monitor and warn the status of industrial equipment.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: An industrial equipment condition monitoring and early warning system based on machine learning, comprising a neural network training module, an optimization pruning strategy generation module, a pruning neural network generation module, a vibration signal acquisition module, a central component energy calculation module, an energy value calculation module and a condition monitoring and early warning module; The neural network training module performs model training on the convolutional neural network to obtain a trained convolutional neural network; Optimize the pruning strategy generation module, build a pruning strategy generation model based on the trained convolutional neural network, and perform multi-objective optimization on the pruning strategy generation model to obtain the optimized pruning strategy; A pruned neural network generation module prunes the trained convolutional neural network according to an optimized pruning strategy to obtain a pruned convolutional neural network; Vibration signal acquisition module, which collects vibration signals of industrial equipment to be monitored; A central component energy calculation module converts the vibration signal into a spatial frequency image and calculates the central component energy of the spatial frequency in the spatial frequency image; An energy value calculation module calculates the energy value according to the central component energy of the spatial frequency in the spatial frequency image; The condition monitoring and early warning module inputs the energy value into the pruned convolutional neural network, performs condition monitoring on the industrial equipment to be monitored, and issues early warnings based on the condition monitoring results.

[0007] Preferably, the neural network training module performs model training on the convolutional neural network to obtain a trained convolutional neural network, including: S11, establishing a training data set, and dividing the training data set into a training set, a validation set, and a test set according to a preset ratio; S12, setting the loss function and optimizer of the convolutional neural network, and inputting the training set into the convolutional neural network for model training; S13, the optimizer updates the network parameters according to the loss function and the network gradient information; S14. If the loss value is less than the preset threshold, the model training ends, and the current convolutional neural network is the trained convolutional neural network. Otherwise, return to S12 and continue to use the training set for model training. S15. Input the validation set into the trained convolutional neural network, evaluate the generalization ability of the model by observing the performance on the validation set, and tune the hyperparameters and structure of the model; S16. Input the test set into the tuned convolutional neural network to evaluate the model performance.

[0008] Preferably, in S11, a training data set is established, and the training data set is divided into a training set, a validation set, and a test set according to a preset ratio, including: S111. Collect historical vibration signals of industrial equipment, convert the historical vibration signals into spatial frequency images, and calculate the central component energy of the spatial frequency in the spatial frequency image; S112, calculating an energy value according to the central component energy of the spatial frequency in the spatial frequency image; S113, determining the state abnormality type corresponding to the historical vibration signal, and marking the energy value corresponding to the historical vibration signal to establish a training data set; S114. Divide the training data set into a training set, a validation set, and a test set according to a preset ratio.

[0009] Preferably, in S111, historical vibration signals of industrial equipment are collected, the historical vibration signals are converted into spatial frequency images, and the central component energy of the spatial frequency in the spatial frequency image is calculated, including: S1111. Collect historical vibration signals of industrial equipment over time; S1112, performing short-time Fourier transform on the historical vibration signal, calculating the intensity information of each frequency component of the historical vibration signal that changes with time, and inserting the calculated intensity information into the time axis and the frequency axis to form a time-frequency image; S1113, performing a two-dimensional Fourier transform on the time frequency image to generate a spatial frequency image; S1114. Calculate the central component energy of the spatial frequency in the spatial frequency image.

[0010] Preferably, calculating the energy value according to the central component energy of the spatial frequency in the spatial frequency image in S112 includes: According to the central component energy of the spatial frequency in the spatial frequency image, the energy value is calculated using the following formula: ; in, E j The energy value is j elements, , S is the total number of frequency points, L is the window length, Indicates rounding down. e i For the i The central component energy at the frequency point, s is the number of frequency points in the window.

[0011] Preferably, the optimizer in S13 updates the network parameters according to the loss function and the network gradient information, including: S131, based on the current network parameters and loss function, using the currently input small batch training set data to calculate the gradient of the loss function with respect to weights and biases; S132. Update the first-order moment and the second-order moment according to the gradient of the loss function with respect to the weight and the bias; S133, performing deviation correction on the updated first-order moment and second-order moment; S134, updating the weight and bias according to the first-order moment and the second-order moment after the deviation correction; Among them, weights and biases are the network parameters that the convolutional neural network needs to learn, the first-order moment is used to store the exponential moving average of the gradient, initialized to 0, and the second-order moment is used to store the exponential moving average of the square of the gradient, initialized to 0.

