Artificial Intelligence-Based Predictive Maintenance Method and Device for Equipment

Abstract

A device predictive maintenance method and device based on artificial intelligence provided by an embodiment of the present invention relate to the technical field of data processing. By identifying the statistical characteristics of data, focusing on the most significant changing features, and adjusting the model focus. Moreover, the training sample set of the feature extraction model is adjusted based on the abnormal evaluation results, which can suppress abnormal data, protect the model from the interference of these data, and optimize the effect of feature extraction. Quantum state modeling can better capture the non-linear relationships in data, avoid information loss of traditional linear dimensionality reduction algorithms, and thus improve the expressiveness of the features after dimensionality reduction. The embodiment of the present invention improves the prediction accuracy and reliability of the model, ensures the data quality input into the classifier model, and reduces the risk of misjudgment through data compression based on quantum state amplitudes and optimized feature extraction.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular, to a device predictive maintenance method and device based on artificial intelligence. Background Art

[0002] With the development of industrial automation and information technology, pre-baked anode production equipment plays an important role in the aluminum electrolysis industry. The operating status of these devices directly affects production efficiency and product quality. Traditional device maintenance methods rely on regular inspections and empirical judgments, which not only take a long time but also make it difficult to achieve real-time and accurate assessment of the device status. In this context, intelligent device health management technology is particularly important. By using advanced data analysis technologies such as machine learning and artificial intelligence, the device status can be monitored in real time, potential faults can be predicted, thereby reducing downtime, lowering maintenance costs, and improving production efficiency.

[0003] Traditional methods usually rely on batch processing or regular sampling, resulting in time delays in data processing and feature extraction. Moreover, when sudden abnormal situations occur during device operation, traditional methods fail to effectively filter these abnormal data, leading to misjudgment of the model. In addition, most traditional dimensionality reduction algorithms are based on linear transformation and cannot effectively process data with non-linear structures, resulting in information loss. Due to information loss and linear limitations, the features after dimensionality reduction are difficult to fully reflect the complexity and diversity of the original data, affecting the performance of the classifier. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a device predictive maintenance method and device based on artificial intelligence, which can improve the prediction accuracy of the device failure state.

[0005] In a first aspect, an embodiment of the present invention provides a device predictive maintenance method based on artificial intelligence, where the method includes: obtaining operation status monitoring data of a target device; analyzing statistical characteristics of the status monitoring data, and identifying key features in the operation status monitoring data based on the statistical characteristics; where the key features are identified by using a pre-constructed feature extraction model, and the training sample set for training the feature extraction model is adjusted based on the result of anomaly assessment; simulating the quantum state of the key features, and compressing the key features based on the quantum state amplitude of the quantum state to generate a parameter to be measured; using a pre-constructed classifier model to perform classification prediction on the parameter to be measured, and outputting a device status prediction result of the parameter to be measured; determining a predicted failure of the target device according to the device status prediction result.

[0006] In combination with the first aspect, the embodiments of the present invention provide a first implementation manner of the first aspect. Among them, the steps of determining the predicted fault of the target device according to the device state prediction result include: obtaining the device prediction state indicated by the device state prediction result; determining the device prediction state as the predicted fault of the target device.

[0007] In combination with the first aspect, the embodiments of the present invention provide a second implementation manner of the first aspect. Among them, the steps of analyzing the statistical characteristics of the state monitoring data and identifying the key features in the running state monitoring data based on the statistical characteristics include: calculating the information gain of each feature of the state monitoring data; determining the statistical characteristics of the state monitoring data based on the information gain; determining the features with information gain exceeding the preset gain threshold as the key features in the running state monitoring data.

[0008] In combination with the first aspect, the embodiments of the present invention provide a third implementation manner of the first aspect. Among them, the steps of adjusting the training sample set of the training feature extraction model based on the result of the anomaly evaluation include: training a preset neural network with a preset training sample set to determine the historical mean and standard deviation of the training sample set; determining the anomaly evaluation result of the training sample set based on the historical mean and standard deviation; adjusting the feature values of the features of the training sample set based on the result of the anomaly evaluation.

[0009] In combination with the first aspect, the embodiments of the present invention provide a fourth implementation manner of the first aspect. Among them, the steps of simulating the quantum state of the key feature, compressing the key feature based on the quantum state amplitude of the quantum state, and generating the parameter to be measured include: converting the key feature into qubits and calculating the quantum noise level of the key feature; determining the probability amplitude corresponding to the qubits; compressing the key feature based on the probability amplitude and the quantum noise level.

[0010] In combination with the first aspect, the embodiments of the present invention provide a fifth implementation manner of the first aspect. Among them, the steps of compressing the key feature based on the probability amplitude and the quantum noise level include: converting the key feature into a low-dimensional feature representation through a preset autoencoder to compress the key feature; wherein, the training method of the autoencoder includes: training a preset autoencoder with a preset training sample set; evaluating the data complexity of the training sample set based on the information entropy of the training sample set; adjusting the dimension of the encoding layer of the autoencoder according to the data complexity and the preset encoding dimension adjustment coefficient; adjusting the output layer of the autoencoder based on the adjusted dimension of the encoding layer.

[0011] In combination with the first aspect, an embodiment of the present invention provides a sixth implementation manner of the first aspect. The method for constructing a classifier model includes: training a preset neural decision tree using a preset training sample set, and calculating a feature influence parameter matrix of the preset training sample set; calculating the feature influence weight of the training sample set based on the feature influence parameter matrix and the learning situation of the neural decision tree; determining the classification error of each decision path of the neural decision tree based on the feature influence weight; optimizing the decision rule of the neural decision tree and determining the matrix update amount of the feature influence parameter matrix based on the classification error; until the neural decision tree meets the preset training accuracy, freezing the parameters of the neural decision tree according to a preset parameter ratio, and constructing a classifier model.

[0012] In combination with the first aspect, an embodiment of the present invention provides a seventh implementation manner of the first aspect. The method further includes: obtaining a running state monitoring sample of a preset device, annotating the running state monitoring sample according to the corresponding fault state of the preset device, and constructing an initial sample set; using the generator of a preset generative adversarial network to perform sample augmentation on the initial sample set to generate an initial augmented sample; using the discriminator of the generative adversarial network to discriminate the initial sample set and the initial augmented sample respectively; updating the generator parameters of the generator based on the discrimination results of the initial sample set and the initial augmented sample to update the initial augmented sample; until the initial augmented sample meets the preset augmentation requirements, constructing a training sample set based on the initial sample set and the initial augmented sample.

[0013] In combination with the first aspect, an embodiment of the present invention provides an eighth implementation manner of the first aspect. The step of updating the generator parameters of the generator based on the discrimination results of the initial sample set and the initial augmented sample includes: determining the loss function of the generator based on the discrimination results of the initial sample set and the initial augmented sample; determining the fitness function of the generative adversarial network based on the dynamic change corresponding to the generative adversarial network; updating the generator parameters of the generator based on the loss function and the fitness function.