[0012] Preferably, in S132, the first-order moment and the second-order moment are updated according to the gradient of the loss function with respect to the weight and the bias, including: According to the gradient of the loss function with respect to weights and biases, the first-order moment is updated using the following formula: ; According to the gradient of the loss function with respect to weights and biases, the second-order moment is updated using the following formula: ; in, m t For the t The first-order moment of the iteration, m t－1 For the t -1st order moment of iteration, is the exponential decay rate of the first-order moment estimate, and the commonly used value is 0.9; v t For the t The second-order moment of the iteration, v t－1 For the t -1 second-order moment of iteration, is the exponential decay rate of the second-order moment estimate, and the commonly used value is 0.999. g t For the t The gradient of the iteration, that is, the gradient of the current loss function with respect to the weight or bias; In S133, the updated first-order moment and second-order moment are subjected to deviation correction, including: The updated first-order moment is corrected using the following formula: ; The updated second-order moment is corrected using the following formula: ; in, m t ’ After bias correction t The first-order moment of the iteration, v t ’ After bias correction t The second moment of the iteration; In S134, the weights and biases are updated according to the first-order moment and the second-order moment after the deviation correction, including: According to the first-order moment and second-order moment after deviation correction, the weights are updated using the following formula: ; According to the first-order moment and second-order moment after deviation correction, the bias is updated using the following formula: ; in, W t For the t The weight of the iteration, W t－1 For the t - weight of 1 iteration; B t For the t The bias of the iteration, B t－1 For the t -1 iteration bias, is the learning rate, is a number close to 0.

[0013] Preferably, the optimized pruning strategy generation module constructs a pruning strategy generation model based on a trained convolutional neural network, and performs multi-objective optimization on the pruning strategy generation model to obtain an optimized pruning strategy, including: S21, determining the objective function generated by the pruning strategy based on the trained convolutional neural network, and the constraint conditions for generating the pruning strategy; S22, constructing a pruning strategy generation model according to the objective function and constraints generated by the pruning strategy; S23, using a multi-objective raccoon optimization algorithm to perform multi-objective optimization on the pruning strategy generation model to obtain an optimized pruning strategy; Among them, the objective functions generated by the pruning strategy include minimum model size, highest accuracy and minimum loss value, and the optimized pruning strategy includes pruning order, pruning ratio and pruning threshold; In S23, the multi-objective raccoon optimization algorithm is used to perform multi-objective optimization on the pruning strategy generation model to obtain the optimized pruning strategy, including: S231, encoding the pruning parameters in the pruning strategy as the search space of the algorithm; S232, initializing algorithm parameters, and randomly generating a certain number of raccoon individuals according to the set parameters to form an initial population, each raccoon individual representing a candidate solution in the solution space; S233. In the search space, some raccoon individuals simulate the behavior of climbing up a tree to scare the iguana, while other raccoon individuals wait under the tree and update the positions of the raccoon individuals according to the hunting strategy, so that the population moves continuously to the search space to explore new solution spaces, compare the fitness values ​​of the raccoon individuals before and after the position update, and replace the original raccoon individuals; S234, when a predator appears, the raccoon individual adjusts its escape direction according to its current position during the escape process to find a better solution, compares the fitness values ​​of the raccoon individual before and after the position update, and replaces the original raccoon individual; S235. According to the Pareto dominance relationship of the multi-objective optimization problem, the raccoon individuals in the population are sorted, non-dominated solutions are selected, and the crowding distance between the non-dominated solutions is calculated; S236. According to the non-dominated solutions and the crowding distances between them, a part of raccoon individuals are selected as the next generation population, and new random raccoon individuals are introduced to maintain the population diversity; S237, judging whether the iteration termination condition is satisfied, if not, returning to S233, otherwise, selecting the global optimal solution from the current non-dominated solutions by calculating the hypervolume as the optimization pruning strategy; Among them, the iguana represents the current optimal solution, and the predator represents the local optimal solution.

[0014] Preferably, in S233, in the search space, some raccoon individuals simulate the behavior of climbing up a tree to scare the iguana, and other raccoon individuals wait under the tree, and the positions of the raccoon individuals are updated according to the hunting strategy, so that the population moves continuously to the search space to explore new solution spaces, including: During the hunting process, the position of the raccoon individual is updated using the following formula: ; in, P n ’ For the n The updated position of each raccoon individual, P n For the n The position of the raccoon individuals before the update, P best For the location of the iguana, r 1 is a random number, is the control factor; In S234, when a predator appears, the raccoon adjusts its escape direction according to its current position during the escape process to find a better solution, including: During the escape process, the position of the raccoon individual is updated using the following formula: ; in, P n ’ For the n The updated position of each raccoon individual, P n For the n The position of the raccoon individuals before the update, is the step size control parameter, d n For the n The escape direction of each raccoon, , r 2. r 3 are all random numbers, LB , UB are the upper and lower bounds of the search space, respectively. is the normalization function.

[0015] A method for monitoring and early warning of industrial equipment status based on machine learning, comprising the following steps: S1. Perform model training on the convolutional neural network to obtain a trained convolutional neural network; S2. Building a pruning strategy generation model based on the trained convolutional neural network, and performing multi-objective optimization on the pruning strategy generation model to obtain an optimized pruning strategy; S3, pruning the trained convolutional neural network according to the optimized pruning strategy to obtain a pruned convolutional neural network; S4, collecting vibration signals of the industrial equipment to be monitored, converting the vibration signals into spatial frequency images, and calculating the central component energy of the spatial frequency in the spatial frequency image; S5, calculating an energy value according to the central component energy of the spatial frequency in the spatial frequency image; S6. Input the energy value into the pruned convolutional neural network, perform status monitoring on the industrial equipment to be monitored, and issue an early warning based on the status monitoring results.

[0016] Compared with the prior art, the industrial equipment status monitoring and early warning system and method based on machine learning provided by the present invention has the following beneficial effects: 1) The vibration signal acquisition module acquires the vibration signal of the industrial equipment to be monitored. The central component energy calculation module converts the vibration signal into a spatial frequency image and calculates the central component energy of the spatial frequency in the spatial frequency image. The energy value calculation module calculates the energy value based on the central component energy of the spatial frequency in the spatial frequency image. The state monitoring and early warning module inputs the energy value into the pruned convolutional neural network to perform state monitoring on the industrial equipment to be monitored and issue an early warning based on the state monitoring results. By performing signal processing and feature extraction on the vibration signal and using the pruned convolutional neural network for feature recognition, more accurate state monitoring and early warning results can be obtained. 2) The neural network training module performs model training on the convolutional neural network to obtain a trained convolutional neural network. The optimized pruning strategy generation module constructs a pruning strategy generation model based on the trained convolutional neural network, and performs multi-objective optimization on the pruning strategy generation model to obtain an optimized pruning strategy. The pruned neural network generation module prunes the trained convolutional neural network according to the optimized pruning strategy to obtain a pruned convolutional neural network. By constructing a pruning strategy generation model and encoding the pruning parameters in the pruning strategy as the search space of the algorithm, the multi-objective optimization algorithm can be used to perform multi-objective optimization on the pruning strategy generation model to obtain an optimized pruning strategy, thereby accurately and effectively removing the redundant structure inside the convolutional neural network to obtain a pruned convolutional neural network with a simpler structure. On the one hand, it is easy to deploy, and on the other hand, it can reduce the amount of calculation when executing tasks and give results faster, meeting the real-time requirements of industrial equipment status monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0018] Figure 1 It is a schematic diagram of the system of the present invention; Figure 2 It is a schematic diagram of the process of the present invention; Figure 3 A schematic diagram of the process of model training for a convolutional neural network in the present invention; Figure 4 A schematic diagram of the process of constructing a pruning strategy generation model and performing multi-objective optimization on the pruning strategy generation model in the present invention. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] An industrial equipment condition monitoring and early warning system based on machine learning, such as Figure 1As shown, it includes a neural network training module, an optimization pruning strategy generation module, a pruning neural network generation module, a vibration signal acquisition module, a central component energy calculation module, an energy value calculation module and a state monitoring and early warning module; The neural network training module performs model training on the convolutional neural network to obtain a trained convolutional neural network; Optimize the pruning strategy generation module, build a pruning strategy generation model based on the trained convolutional neural network, and perform multi-objective optimization on the pruning strategy generation model to obtain the optimized pruning strategy; A pruned neural network generation module prunes the trained convolutional neural network according to an optimized pruning strategy to obtain a pruned convolutional neural network; Vibration signal acquisition module, which collects vibration signals of industrial equipment to be monitored; A central component energy calculation module converts the vibration signal into a spatial frequency image and calculates the central component energy of the spatial frequency in the spatial frequency image; An energy value calculation module calculates the energy value according to the central component energy of the spatial frequency in the spatial frequency image; The condition monitoring and early warning module inputs the energy value into the pruned convolutional neural network, performs condition monitoring on the industrial equipment to be monitored, and issues early warnings based on the condition monitoring results.