[0014] Second aspect, an embodiment of the present invention provides an apparatus for predictive maintenance of a device based on artificial intelligence. The apparatus includes: a data acquisition module for acquiring operation status monitoring data of a target device; a feature extraction module for analyzing the statistical characteristics of the status monitoring data and identifying key features in the operation status monitoring data. Among them, a pre-constructed feature extraction model is used to identify key features, and the training sample set for training the feature extraction model is adjusted based on the result of anomaly evaluation; a data processing module for simulating the quantum state of the key features, compressing the key features based on the quantum state amplitude of the quantum state to generate a parameter to be measured; an execution module for classifying and predicting the parameter to be measured by using a pre-constructed classifier model and outputting a prediction result of the device status of the parameter to be measured; an output module for determining a predicted fault of the target device according to the device status prediction result.

[0015] The embodiments of the present invention bring the following beneficial effects: An apparatus and method for predictive maintenance of a device based on artificial intelligence provided by the embodiments of the present invention identify the statistical characteristics of data, focus on the most significant features of changes, and adjust the focus of the model. Moreover, the training sample set of the feature extraction model is adjusted based on the result of anomaly evaluation, which can suppress abnormal data, protect the model from being interfered by these data, and optimize the effect of feature extraction. Quantum state modeling can better capture the non-linear relationship in the data, avoid information loss of traditional linear dimensionality reduction algorithms, and thus improve the expressiveness of the features after dimensionality reduction. The embodiments of the present invention improve the prediction accuracy and reliability of the model, ensure the data quality input to the classifier model, and reduce the risk of misjudgment through data compression based on quantum state amplitude and optimized feature extraction.

[0016] Other features and advantages of the present invention will be described in the following specification, and will, in part, be obvious from the specification, or be understood by implementing the present invention. The objectives and other advantages of the present invention are realized and obtained by the structures specifically pointed out in the specification and the drawings.

[0017] To make the above objectives, features, and advantages of the present invention more obvious and understandable, the following preferred embodiments are specifically described below in conjunction with the accompanying drawings. Description of the Drawings

[0018] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0019] Figure 1Flowchart of a method for predictive maintenance of devices based on artificial intelligence provided by an embodiment of the present invention;

[0020] Figure 2 Flowchart of another method for predictive maintenance of devices based on artificial intelligence provided by an embodiment of the present invention;

[0021] Figure 3 Flowchart of a method for constructing a training sample set provided by an embodiment of the present invention;

[0022] Figure 4 Schematic structural diagram of a device for predictive maintenance of devices based on artificial intelligence provided by an embodiment of the present invention;

[0023] Figure 5 Schematic structural diagram of an electronic device provided by an embodiment of the present invention. Detailed implementation manners

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following describes the implementation manners of the present disclosure through specific specific examples. Those skilled in the art can easily understand other advantages and effects of the present disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all the embodiments. The present disclosure can also be implemented or applied through other different specific implementation manners. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that, without conflict, the following embodiments and the features in the embodiments can be combined with each other. Based on the embodiments in the present disclosure, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present disclosure.

[0025] It should be noted that the following describes various aspects of the embodiments within the scope of the appended claims. It should be obvious that the aspects described in the present invention can be embodied in a wide variety of forms, and any specific structure and / or function described in the present invention is illustrative only. Based on the present disclosure, those skilled in the art should understand that one aspect described in the present invention can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects described in the present invention can be used to implement the device and / or practice the method. Additionally, this device and / or this method can be implemented using other structures and / or functions in addition to one or more of the aspects described in the present invention.

[0026] It should also be noted that the illustrations provided in the following embodiments only schematically illustrate the basic concept of the present disclosure. The diagrams only show the components related to the present disclosure, rather than being drawn according to the number, shape, and size of the components in actual implementation. The type, quantity, and proportion of each component in actual implementation can be arbitrarily changed, and the component layout type may also be more complex. Additionally, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0027] The present invention provides an artificial intelligence-based method and device for predictive maintenance of equipment, which can improve the prediction accuracy of the equipment fault state.

[0028] For ease of understanding, first, an artificial intelligence-based method for predictive maintenance of equipment provided by an embodiment of the present invention will be described. Figure 1 The flowchart of an artificial intelligence-based method for predictive maintenance of equipment provided by an embodiment of the present invention is shown. Referring to Figure 1 , the method includes the following steps:

[0029] Step S102, obtain the operation status monitoring data of the target equipment.

[0030] The data collection of the present invention is derived from the data obtained by various sensors during the operation of the equipment, including but not limited to temperature sensors, vibration sensors, pressure sensors, and current sensors. All the collected data is stored in a standardized JSON format. In one embodiment, the attributes of the data include: Ta (temperature), representing the working temperature of the equipment component; Va (vibration frequency), representing the vibration frequency generated during the operation of the equipment; Pa (pressure), representing the working pressure of the liquid or gas in the equipment; Ia (current), representing the magnitude of the current during the operation of the equipment; Ra (rotation speed), representing the rotation speed of the equipment component; Sa (sound level), representing the sound level generated during the operation of the equipment; Ha (humidity), representing the humidity of the surrounding environment of the equipment; Oa (oil quality), representing the quality status of the lubricating oil; Ua (usage duration), representing the operation duration of the equipment since the last maintenance; Ca (configuration parameters), representing the configuration parameters of the equipment operation. It should be noted that this embodiment is only to illustrate a data format and type of the present invention. In actual applications, the attributes of the data are usually more than 10 attributes, and the number of data attributes may reach dozens or even hundreds.

[0031] Step S104, analyze the statistical characteristics of the status monitoring data, and identify the key features in the operation status monitoring data based on the statistical characteristics.

[0032] In the embodiments of the present invention, detailed statistical analysis is performed on the state monitoring data to identify important statistics. In the embodiments of the present invention, by analyzing the statistical characteristics of the input data, the most significant features of the change are identified, the features that have an important impact on the change of the model output are captured, and according to the importance of the captured features, the focus of the network is adjusted to preferentially process those key features. Based on the statistical characteristics, the key features that can reflect the change of the device operation state are selected.

[0033] Among them, in the embodiments of the present invention, a pre-constructed feature extraction model is used to identify key features, and the training sample set for training the feature extraction model is adjusted based on the results of anomaly evaluation, and the corresponding features are suppressed to protect the model from the interference of these data and ensure the effect of feature extraction.

[0034] Step S106, simulate the quantum state of the key features, and based on the quantum state amplitudes of the quantum states, perform data compression on the key features to generate the parameters to be measured.

[0035] The quantum state amplitudes can significantly reduce the data dimension while ensuring the integrity of information, and improve the calculation efficiency. In the embodiments of the present invention, data compression is combined with the quantum state to better capture the non-linear relationships in the data and avoid the information loss of traditional linear dimensionality reduction algorithms.

[0036] Step S108, use the pre-constructed classifier model to perform classification prediction on the parameters to be measured, and output the prediction result of the device state of the parameters to be measured.

[0037] Step S110, determine the predicted fault of the target device according to the prediction result of the device state.

[0038] Among them, the predicted device state indicated by the prediction result of the device state can be obtained, and the predicted device state is determined as the predicted fault of the target device. In one embodiment, the collected original data is input into the trained feature extraction and feature dimensionality reduction model for feature processing. Further, the processed features are input into the classifier model for classification, and then the classification result is obtained. In this embodiment, the classification categories include the early-stage faults of "bearing wear", "motor overheating", "insufficient lubrication", etc., and correspondingly, they are output in the form of corresponding labels, such as the specific numbers (0-9) indicated by the labels.