[0021] The above technical solution can obtain more accurate status monitoring and early warning results by performing signal processing and feature extraction on vibration signals and using pruned convolutional neural networks for feature recognition.

[0022] ① The neural network training module performs model training on the convolutional neural network to obtain a trained convolutional neural network, such as Figure 3 As shown, including: S11, establishing a training data set, and dividing the training data set into a training set, a validation set, and a test set according to a preset ratio; S12, setting the loss function and optimizer of the convolutional neural network, and inputting the training set into the convolutional neural network for model training; S13, the optimizer updates the network parameters according to the loss function and the network gradient information; S14. If the loss value is less than the preset threshold, the model training ends, and the current convolutional neural network is the trained convolutional neural network. Otherwise, return to S12 and continue to use the training set for model training. S15. Input the validation set into the trained convolutional neural network, evaluate the generalization ability of the model by observing the performance on the validation set, and tune the hyperparameters and structure of the model; S16. Input the test set into the tuned convolutional neural network to evaluate the model performance.

[0023] 1) In S11, a training data set is established and divided into a training set, a validation set, and a test set according to a preset ratio, including: S111. Collect historical vibration signals of industrial equipment, convert the historical vibration signals into spatial frequency images, and calculate the central component energy of the spatial frequency in the spatial frequency image; S112, calculating an energy value according to the central component energy of the spatial frequency in the spatial frequency image; S113, determining the state abnormality type corresponding to the historical vibration signal, and marking the energy value corresponding to the historical vibration signal to establish a training data set; S114. Divide the training data set into a training set, a validation set, and a test set according to a preset ratio.

[0024] Specifically, in S111, historical vibration signals of industrial equipment are collected, the historical vibration signals are converted into spatial frequency images, and the central component energy of the spatial frequency in the spatial frequency image is calculated, including: S1111. Collect historical vibration signals of industrial equipment over time; S1112, performing short-time Fourier transform on the historical vibration signal, calculating the intensity information of each frequency component of the historical vibration signal that changes with time, and inserting the calculated intensity information into the time axis and the frequency axis to form a time-frequency image; S1113, performing a two-dimensional Fourier transform on the time frequency image to generate a spatial frequency image; S1114. Calculate the central component energy of the spatial frequency in the spatial frequency image.

[0025] Specifically, in S112, the energy value is calculated according to the central component energy of the spatial frequency in the spatial frequency image, including: According to the central component energy of the spatial frequency in the spatial frequency image, the energy value is calculated using the following formula: ; in, E j The energy value is j elements, , S is the total number of frequency points, L is the window length, Indicates rounding down. e i For the i The central component energy at the frequency point, s is the number of frequency points in the window.

[0026] 2) In S13, the optimizer updates the network parameters according to the loss function and network gradient information, such as Figure 3As shown, including: S131, based on the current network parameters and loss function, using the currently input small batch training set data to calculate the gradient of the loss function with respect to weights and biases; S132. Update the first-order moment and the second-order moment according to the gradient of the loss function with respect to the weight and the bias; S133, performing deviation correction on the updated first-order moment and second-order moment; S134, updating the weight and bias according to the first-order moment and the second-order moment after the deviation correction; Among them, weights and biases are the network parameters that the convolutional neural network needs to learn, the first-order moment is used to store the exponential moving average of the gradient, initialized to 0, and the second-order moment is used to store the exponential moving average of the square of the gradient, initialized to 0.