[0039] In summary, the predictive maintenance method for devices based on artificial intelligence provided by the embodiments of the present invention identifies the statistical characteristics of data, focuses on the most significant changing features, and adjusts the model focus. Moreover, the training sample set of the feature extraction model is adjusted based on the anomaly evaluation results, which can suppress abnormal data, protect the model from being interfered by such data, and optimize the effect of feature extraction. Quantum state modeling can better capture the non-linear relationships in the data, avoid information loss of traditional linear dimensionality reduction algorithms, and thus enhance the expressiveness of the features after dimensionality reduction. The embodiments of the present invention improve the prediction accuracy and reliability of the model, ensure the data quality input to the classifier model, and reduce the risk of misjudgment through data compression based on quantum state amplitudes and optimized feature extraction.

[0040] Further, in combination with the above embodiments, the embodiments of the present invention also provide another predictive maintenance method for devices based on artificial intelligence. Figure 2 The flowchart of another predictive maintenance method for devices based on artificial intelligence provided by the embodiments of the present invention is shown. Referring to Figure 2 , this method includes the following steps:

[0041] Step S202, obtain the operation status monitoring data of the target device.

[0042] Step S204, calculate the information gain of each feature of the status monitoring data.

[0043] Step S206, determine the statistical characteristics of the status monitoring data based on the information gain.

[0044] Step S208, determine the key features in the operation status monitoring data for the features whose information gain exceeds the preset gain threshold.

[0045] In specific implementation, the embodiments of the present invention perform feature extraction on the operation status monitoring data through a preset feature extraction model to determine the key features in the operation status monitoring data. Specifically, feature selection is realized by real-time analyzing the statistical characteristics in the data stream, calculating the information gain of each feature, and selecting the features whose information gain exceeds the threshold, expressed as:

[0046]

[0047] In the formula, s i is the data feature, is the feature value corresponding to the data feature s i ; Y is the target variable, specifically the device fault status label, is the preset information gain threshold; is the information gain of the feature. Preferably, is set to 0.1.

[0048] Further, the present invention selects features based on information gain, and the information gain is implemented by mutual information. The calculation method is expressed as:

[0049]

[0050] In the formula, is the entropy of the target variable , is the conditional entropy of when the given feature is present.

[0051] Existing feature extraction methods are difficult to capture the changes of key features in real time, and lack a suppression strategy for abnormal data. They are easily affected by abnormal data, resulting in a decline in model performance and having an adverse impact on the accuracy of fault prediction. In the feature extraction process of the embodiments of the present invention, not only the statistical characteristics of features are concerned, but also an adaptive abnormal feature suppression strategy is used to enhance the robustness of the model to abnormal data for the dynamic suppression mechanism affected by abnormal data in the feature extraction model.

[0052] Further, the embodiments of the present invention also adjust the training sample set for training the feature extraction model based on the results of anomaly assessment. By identifying and adjusting the abnormal feature values that may lead to a decline in model performance, the influence is dynamically adjusted, thereby protecting the model from being interfered by these data. The method of anomaly assessment is expressed as:

[0053]

[0054] In the formula, and are respectively the historical mean and standard deviation of the feature values corresponding to the data feature ; is the anomaly assessment result. When it is close to 1, it represents normal data, and when it is close to 0, it represents abnormal data, and it is divided by a threshold of 0.5.

[0055] Feature suppression and adjustment based on the results of anomaly assessment are expressed as:

[0056]

[0057] In the formula, is a parameter dynamically adjusted based on the results of anomaly assessment.

[0058] Further, the calculation method of the parameter dynamically adjusted based on the results of anomaly assessment is expressed as:

[0059]

[0060] In the formula, is the minimum value function; is the maximum value function; is a threshold parameter for anomaly evaluation, used to determine the conversion boundary of features from anomaly to normal. Preferably, is set to 0.5.

[0061] Furthermore, the embodiments of the present invention also evaluate the feature uncertainty of the neural network for the training sample set. Based on the feature uncertainty, the network parameters of the neural network are updated. Until the neural network meets the preset training requirements, a feature extraction model is constructed.

[0062] Among them, the uncertainty estimation is realized through the Bayesian posterior variance. Specifically, the variance of the parameters of the neural network helps to control the sensitivity of the model to features with high uncertainty, and the calculation method is expressed as:

[0063]

[0064] In the formula, is the mean of the posterior distribution ; is the given data set, indicating all the data that the network has received; is the probability distribution function of the parameter under the condition of the given data set ; represents the differential of the parameter .

[0065] The uncertainty measure is used to evaluate the confidence level of the current model for a specific input. If the uncertainty of some features is too high, the model will automatically adjust its parameters to learn these features more finely and reduce the uncertainty of prediction. The way to adjust the network parameters is expressed as:

[0066]

[0067] In the formula, is the learning rate of the neural network; is the regularization learning rate of the neural network; is the L2 regularization; is the parameter of the neural network at the th iteration; is the parameter of the neural network at the th iteration; is the regularization coefficient of the neural network, represents the variance of the parameters of the neural network; is the posterior probability distribution function obtained when the input is and the parameter is ; ; is the input feature of the neural network at the th iteration.

[0068] Among them, the parameters of the neural network include the weights and biases of the neural network. In one embodiment, the initialization method is expressed as:

[0069]

[0070] In the formula, represents the parameters of the neural network model for feature extraction, represents the variance of the neural network initialization; represents being subject to a specific distribution; represents the normal distribution; represents the initial parameters of the neural network model for feature extraction.

[0071] Among them, the regularization learning rate of the neural network is set by an adaptive adjustment method, and the calculation method is expressed as:

[0072]

[0073] In the formula, is the regularization learning rate of the neural network at the (t + 1)-th iteration; is the regularization learning rate of the neural network at the t-th iteration; and are the decay rates of the estimated first and second moments at the t-th iteration, and are set to values close to 1.

[0074] In summary, repeat the above steps iteratively until the preset iteration stop condition is met, which indicates that the model training is completed. In one embodiment, the preset iteration stop condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0075] In summary, the embodiment of the present invention uses a 6-layer fully connected neural network for feature extraction. In the prior art, some solutions use a neural network for feature extraction. In some neural network structures, problems such as gradient disappearance, gradient explosion, or getting stuck in a local optimal solution may be encountered, affecting the training stability and the performance of the model. The present invention adopts a neural network algorithm based on dynamic feature capture, which can effectively extract key features that may indicate equipment failures and enhance the adaptability of the model to features of different fault types.

[0076] Step S210, simulate the quantum state of the key feature, and based on the quantum state amplitude of the quantum state, compress the key feature to generate the parameter to be measured.

[0077] Existing dimensionality reduction algorithms are limited by the linear constraints of the encoding layer, resulting in significant information loss. It is difficult to efficiently compress complex device data, and the feature expressiveness after dimensionality reduction is limited, affecting subsequent classification results. In the embodiments of the present invention, key features are converted into qubits, and the quantum noise level of the key features is calculated. The probability amplitude corresponding to the qubit is determined, and based on the probability amplitude and the quantum noise level, the key features are compressed. In quantum computing, a qubit can be in a superposition state between the 0 and 1 states, and its probability amplitude defines the probability of observing 0 or 1 if a measurement is made, including not only amplitude information but also phase information. Multiple quantum systems interact to form an entangled state, and their probability amplitudes jointly describe the quantum state of the entire system, reflecting the non-classical correlation between quantum systems. Quantum noise refers to any factor that interferes with ideal quantum operations and affects the evolution of the quantum state. Compressing data based on these parameters can retain the most critical information.

[0078] Specifically, in the embodiments of the present invention, a preset autoencoder is used to convert key features into low-dimensional feature representations and compress the key features. Among them, in the encoding process of the autoencoder in the embodiments of the present invention, the feature representation of the data is further compressed by simulating the amplitude of qubits to optimize the representation efficiency of the data. Specifically, the data features are further compressed using the amplitude of qubits, expressed as:

[0079]

[0080] In the formula, is the quantum probability amplitude function for simulating the quantum state; is the normalization constant to ensure that the sum of probabilities is 1; is the parameter that controls the compression sensitivity of the quantum state; is the th element of represents corresponding to the quantum noise level.

[0081] In the embodiments of the present invention, the quantum probability amplitude function is used to simulate the quantum state and determine the probability amplitude of the qubit. Among them, the quantum probability amplitude function assigns an importance weight to each feature in the form of a probability distribution. Important features are assigned higher weights, while unimportant features are compressed or even ignored, ensuring that the data after dimensionality reduction retains the most critical information.

[0082] Furthermore, the quantum noise level allows the model to have different compression levels in different parts of the feature representation, thereby more finely controlling information loss. The calculation method is expressed as:

[0083]

[0084] In the formula, It is a coefficient for adjusting the sensitivity of quantum noise response. By introducing randomness through quantum noise, overfitting of the model to the input data can be effectively avoided. Especially in the case of a high feature dimension, quantum noise improves the generalization performance of the model by suppressing non-critical features.

[0085] Among them, the embodiments of the present invention also train the autoencoder to ensure that the model can balance the importance of various types of data during the learning process. In a specific implementation, the training method of the autoencoder is as follows:

[0086] 1) Use a preset training sample set to train a preset autoencoder.

[0087] First, initialize the parameters of the autoencoder algorithm model. In one embodiment, a strategy based on quantum probability distribution is used for initializing the parameters of the autoencoder algorithm, expressed as:

[0088]

[0089]

[0090] In the formula, represents the weight of the autoencoder; represents the bias of the autoencoder; is the number of features input to the autoencoder; is the first randomly initialized parameter, is the second randomly initialized parameter; represents the imaginary unit, used to simulate the complex representation of the quantum state.

[0091] During the training process, the data input to the autoencoder can be the data features extracted by using the above-mentioned feature extraction model. It is converted into a low-dimensional feature representation through the encoder part, and the encoder learns how to map the high-dimensional input data to a low-dimensional latent space, expressed as:

[0092]

[0093] In the formula, is the low-dimensional feature representation encoded by the autoencoder; is the Sigmoid activation function; is the data vector input to the autoencoder.

[0094] Among them, the embodiments of the present invention use the reconstruction error and the asymmetric regularization term to calculate the loss, and adjust the network parameters through the backpropagation algorithm to minimize the overall loss. Then, the calculation method of the parameter update amount of the autoencoder is expressed as:

[0095]

[0096]

[0097] Wherein, and are the update amounts of the weights and biases of the autoencoder respectively; is the learning rate of the autoencoder; is the regularization coefficient for controlling the sparsity of the weights; represents the L2 norm; represents the L1 norm, which is used to achieve the sparsity of the weights; is the reconstructed data feature of the decoder in the autoencoder; is the j-th element of the weight vector of the autoencoder. Preferably, is set to 0.3.

[0098] 2) Based on the information entropy of the training sample set, evaluate the data complexity of the training sample set.

[0099] The complexity of the input data is evaluated through the information entropy of the input data, and the calculation method is expressed as:

[0100]

[0101] Wherein, represents the probability distribution of the i-th feature in

[0102] 3) According to the data complexity and the preset coding dimension adjustment coefficient, adjust the dimension of the coding layer of the autoencoder.

[0103] Define that the adjustment of the dynamic dimension of the coding layer is realized based on the complexity of the input data, and is expressed as:

[0104]

[0105] Wherein, is the basic coding dimension; is the coding dimension adjustment coefficient, which is used to control the sensitivity of the dimension adjustment; is the complexity of the input data.

[0106] 4) Based on the adjusted coding layer dimension, adjust the output layer of the autoencoder.

[0107] Use the dynamic dimension to adjust the output layer of the encoder, and it is expressed as:

[0108]

[0109] Wherein, is the dimensionality-reduced vector after dynamic dimension adjustment, which is input into the decoder of the autoencoder as the feature vector after feature dimensionality reduction for the next iteration; is the noise vector for the t-th iteration, which is used to enhance the robustness of the autoencoder model during training.

[0110] Furthermore, the adjustment of the noise vector is realized based on the current reconstruction error of the model, and the calculation method is expressed as:

[0111]

[0112] In the formula, is the basic noise level; is the noise adjustment coefficient, which is used to control the sensitivity of the error to the noise adjustment; is the model reconstruction error calculated in the t-th iteration. Preferably, is set to 0.01.

[0113] In summary, the embodiment of the present invention uses a dynamic coding space adjustment strategy to continuously adapt to the complexity and variability of input data during the iteration process by dynamically adjusting the dimension and configuration of the coding layer, so as to achieve more refined feature representation and compression. Repeat the above steps iteratively until the preset iteration stop condition is met, which means the model training is completed. In one embodiment, the preset iteration stop condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0114] In summary, the present invention adopts an autoencoding neural network algorithm based on quantum-inspired sparse coding as a feature dimensionality reduction model. On the basis of the traditional autoencoder algorithm, the probability amplitude of quantum bits is used to simulate the distribution of data features, so as to achieve efficient and meaningful feature dimensionality reduction. The feature dimensionality reduction of the traditional autoencoding network is often limited by the linear limitation and information loss of the coding layer. The present invention uses a quantum-inspired method to use a non-linear quantum probability distribution to enhance the coding efficiency, optimize the feature compression and reconstruction process, and improve the model's ability to process complex data.

[0115] Step S212, using the pre-constructed classifier model to classify and predict the parameter to be measured, and output the device state prediction result of the parameter to be measured.

[0116] Combined with the above embodiments, the existing classifier model still has the problem of a single parameter update method, which is prone to overfitting during training, lacks a dynamic adjustment mechanism, is difficult to adapt to the learning needs at different stages and the changes in the device operating state, and limits the classification accuracy and the generalization ability of the model. The present invention uses the neural decision tree algorithm as the classifier model. The neural decision tree combines the intuitive classification rules of the traditional decision tree and the non-linear learning ability of deep learning, optimizes the feature selection and decision-making path in the classification process. On the basis of the traditional neural decision tree, the present invention dynamically adjusts the update frequency and freezing state of the model parameters by monitoring the learning progress and classification effect of the model in real time, so as to adapt to the learning needs at different stages, reduce the risk of overfitting, and improve the generalization ability of the model.

[0117] Specifically, the construction method of the classifier model in the embodiment of the present invention is as follows:

[0118] 1) Use a preset training sample set to train a preset neural decision tree, and calculate the feature influence parameter matrix of the preset training sample set.

[0119] 2) Calculate the feature influence weights of the training sample set based on the feature influence parameter matrix and the learning situation of the neural decision tree.