[0027] Specifically, in S132, the first-order moment and the second-order moment are updated according to the gradient of the loss function with respect to the weight and the bias, including: According to the gradient of the loss function with respect to weights and biases, the first-order moment is updated using the following formula: ; According to the gradient of the loss function with respect to weights and biases, the second-order moment is updated using the following formula: ; in, m t For the t The first-order moment of the iteration, m t－1 For the t -1st order moment of iteration, is the exponential decay rate of the first-order moment estimate, and the commonly used value is 0.9; v t For the t The second-order moment of the iteration, v t－1 For the t -1 second-order moment of iteration, is the exponential decay rate of the second-order moment estimate, and the commonly used value is 0.999. g t For the t The gradient of the iteration, that is, the gradient of the current loss function with respect to the weight or bias.

[0028] Specifically, in S133, deviation correction is performed on the updated first-order moment and second-order moment, including: The updated first-order moment is corrected using the following formula: ; The updated second-order moment is corrected using the following formula: ; in, m t ’ After bias correction t The first-order moment of the iteration, v t ’ After bias correction t The second moment of the iteration.

[0029] Specifically, in S134, the weights and biases are updated according to the first-order moment and the second-order moment after the deviation correction, including: According to the first-order moment and second-order moment after deviation correction, the weights are updated using the following formula: ; According to the first-order moment and second-order moment after deviation correction, the bias is updated using the following formula: ; in, W t For the t The weight of the iteration, W t－1 For the t - weight of 1 iteration; B t For the t The bias of the iteration, B t－1 For the t -1 iteration bias, is the learning rate, is a number close to 0.

[0030] ② The optimized pruning strategy generation module builds a pruning strategy generation model based on the trained convolutional neural network, and performs multi-objective optimization on the pruning strategy generation model to obtain the optimized pruning strategy, such as Figure 4 As shown, including: S21, determining the objective function generated by the pruning strategy based on the trained convolutional neural network, and the constraint conditions for generating the pruning strategy; S22, constructing a pruning strategy generation model according to the objective function and constraints generated by the pruning strategy; S23, using a multi-objective raccoon optimization algorithm to perform multi-objective optimization on the pruning strategy generation model to obtain an optimized pruning strategy; Among them, the objective functions generated by the pruning strategy include minimum model size, highest accuracy and minimum loss value, and the optimized pruning strategy includes pruning order, pruning ratio and pruning threshold.

[0031] In S23, the multi-objective raccoon optimization algorithm is used to perform multi-objective optimization on the pruning strategy generation model to obtain the optimized pruning strategy, such as Figure 4 As shown, including: S231, encoding the pruning parameters in the pruning strategy as the search space of the algorithm; S232, initializing algorithm parameters, and randomly generating a certain number of raccoon individuals according to the set parameters to form an initial population, each raccoon individual representing a candidate solution in the solution space; S233. In the search space, some raccoon individuals simulate the behavior of climbing up a tree to scare the iguana, while other raccoon individuals wait under the tree and update the positions of the raccoon individuals according to the hunting strategy, so that the population moves continuously to the search space to explore new solution spaces, compare the fitness values ​​of the raccoon individuals before and after the position update, and replace the original raccoon individuals; S234, when a predator appears, the raccoon individual adjusts its escape direction according to its current position during the escape process to find a better solution, compares the fitness values ​​of the raccoon individual before and after the position update, and replaces the original raccoon individual; S235. According to the Pareto dominance relationship of the multi-objective optimization problem, the raccoon individuals in the population are sorted, non-dominated solutions are selected, and the crowding distance between the non-dominated solutions is calculated; S236. According to the non-dominated solutions and the crowding distances between them, a part of raccoon individuals are selected as the next generation population, and new random raccoon individuals are introduced to maintain the population diversity; S237, judging whether the iteration termination condition is satisfied, if not, returning to S233, otherwise, selecting the global optimal solution from the current non-dominated solutions by calculating the hypervolume as the optimization pruning strategy; Among them, the iguana represents the current optimal solution, and the predator represents the local optimal solution.

[0032] Specifically, in S233, in the search space, some raccoon individuals simulate the behavior of climbing up a tree to scare the iguana, while other raccoon individuals wait under the tree and update the positions of the raccoon individuals according to the hunting strategy, so that the population moves continuously to the search space to explore new solution spaces, including: During the hunting process, the position of the raccoon individual is updated using the following formula: ; in, P n ’ For the n The updated position of each raccoon individual, P n For the n The position of the raccoon individuals before the update,P best For the location of the iguana, r 1 is a random number, is the control factor.