[0120] First, initialize the structure and parameters of the neural decision tree. In one embodiment, the initialization method is expressed as:

[0121]

[0122]

[0123] In the formula, represents the weight of the neural decision tree; represents the dimension of the data input to the neural decision tree; represents the number of decision nodes of the neural decision tree; represents the bias of the neural decision tree; represents a random uniform distribution; represents being subject to a specific distribution; represents a vector of all zeros.

[0124] The data input to the neural decision tree is propagated forward. Each decision node calculates the decision condition according to its weight and activation function to determine the data flow direction. During the forward propagation process, the weights of each input feature in the model decision are dynamically adjusted to achieve real-time optimization of the influence of different features, so that the model can more accurately adapt to the characteristics of different device states, thereby improving the classification accuracy and the adaptability of the model.

[0125] Specifically, the influence of features is evaluated through a feature influence evaluation function. According to the current input features and the learning effect of historical data, the influence weight of each feature is calculated, and the calculation method is expressed as:

[0126]

[0127] In the formula, is the influence weight of the feature; is the feature vector corresponding to the data input to the neural decision tree; is the Softmax function; is the feature influence parameter matrix to be learned; is the feature influence bias term; is the Sigmoid activation function, which is used to ensure the non-linearity of the output and the expression ability of the model. The output of the Softmax function ensures that the sum of all feature weights is 1, making it interpretable as a probability distribution. Among them, the feature influence parameter matrix to be learned is constructed based on the feature influence weights in the training process. Both the feature influence parameter matrix to be learned and the feature influence bias term are learning parameters. For example, the parameters can be updated by the method of error backpropagation.

[0128] 3) Based on the feature influence weights, determine the classification error of each decision path of the neural decision tree.

[0129] The features input to the neural decision tree will be weighted by the corresponding influence weights to adjust the contribution of each feature to the decision node. The calculation method is expressed as:

[0130]

[0131] In the formula, represents the product of the corresponding elements, is the feature input of the weighted neural decision tree. By adjusting the influence of the input features, the model can respond more flexibly to important features, thereby improving the decision-making accuracy.

[0132] Furthermore, in the forward propagation process of the neural decision tree, the calculation method of the feature is expressed as:

[0133]

[0134] In the formula, represents the output of each decision node of the neural decision tree; is the activation function based on sensitivity enhancement.

[0135] In one embodiment, the activation function based on sensitivity enhancement enhances the non-linearity and adaptability of the model by controlling the curvature of the activation function. Let the input of the activation function based on sensitivity enhancement For , the calculation method of the activation function based on sensitivity enhancement is expressed as:

[0136]

[0137] In the formula, controls the curvature of the activation function and enhances the sensitivity of the model to input changes; is the threshold parameter, which determines the activation offset.

[0138] Furthermore, the effect of each decision path is evaluated by using the data input to the neural decision tree. In one embodiment, the loss function of the neural decision tree uses a regularization term to prevent overfitting and enhance the adaptability of the model to new data. The calculation method is expressed as:

[0139]

[0140] In the formula, is the number of training samples currently input to the neural decision tree, is the true label of the i-th sample, is the label of the i-th sample predicted by the neural decision tree model, is the regularization coefficient, is the -th column of the weight matrix of the neural decision tree; is the L2 norm; is the information entropy calculation function.

[0141] In the embodiment of the present invention, by calculating the information entropy of each classification in each iteration and adjusting the loss function accordingly, the processing ability of the model for uncertain and unbalanced data is strengthened, and the classification effect is further improved. The calculation method of the information entropy calculation function is expressed as:

[0142]

[0143] In the formula, is the prediction probability of the neural decision tree model for the k-th class, is the total number of categories of the samples input to the neural decision tree.

[0144] 4) Based on the classification error, optimize the decision rules of the neural decision tree and determine the matrix update amount of the feature influence parameter matrix.

[0145] In the embodiments of the present invention, the parameters of the neural decision tree are optimized according to the classification error. The neural decision tree combines the intuitive classification rules of the traditional decision tree and the non-linear learning ability of deep learning, and can not only implement the decision path implementation mode based on the traditional decision tree, but also combine the non-linear feature transformation mode of deep learning to realize the construction of classification rules. Specifically, the gradient descent method is used to update the weights and biases of the neural decision tree, and the calculation method of the update amount is expressed as:

[0146]

[0147]

[0148] In the formula, and are the update amounts of the weights and biases of the neural decision tree respectively; is the learning rate of the neural decision tree; is the loss function of the neural decision tree.

[0149] As the number of iterations increases, the learning rate of the neural decision tree is dynamically adjusted to improve the convergence speed of the model. The adjustment method is expressed as:

[0150]

[0151] is the learning rate of the neural decision tree at the th iteration, is the initial learning rate of the neural decision tree, The learning rate of the neural decision tree controls the attenuation rate. Preferably, is set to 0.01, is set to 0.3.

[0152] The update of the feature influence parameter matrix to be learned depends on the error between the model output and the target output, and is updated using the backpropagation algorithm. The calculation method of the update amount is expressed as:

[0153]

[0154] In the formula, is the update amount of the feature influence parameter matrix to be learned.

[0155] 5) Until the neural decision tree meets the preset training accuracy, freeze the parameters of the neural decision tree according to the preset parameter ratio to construct a classifier model.

[0156] When the model performs well, gradually freeze some parameters to reduce the complexity of training and prevent overfitting. Specifically, when the training accuracy of the model reaches the preset threshold, the calculation method of the proportion of frozen parameters is expressed as:

[0157]

[0158] In the formula, is the parameter ratio of the parameter-frozen neural decision tree in the t-th iteration; controls the freezing rate, and gradually increases the ratio of frozen parameters with the number of iterations to stabilize the training process; is the current number of iterations.

[0159] Repeat the above steps iteratively until the preset iteration stop condition is met, which indicates that the model training is completed. In one embodiment, the preset iteration stop condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.

[0160] Step S214, determine the predicted fault of the target device according to the device status prediction result.

[0161] The embodiment of the present invention uses the trained model to process new samples to achieve predictive maintenance of the device.

[0162] In summary, another device predictive maintenance method based on artificial intelligence provided by the embodiment of the present invention uses a neural network based on dynamic feature capture for feature extraction, identifies key features by real-time analyzing the statistical characteristics of input data, preferentially processes features that have an important impact on the prediction result to improve the sensitivity of the model to the fault state. In addition, through an adaptive abnormal feature suppression strategy, the impact of abnormal data is dynamically suppressed to protect the model stability, and the ability of the model to handle high-uncertainty data is further enhanced through uncertainty measure evaluation.

[0163] In addition, the quantum-inspired sparse coding autoencoder algorithm is used for feature dimension reduction, and the efficient compression of data features is realized by simulating the probability distribution of quantum bits, which solves the problems of large information loss and low coding efficiency in traditional dimension reduction methods. In addition, through a dynamic coding space adjustment strategy, the dimension of the coding layer and the noise response sensitivity are dynamically adjusted according to the complexity of the input data, effectively enhancing the adaptability of the model to different complexity data.

[0164] The neural decision tree model is also used as a classifier. By combining the regularity of the decision tree and the non-linear learning ability of deep learning, the optimization of the fault classification process is realized, and the rationality of feature selection and decision path in the classification process is improved. In addition, through a parameter dynamic adjustment and freezing mechanism, the learning progress of the model is monitored in real time, the parameter update frequency and freezing state are dynamically adjusted, enhancing the generalization ability of the model and reducing the overfitting phenomenon.

[0165] Further, the embodiment of the present invention trains the above model by using a preset training sample set. In addition, the embodiment of the present invention also provides another method for predictive maintenance of equipment based on artificial intelligence, and mainly describes the construction method of the training sample set. Figure 3 shows a flowchart of a method for constructing a training sample set provided by an embodiment of the present invention. Refer to Figure 3 , the method includes the following steps:

[0166] Step S10, obtain the operation status monitoring samples of the preset equipment, and label the operation status monitoring samples according to the corresponding fault status of the preset equipment to construct an initial sample set.

[0167] The operation status monitoring samples can refer to the above embodiments. Further, the embodiment of the present invention also labels the collected data. The labeling method of the present invention is manual labeling. In one embodiment, the labeling categories include the early-stage faults of bearing wear, the early-stage faults of motor overheating, the early-stage faults of insufficient lubrication, and other early-stage states of various faults.

[0168] Step S11, use the generator of the preset generative adversarial network to expand the initial sample set to generate initial expanded samples.

[0169] It can be understood that in the task of the present invention, the acquisition, labeling, and preprocessing of training data are time-consuming and laborious, and insufficient training samples are likely to lead to poor generalization ability of the model and affect the accuracy of the model at the same time. The existing data augmentation methods mainly rely on a small amount of sampling technology, which cannot effectively augment the equipment operation data, and the quality and diversity of the generated data are not high, resulting in insufficient generalization ability and prediction accuracy of the model. The present invention uses a generative adversarial network for sample generation, and then realizes data augmentation. On the basis of the traditional generative adversarial network, the mutation mechanism of the evolutionary algorithm is used to optimize the learning process of the generator and discriminator in the generative adversarial network, so as to improve the quality and diversity of the generated data. In addition, the game learning strategy is used to further refine the adversarial process of the generator and discriminator, so that each update more accurately reflects the complexity of the data set and its inherent distribution characteristics.

[0170] Among them, the parameters of the generator and discriminator are randomly initialized. In one embodiment, the initialization method is expressed as:

[0171]

[0172] In the formula, represents the parameters of the generator, represents the parameters of the discriminator, represents the initialized variance; represents being subject to a specific distribution; Represents the variance of the normal distribution for initializing the parameters of the generative adversarial network. Preferably, It is set to 0.01.

[0173] Step S12: Use the discriminator of the generative adversarial network to discriminate the initial sample set and the initial augmented samples respectively.

[0174] In the standard training framework of the generative adversarial network, the generator attempts to generate new data sufficient to deceive the discriminator, and the discriminator tries to distinguish between real data and generated data. The loss function of the traditional generative adversarial network is expressed as:

[0175]

[0176] In the formula, Represents real data, Represents the input from the random noise distribution ; Represents the random noise distribution; Represents the expectation; Is the binary part of the real data; Represents the discriminator; Represents the generator; Represents the discriminator function; Represents the generator function; Is the determination output of the discriminator for real data, Is the fake data output by the generator.

[0177] Based on the traditional generative adversarial network, the present invention is optimized using a game learning strategy, enabling the two networks to continuously improve their strategies and counter-strategies during the training process. Specifically, by fully considering multiple discrimination strategies in the discrimination process of the discriminator, the discrimination strategies for real data and the data generated by the generator are different. Based on this, the loss function is calculated, and through the constraint of the loss function during the training process, the game confrontation ability can be improved, and the ability of the generator can be enhanced. Define the discrimination process of the discriminator for real data as:

[0178]

[0179] In the formula, Is the scoring function of the discriminator for real data, Are the parameters of the scoring function of the discriminator for real data.

[0180] Furthermore, the implementation manner of the scoring function of the discriminator for real data is expressed as:

[0181]

[0182] In the formula, The number of real data samples input into the discriminator for the current batch; The hyperparameter of the scoring function of the discriminator for real data; The hyperbolic tangent function; The weight training parameter of the scoring function of the discriminator for real data; Is The transpose of the parameter; The bias training parameter of the scoring function of the discriminator for real data.

[0183] Moreover, the discrimination process of the discriminator for the generated data is defined as:

[0184]

[0185] In the formula, Is the scoring function of the discriminator for the generated data, Is the parameter of the scoring function of the discriminator for the generated data.

[0186] Furthermore, the calculation method of the scoring function of the discriminator for the generated data is expressed as:

[0187]

[0188] In the formula, Is the hyperparameter of the scoring function of the discriminator for the generated data; Is the weight training parameter of the scoring function of the discriminator for the generated data; Is The transpose of the parameter; Is the bias training parameter of the scoring function of the discriminator for the generated data; Is the number of generated data samples generated by the generator input into the discriminator for the current batch.

[0189] In summary, the discriminator in the embodiment of the present invention uses different discrimination functions for real data and generated data, enabling the discriminator to more accurately model data with different properties. The distributions of real data and generated data may have essential differences. Therefore, designing appropriate discrimination functions separately can better capture these differences and improve the discrimination ability. At the same time, the authenticity of the generated data is examined from multiple perspectives, thereby improving the robustness of the entire adversarial process. The authenticity of the generated data can be examined from multiple perspectives, thereby improving the robustness of the entire adversarial process.

[0190] Step S13: Update the generator parameters of the generator based on the discrimination results of the initial sample set and the discrimination results of the initial augmented samples to update the initial augmented samples.

[0191] In the embodiments of the present invention, the loss function of the generator is determined based on the discrimination results of the initial sample set and the discrimination results of the initial augmented samples; the fitness function of the generative adversarial network is determined based on the dynamic changes corresponding to the generative adversarial network; and the generator parameters of the generator are updated based on the loss function and the fitness function.

[0192] The evolutionary strategy is used to adjust the parameters of the generator to improve its generation quality. The evolutionary algorithm evaluates and selects the best individual, expressed as:

[0193]

[0194] In the formula, is the generator parameter at the t-th iteration, is the generator parameter at the (t + 1)-th iteration; is the learning rate of the generator at the t-th iteration, is the loss function of the generative adversarial network; represents the gradient with respect to the generator parameters; is the fitness function of the generative adversarial network. In the traditional generative adversarial network during the adversarial training process, a fixed learning rate is usually used for training, which may lead to too large a learning rate in the initial stage of training and unable to be finely adjusted, or too small a learning rate in the later stage of training and slow convergence speed. To solve this problem, the present invention adopts a dynamic learning rate adjustment mechanism based on loss feedback to adjust the dynamic learning rate, expressed as:

[0195]

[0196] In the formula, is the base learning rate, represents the change in the loss of the discriminator in two consecutive iterations; is the learning rate adjustment factor of the generative adversarial network, used to control the speed of learning rate reduction. Preferably, is set to 0.01, is set to 0.3.