[0033] Specifically, when a predator appears in S234, the raccoon individual adjusts its escape direction according to its current position during the escape process to find a better solution, including: During the escape process, the position of the raccoon individual is updated using the following formula: ; in, P n ’ For the n The updated position of each raccoon individual, P n For the n The position of the raccoon individuals before the update, is the step size control parameter, d n For the n The escape direction of each raccoon, , r 2. r 3 are all random numbers, LB , UB are the upper and lower bounds of the search space, respectively. is the normalization function.

[0034] The above technical solution, by constructing a pruning strategy generation model and encoding the pruning parameters in the pruning strategy as the search space of the algorithm, can use the multi-objective optimization algorithm to perform multi-objective optimization on the pruning strategy generation model to obtain an optimized pruning strategy, and then accurately and effectively remove the redundant structure inside the convolutional neural network to obtain a pruned convolutional neural network with a simpler structure. On the one hand, it is easy to deploy, and on the other hand, it can reduce the amount of calculation when executing tasks and give results faster, meeting the real-time requirements of industrial equipment status monitoring.

[0035] In the technical solution of this application, on the basis of disclosing the above-mentioned industrial equipment status monitoring and early warning system based on machine learning, a method for industrial equipment status monitoring and early warning based on machine learning is also disclosed, such as Figure 2 As shown, the following steps are included: S1. Perform model training on the convolutional neural network to obtain a trained convolutional neural network; S2. Building a pruning strategy generation model based on the trained convolutional neural network, and performing multi-objective optimization on the pruning strategy generation model to obtain an optimized pruning strategy; S3, pruning the trained convolutional neural network according to the optimized pruning strategy to obtain a pruned convolutional neural network; S4, collecting vibration signals of the industrial equipment to be monitored, converting the vibration signals into spatial frequency images, and calculating the central component energy of the spatial frequency in the spatial frequency image; S5, calculating an energy value according to the central component energy of the spatial frequency in the spatial frequency image; S6. Input the energy value into the pruned convolutional neural network, perform status monitoring on the industrial equipment to be monitored, and issue an early warning based on the status monitoring results.

[0036] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An industrial equipment condition monitoring and early warning system based on machine learning, characterized by: It includes a neural network training module, an optimized pruning strategy generation module, a pruning neural network generation module, a vibration signal acquisition module, a central component energy calculation module, an energy value calculation module and a status monitoring and early warning module; The neural network training module performs model training on the convolutional neural network to obtain a trained convolutional neural network; Optimize the pruning strategy generation module, build a pruning strategy generation model based on the trained convolutional neural network, and perform multi-objective optimization on the pruning strategy generation model to obtain the optimized pruning strategy; A pruned neural network generation module prunes the trained convolutional neural network according to an optimized pruning strategy to obtain a pruned convolutional neural network; Vibration signal acquisition module, which collects vibration signals of industrial equipment to be monitored; A central component energy calculation module converts the vibration signal into a spatial frequency image and calculates the central component energy of the spatial frequency in the spatial frequency image; An energy value calculation module calculates the energy value according to the central component energy of the spatial frequency in the spatial frequency image; The condition monitoring and early warning module inputs the energy value into the pruned convolutional neural network, performs condition monitoring on the industrial equipment to be monitored, and issues early warnings based on the condition monitoring results.

2. The industrial equipment status monitoring and early warning system based on machine learning according to claim 1 is characterized in that: The neural network training module performs model training on the convolutional neural network to obtain a trained convolutional neural network, including: S11, establishing a training data set, and dividing the training data set into a training set, a validation set, and a test set according to a preset ratio; S12, setting the loss function and optimizer of the convolutional neural network, and inputting the training set into the convolutional neural network for model training; S13, the optimizer updates the network parameters according to the loss function and the network gradient information; S14. If the loss value is less than the preset threshold, the model training ends, and the current convolutional neural network is the trained convolutional neural network. Otherwise, return to S12 and continue to use the training set for model training. S15. Input the validation set into the trained convolutional neural network, evaluate the generalization ability of the model by observing the performance on the validation set, and tune the hyperparameters and structure of the model; S16. Input the test set into the tuned convolutional neural network to evaluate the model performance.