[0197] The fitness function of the generative adversarial network aims to prompt the generator to not only deceive the discriminator but also increase the diversity of the generated samples. The calculation method is expressed as:

[0198]

[0199] In the formula, is the parameter controlling the fitness influence; is the mutation rate at the t-th iteration; represents the variance. Preferably, Set to 0.3. Automatically adjust the mutation rate according to the training results of the previous iteration to adapt to the dynamic changes during the training process, optimize the exploration and exploitation balance of the algorithm, and prevent premature convergence to local optimal solutions. The specific adjustment strategy is as follows:

[0200]

[0201] In the formula, is the mutation rate of the (t + 1)-th iteration, is the mutation rate adjustment coefficient; is the evaluation index, which can specifically be set as the diversity index of the generated data, is the target value of the preset evaluation index. Preferably, is set to 0.2. Select the individual with the best performance based on the fitness function for reproduction to generate a new generation of network parameters, and at the same time eliminate the individuals with poor performance, which is expressed as:

[0202]

[0203] In the formula, are the parameters of the generator corresponding to the individuals with good performance selected based on the fitness function; is the selection function, which selects the parameter set with the best performance from the current population according to the fitness function for update.

[0204] Step S14, until the initial augmented samples meet the preset augmentation requirements, construct a training sample set based on the initial sample set and the initial augmented samples.

[0205] Repeat the above steps iteratively until the preset stop iteration condition is met, which means the model training is completed. In one embodiment, the preset stop iteration condition is to reach the preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times. After the data augmentation model training is completed, use the trained data augmentation model to increase the number of samples. In one embodiment, assume the original collected samples are 800, and the data augmentation model augments and generates 200 samples, then the augmented data set contains 1000 samples. It should be noted that the trained sample set can be used to train the feature extraction model. Further, the autoencoder can be trained after feature extraction of the trained sample set. The trained autoencoder can also be used to compress the data of its training sample set and further train the classifier model.

[0206] In summary, another artificial intelligence-based equipment predictive maintenance method provided by an embodiment of the present invention adopts a generative adversarial network in conjunction with an evolutionary algorithm for data expansion, and dynamically adjusts the learning process of the generator and the discriminator through an evolutionary strategy, thereby solving the problem of poor model generalization ability and insufficient precision caused by insufficient training data. In addition, the generator and discriminator of the generative adversarial network are optimized through a game learning strategy, so that the generated data is more in line with the complex distribution characteristics of actual equipment data, thereby improving the quality and diversity of generated samples.

[0207] On the basis of the above embodiments, the present invention also provides an artificial intelligence-based equipment predictive maintenance device. Figure 4 A schematic diagram of the structure of an artificial intelligence-based equipment predictive maintenance device provided by an embodiment of the present invention is shown, referring to Figure 4 The device includes: a data acquisition module 100, which is used to acquire the operating status monitoring data of the target device; a feature extraction module 200, which is used to analyze the statistical characteristics of the status monitoring data and identify the key features in the operating status monitoring data; a data processing module 300, which is used to simulate the quantum state of the key features, and based on the quantum state amplitude of the quantum state, the key features are compressed to generate the parameters to be measured; an execution module 400, which is used to classify and predict the parameters to be measured using a pre-built classifier model, and output the device status prediction results of the parameters to be measured; an output module 500, which is used to determine the predicted fault of the target device according to the device status prediction results.

[0208] An artificial intelligence-based equipment predictive maintenance device provided in an embodiment of the present invention has the same technical features as an artificial intelligence-based equipment predictive maintenance method provided in the above embodiment, so it can also solve the same technical problems and achieve the same technical effects.

[0209] The output module 500 is also used to determine the predicted fault of the target device according to the device state prediction result, including: obtaining the device predicted state indicated by the device state prediction result; and determining the device predicted state as the predicted fault of the target device.

[0210] The feature extraction module 200 is also used to calculate the information gain of each feature of the state monitoring data; based on the information gain, determine the statistical characteristics of the state monitoring data; and determine the features whose information gain exceeds the preset gain threshold as key features in the operating state monitoring data. The feature extraction module 200 is also used to train the preset neural network using the preset training sample set, determine the historical mean and standard deviation of the training sample set; determine the abnormal evaluation results of the training sample set based on the historical mean and standard deviation; and adjust the feature values ​​of the features of the training sample set based on the results of the abnormal evaluation.

[0211] The above data processing module 300 is further configured to convert the key features into qubits, calculate the quantum noise level of the key features; determine the probability amplitude corresponding to the qubits; and compress the key features based on the probability amplitude and the quantum noise level. The above data processing module 300 is further configured to convert the key features into low-dimensional feature representations through a preset autoencoder, and perform data compression on the key features; wherein, the training method of the autoencoder includes: training a preset autoencoder by using a preset training sample set; evaluating the data complexity of the training sample set based on the information entropy of the training sample set; adjusting the dimension of the encoding layer of the autoencoder according to the data complexity and a preset encoding dimension adjustment coefficient; and adjusting the output layer of the autoencoder based on the adjusted dimension of the encoding layer.

[0212] The above execution module 400 is further configured to train a preset neural decision tree by using a preset training sample set, and calculate a feature influence parameter matrix of the preset training sample set; calculate the feature influence weight of the training sample set based on the feature influence parameter matrix and the learning situation of the neural decision tree; determine the classification error of each decision path of the neural decision tree based on the feature influence weight; optimize the decision rule of the neural decision tree based on the classification error, and determine the matrix update amount of the feature influence parameter matrix; until the neural decision tree meets the preset training accuracy, freeze the parameters of the neural decision tree according to a preset parameter ratio, and construct a classifier model. The above execution module 400 is further configured to obtain the operation state monitoring samples of a preset device, label the operation state monitoring samples according to the fault state corresponding to the preset device, and construct an initial sample set; use the generator of a preset generative adversarial network to perform sample augmentation on the initial sample set to generate initial augmented samples; use the discriminator of the generative adversarial network to discriminate the initial sample set and the initial augmented samples respectively; update the generator parameters of the generator based on the discrimination results of the initial sample set and the initial augmented samples to update the initial augmented samples; until the initial augmented samples meet the preset augmentation requirements, construct a training sample set based on the initial sample set and the initial augmented samples. The above execution module 400 is further configured to determine the loss function of the generator based on the discrimination results of the initial sample set and the initial augmented samples; determine the fitness function of the generative adversarial network based on the dynamic changes corresponding to the generative adversarial network; and update the generator parameters of the generator based on the loss function and the fitness function.

[0213] An embodiment of the present invention further provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of the method shown above are implemented. Figures 1 to 3 An embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is run by a processor, the steps of the method shown above are executed.Figures 1 to 3 Steps of the method shown. An embodiment of the present invention also provides a schematic structural diagram of an electronic device, as Figure 5 shown, which is a schematic structural diagram of the electronic device. Among them, the electronic device includes a processor 51 and a memory 50. The memory 50 stores computer-executable instructions that can be executed by the processor 51. The processor 51 executes the computer-executable instructions to implement the above Figures 1 to 3 shown method. In Figure 5 the embodiment shown, the electronic device further includes a bus 52 and a communication interface 53. Among them, the processor 51, the communication interface 53, and the memory 50 are connected through the bus 52. Among them, the memory 50 may include a high-speed random access memory (RAM, Random Access Memory), and may also include a non-volatile memory, such as at least one disk memory. Through at least one communication interface 53 (which can be wired or wireless), a communication connection is realized between the system network element and at least one other network element, and the Internet, wide area network, local area network, metropolitan area network, etc. can be used. The bus 52 can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc., and can also be an AMBA (Advanced Microcontroller Bus Architecture) bus. Among them, AMBA defines three buses, including an APB (Advanced Peripheral Bus) bus, an AHB (Advanced High-performance Bus) bus, and an AXI (Advanced eXtensible Interface) bus. The bus 52 can be divided into an address bus, a data bus, a control bus, etc. For the convenience of representation, Figure 5It is represented only by a bidirectional arrow, but it does not mean that there is only one bus or one type of bus. The processor 51 may be an integrated circuit chip with signal processing capabilities. In the implementation process, the steps of the above method can be completed by the integrated logic circuit of the hardware in the processor 51 or the instructions in the form of software. The above-mentioned processor 51 may be a general-purpose processor, including a central processing unit (CPU for short), a network processor (NP for short), etc.; it may also be a digital signal processor (DSP for short), an application specific integrated circuit (ASIC for short), a field-programmable gate array (FPGA for short), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as being executed and completed by a hardware decoding processor, or executed and completed by a combination of hardware and software modules in the decoding processor. The software module may be located in a mature storage medium in the art such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor 51 reads the information in the memory and combines its hardware to complete the foregoing Figures 1 to 3 any of the shown methods.