3. The industrial equipment condition monitoring and early warning system based on machine learning according to claim 2 is characterized in that: In S11, a training data set is established and divided into a training set, a validation set, and a test set according to a preset ratio, including: S111. Collect historical vibration signals of industrial equipment, convert the historical vibration signals into spatial frequency images, and calculate the central component energy of the spatial frequency in the spatial frequency image; S112, calculating an energy value according to the central component energy of the spatial frequency in the spatial frequency image; S113, determining the state abnormality type corresponding to the historical vibration signal, and marking the energy value corresponding to the historical vibration signal to establish a training data set; S114. Divide the training data set into a training set, a validation set, and a test set according to a preset ratio.

4. The industrial equipment condition monitoring and early warning system based on machine learning according to claim 3 is characterized in that: S111 collects historical vibration signals of industrial equipment, converts the historical vibration signals into spatial frequency images, and calculates the central component energy of the spatial frequency in the spatial frequency image, including: S1111. Collect historical vibration signals of industrial equipment over time; S1112, performing short-time Fourier transform on the historical vibration signal, calculating the intensity information of each frequency component of the historical vibration signal that changes with time, and inserting the calculated intensity information into the time axis and the frequency axis to form a time-frequency image; S1113, performing a two-dimensional Fourier transform on the time frequency image to generate a spatial frequency image; S1114. Calculate the central component energy of the spatial frequency in the spatial frequency image.

5. The industrial equipment status monitoring and early warning system based on machine learning according to claim 3 is characterized in that: In S112, the energy value is calculated according to the central component energy of the spatial frequency in the spatial frequency image, including: According to the central component energy of the spatial frequency in the spatial frequency image, the energy value is calculated using the following formula: ； in, E j The energy value is j elements, , S is the total number of frequency points, L is the window length, Indicates rounding down. e i For the i The central component energy at the frequency point, s is the number of frequency points in the window.

6. The industrial equipment condition monitoring and early warning system based on machine learning according to claim 2 is characterized in that: In S13, the optimizer updates the network parameters according to the loss function and network gradient information, including: S131, based on the current network parameters and loss function, using the currently input small batch training set data to calculate the gradient of the loss function with respect to weights and biases; S132. Update the first-order moment and the second-order moment according to the gradient of the loss function with respect to the weight and the bias; S133, performing deviation correction on the updated first-order moment and second-order moment; S134, updating the weight and bias according to the first-order moment and the second-order moment after the deviation correction; Among them, weights and biases are the network parameters that the convolutional neural network needs to learn, the first-order moment is used to store the exponential moving average of the gradient, initialized to 0, and the second-order moment is used to store the exponential moving average of the square of the gradient, initialized to 0.

7. The industrial equipment condition monitoring and early warning system based on machine learning according to claim 6 is characterized in that: In S132, the first-order moment and the second-order moment are updated according to the gradient of the loss function with respect to the weight and the bias, including: According to the gradient of the loss function with respect to weights and biases, the first-order moment is updated using the following formula: ； According to the gradient of the loss function with respect to weights and biases, the second-order moment is updated using the following formula: ； in, m t For the t The first-order moment of the iteration, m t－1 For the t -1st order moment of iteration, is the exponential decay rate of the first-order moment estimate, and the commonly used value is 0.9; v t For the t The second-order moment of the iteration, v t－1 For the t -1 second-order moment of iteration, is the exponential decay rate of the second-order moment estimate, and the commonly used value is 0.

999. g t For the t The gradient of the iteration, that is, the gradient of the current loss function with respect to the weight or bias; In S133, the updated first-order moment and second-order moment are corrected for deviations, including: The updated first-order moment is corrected using the following formula: ； The updated second-order moment is corrected using the following formula: ； in, m t ’ After bias correction t The first-order moment of the iteration, v t ’ After bias correction t The second moment of the iteration; In S134, the weights and biases are updated according to the first-order moment and the second-order moment after the deviation correction, including: According to the first-order moment and second-order moment after deviation correction, the weights are updated using the following formula: ； According to the first-order moment and second-order moment after deviation correction, the bias is updated using the following formula: ； in, W t For the t The weight of the iteration, W t－1 For the t - weight of 1 iteration; B t For the t The bias of the iteration, B t－1 For the t -1 iteration bias, is the learning rate, is a number close to 0.

8. The industrial equipment condition monitoring and early warning system based on machine learning according to claim 1 is characterized in that: The optimized pruning strategy generation module constructs a pruning strategy generation model based on the trained convolutional neural network, and performs multi-objective optimization on the pruning strategy generation model to obtain an optimized pruning strategy, including: S21. Determine the objective function of the pruning strategy generation based on the trained convolutional neural network, and the constraint conditions of the pruning strategy generation; S22, constructing a pruning strategy generation model according to the objective function and constraints generated by the pruning strategy; S23, using a multi-objective raccoon optimization algorithm to perform multi-objective optimization on the pruning strategy generation model to obtain an optimized pruning strategy; Among them, the objective functions generated by the pruning strategy include minimum model size, highest accuracy and minimum loss value, and the optimized pruning strategy includes pruning order, pruning ratio and pruning threshold; In S23, the multi-objective raccoon optimization algorithm is used to perform multi-objective optimization on the pruning strategy generation model to obtain the optimized pruning strategy, including: S231, encoding the pruning parameters in the pruning strategy as the search space of the algorithm; S232, initializing algorithm parameters, and randomly generating a certain number of raccoon individuals according to the set parameters to form an initial population, each raccoon individual representing a candidate solution in the solution space; S233. In the search space, some raccoon individuals simulate the behavior of climbing up a tree to scare the iguana, while other raccoon individuals wait under the tree and update the positions of the raccoon individuals according to the hunting strategy, so that the population moves continuously to the search space to explore new solution spaces, compare the fitness values ​​of the raccoon individuals before and after the position update, and replace the original raccoon individuals; S234, when a predator appears, the raccoon individual adjusts its escape direction according to its current position during the escape process to find a better solution, compares the fitness values ​​of the raccoon individual before and after the position update, and replaces the original raccoon individual; S235. According to the Pareto dominance relationship of the multi-objective optimization problem, the raccoon individuals in the population are sorted, non-dominated solutions are selected, and the crowding distance between the non-dominated solutions is calculated; S236. According to the non-dominated solutions and the crowding distances between them, a part of raccoon individuals are selected as the next generation population, and new random raccoon individuals are introduced to maintain the population diversity; S237, judging whether the iteration termination condition is satisfied, if not, returning to S233, otherwise, selecting the global optimal solution from the current non-dominated solutions by calculating the hypervolume as the optimization pruning strategy; Among them, the iguana represents the current optimal solution, and the predator represents the local optimal solution.

9. The industrial equipment condition monitoring and early warning system based on machine learning according to claim 8 is characterized in that: In S233, in the search space, some raccoons simulate climbing up a tree to scare the iguana, while other raccoons wait under the tree and update their positions according to the hunting strategy, so that the population moves continuously to the search space to explore new solution spaces, including: During the hunting process, the position of the raccoon individual is updated using the following formula: ； in, P n ’ For the n The updated position of each raccoon individual, P n For the n The position of the raccoon individuals before the update, P best For the location of the iguana, r 1 is a random number, is the control factor; In S234, when a predator appears, the raccoon adjusts its escape direction according to its current position during the escape process to find a better solution, including: During the escape process, the position of the raccoon individual is updated using the following formula: ； in, P n ’ For the n The updated position of each raccoon individual, P n For the n The position of the raccoon individuals before the update, is the step size control parameter, d n For the n The escape direction of each raccoon, , r 2. r 3 are all random numbers, LB , UB are the upper and lower bounds of the search space, respectively. is the normalization function.

10. A method for monitoring and early warning of industrial equipment status based on machine learning, applied to the industrial equipment status monitoring and early warning system based on machine learning according to claim 1, characterized in that: The following steps are involved: S1. Perform model training on the convolutional neural network to obtain a trained convolutional neural network; S2. Building a pruning strategy generation model based on the trained convolutional neural network, and performing multi-objective optimization on the pruning strategy generation model to obtain an optimized pruning strategy; S3, pruning the trained convolutional neural network according to the optimized pruning strategy to obtain a pruned convolutional neural network; S4, collecting vibration signals of the industrial equipment to be monitored, converting the vibration signals into spatial frequency images, and calculating the central component energy of the spatial frequency in the spatial frequency image; S5, calculating an energy value according to the central component energy of the spatial frequency in the spatial frequency image; S6. Input the energy value into the pruned convolutional neural network, perform status monitoring on the industrial equipment to be monitored, and issue an early warning based on the status monitoring results.

Images