[0214] A computer program product of a device predictive maintenance method and device based on artificial intelligence provided by an embodiment of the present invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the method described in the foregoing method embodiments. For the specific implementation, reference can be made to the method embodiments and will not be elaborated herein. Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working process of the above-described system can refer to the corresponding process in the foregoing method embodiments and will not be elaborated herein. In addition, in the description of the embodiments of the present invention, unless otherwise clearly defined and limited, the terms "installation", "connection", and "connection" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the communication inside two components. For those skilled in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations. If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or part of this technical solution can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program code. In the description of the present invention, it should be noted that the orientation or positional relationship indicated by the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation to the present invention. In addition, the terms "first", "second", and "third" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.Finally, it should be noted that the above embodiments are only specific embodiments of the present invention, used to illustrate the technical solutions of the present invention, rather than limiting it. The protection scope of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art within the technical scope disclosed by the present invention can still modify the technical solutions recorded in the foregoing embodiments or can easily think of changes, or perform equivalent replacements on some of the technical features; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. An artificial intelligence-based equipment predictive maintenance method, characterized in that: The method comprises: Obtain the operating status monitoring data of the target device; Analyzing the statistical characteristics of the state monitoring data, and identifying key features in the operating state monitoring data based on the statistical characteristics; wherein the key features are identified using a pre-built feature extraction model, and a training sample set for training the feature extraction model is adjusted based on the result of the abnormality assessment; Simulating the quantum state of the key feature, and based on the quantum state amplitude of the quantum state, performing data compression on the key feature to generate a parameter to be measured; Using a pre-built classifier model to classify and predict the parameters to be measured, and output the device state prediction results of the parameters to be measured; Determining a predicted fault of the target device according to the device status prediction result; The step of simulating the quantum state of the key feature, compressing the data of the key feature based on the quantum state amplitude of the quantum state, and generating the parameter to be measured includes: Converting the key feature into a quantum bit and calculating the quantum noise level of the key feature; Determining a probability amplitude corresponding to the quantum bit; The key features are compressed based on the probability amplitude and the quantum noise level.

2. The method according to claim 1, characterized in that The step of determining the predicted fault of the target device according to the device state prediction result includes: Acquire the predicted state of the device indicated by the device state prediction result; The predicted device state is determined as a predicted failure of the target device.

3. The method according to claim 1, characterized in that The step of analyzing the statistical characteristics of the condition monitoring data and identifying key features in the operating condition monitoring data based on the statistical characteristics comprises: Calculating the information gain of each feature of the condition monitoring data; Based on the information gain, determining the statistical characteristics of the condition monitoring data; The feature whose information gain exceeds a preset gain threshold is determined as a key feature in the operating status monitoring data.

4. The method according to claim 3, characterized in that The step of adjusting the training sample set for training the feature extraction model based on the result of the abnormality assessment comprises: Using a preset training sample set to train a preset neural network, and determining a historical mean and a standard deviation of the training sample set; Determining an abnormality assessment result of the training sample set based on the historical mean and standard deviation; Based on the result of the abnormality assessment, feature values ​​of the features of the training sample set are adjusted.

5. The method according to claim 1, characterized in that The step of compressing the key features based on the probability amplitude and the quantum noise level comprises: The key features are converted into low-dimensional feature representations by a preset autoencoder, and data compression is performed on the key features; The training method of the autoencoder includes: Using a preset training sample set to train a preset autoencoder; Based on the information entropy of the training sample set, evaluating the data complexity of the training sample set; Adjusting the encoding layer dimension of the autoencoder according to the data complexity and a preset encoding dimension adjustment coefficient; Based on the adjusted encoding layer dimension, an output layer of the autoencoder is adjusted.

6. The method according to claim 1, characterized in that The method for constructing the classifier model comprises: Using a preset training sample set to train a preset neural decision tree, and calculating a feature influence parameter matrix of the preset training sample set; Based on the feature influence parameter matrix and the learning status of the neural decision tree, calculating the feature influence weight of the training sample set; Determining the classification error of each decision path of the neural decision tree based on the feature influence weight; Based on the classification error, optimizing the decision rule of the neural decision tree, and determining the matrix update amount of the feature influence parameter matrix; Until the neural decision tree meets the preset training accuracy, the parameters of the neural decision tree are frozen according to the preset parameter ratio to construct a classifier model.

7. The method according to claim 1, characterized in that The method further comprises: Acquire operation status monitoring samples of preset devices, annotate the operation status monitoring samples according to the fault status corresponding to the preset devices, and construct an initial sample set; Performing sample expansion on the initial sample set using a preset generator of a generative adversarial network to generate initial expanded samples; Using the discriminator of the generative adversarial network to discriminate the initial sample set and the initial expanded samples respectively; Based on the discrimination result of the initial sample set and the discrimination result of the initial expanded sample, the generator parameter of the generator is updated to update the initial expanded sample; Until the initial expanded samples meet the preset expansion requirements, a training sample set is constructed based on the initial sample set and the initial expanded samples.

8. The method according to claim 7, characterized in that The step of updating the generator parameters of the generator based on the discrimination result of the initial sample set and the discrimination result of the initial expanded sample comprises: Determining the loss function of the generator based on the discrimination result of the initial sample set and the discrimination result of the initial expanded sample; Determining a fitness function of the generative adversarial network based on dynamic changes corresponding to the generative adversarial network; Based on the loss function and the fitness function, the generator parameters of the generator are updated.

9. An artificial intelligence-based equipment predictive maintenance device, characterized in that: The device comprises: A data acquisition module is used to acquire the operating status monitoring data of the target device; A feature extraction module, used to analyze the statistical characteristics of the state monitoring data and identify key features in the operating state monitoring data; wherein the key features are identified using a pre-built feature extraction model, and a training sample set for training the feature extraction model is adjusted based on the result of anomaly assessment; A data processing module, used for simulating the quantum state of the key feature, and performing data compression on the key feature based on the quantum state amplitude of the quantum state to generate a parameter to be measured; An execution module, used to classify and predict the parameters to be measured using a pre-built classifier model, and output a device state prediction result of the parameters to be measured; An output module, used to determine the predicted fault of the target device according to the device state prediction result; Wherein, the data processing module is also used to convert the key feature into a quantum bit, calculate the quantum noise level of the key feature; determine the probability amplitude corresponding to the quantum bit; and compress the key feature based on the probability amplitude and the quantum noise level.

Images