Intelligent detection methods and systems for industrial equipment operating performance

Abstract

This invention discloses an intelligent detection method and system for the operational performance of industrial equipment, relating to the field of industrial equipment detection. The method includes: collecting a first vibration sample dataset; applying a random resampling transformation to the vibration signal samples to simulate rotational speed changes; and pre-training a prior feature extractor; constructing a fault classifier and a rotational speed discriminator; using a second vibration sample dataset, jointly training and optimizing the prior feature extractor through an adversarial domain adaptation approach to obtain a target fault classifier and a target feature extractor; and collecting real-time vibration signals, inputting them into the target feature extractor for feature extraction, and then inputting them back into the target fault classifier to obtain equipment operation detection information. This invention addresses the technical problems of inaccurate feature extraction, poor model generalization ability, and unreliable detection results in existing equipment operation performance detection methods when dealing with rotational speed changes and differences in data distribution under different operating conditions. It achieves the technical effect of improving the accuracy of feature extraction and the generalization ability of the model, thereby improving the reliability of the detection results.

Description

Technical Field

[0001] This application relates to the field of industrial equipment testing, and in particular to intelligent testing methods and systems for the operational performance of industrial equipment. Background Technology

[0002] In industrial production, the stability of industrial equipment operation directly affects production efficiency, product quality, and production safety. Therefore, accurate detection of industrial equipment operation performance is of great significance. Currently, the solution to the problem of industrial equipment operation performance detection mainly relies on a combination of traditional signal processing techniques and conventional machine learning models. This approach first uses signal processing techniques to extract features from the equipment's vibration signals, and then inputs the extracted features into a machine learning model for fault classification and performance evaluation. However, due to the complex and variable operating conditions of industrial equipment and frequent speed fluctuations, traditional signal processing techniques struggle to effectively handle such non-stationary signals. Specifically, when speed changes, traditional methods based on fixed-frequency analysis cannot accurately capture fault features, resulting in inaccurate and incomplete feature extraction. Moreover, conventional machine learning models have limited generalization ability when faced with differences in data distribution under different operating conditions, making it difficult to maintain stable detection performance under different speeds and loads, leading to false positives and false negatives, thus affecting the accuracy and reliability of equipment operation performance detection.

[0003] At present, the detection methods for the operating performance of industrial equipment have technical problems such as inaccurate feature extraction and poor model generalization ability when dealing with changes in rotational speed and differences in data distribution under different operating conditions, resulting in unreliable equipment operation detection results. Summary of the Invention

[0004] This application provides an intelligent detection method and system for the operational performance of industrial equipment. It employs a first vibration sample dataset of the industrial equipment under inspection, performs random resampling transformation on the vibration signal samples in this dataset to simulate rotational speed changes, and pre-trains a prior feature extractor to construct a fault classifier and a rotational speed discriminator. Using a second vibration sample dataset, it jointly trains and optimizes the prior feature extractor through adversarial domain adaptation, achieving two training objectives: obtaining a target fault classifier and a target feature extractor. The system also collects real-time vibration signals from the industrial equipment under inspection, first inputting these signals into the target feature extractor to extract features, and then inputting them into the target fault classifier to obtain equipment operation detection information. These technical means address the technical problems of inaccurate feature extraction and poor model generalization ability in existing industrial equipment operational performance detection methods when dealing with rotational speed changes and differences in data distribution under different operating conditions, leading to unreliable equipment operation detection results. This method achieves the technical effect of improving the accuracy of feature extraction and the generalization ability of the model, thereby improving the reliability of equipment operation detection results.

[0005] This application provides an intelligent detection method for the operating performance of industrial equipment, comprising: collecting a first vibration sample dataset of the industrial equipment to be inspected; applying a random resampling transformation to the vibration signal samples in the first vibration sample dataset to simulate rotational speed changes; pre-training a prior feature extractor; constructing a fault classifier and a rotational speed discriminator; and jointly training and optimizing the prior feature extractor using a second vibration sample dataset through an adversarial domain adaptation approach to obtain a target fault classifier and a target feature extractor; wherein, the first training objective of the adversarial domain adaptation is to optimize the prior feature extractor and the fault classifier to maximize the fault classification accuracy, and the second training objective is to optimize the prior feature extractor to minimize the rotational speed classification accuracy of the rotational speed discriminator; real-time acquisition of the real-time vibration signal of the industrial equipment to be inspected, inputting it into the target feature extractor for feature extraction, and then inputting it into the target fault classifier to obtain equipment operation detection information.

[0006] In one possible implementation, a first vibration sample dataset of the industrial equipment to be inspected is collected. A random resampling transformation is applied to the vibration signal samples in the first vibration sample dataset to simulate rotational speed changes. A prior feature extractor is pre-trained, and the following processing is performed: The first vibration sample dataset contains multiple sets of vibration signal samples with labeled fault types collected at different rotational speeds; a model architecture of the prior feature extractor is constructed, and the model architecture of the prior feature extractor is pre-trained based on the first vibration sample dataset and a self-supervised learning task, so that the prior feature extractor has the ability to perceive rotational speed changes from vibration signals.

[0007] In a possible implementation, a model architecture based on a prior feature extractor is pre-trained using a self-supervised learning task, and the following processes are performed: a random resampling transformation is applied to the vibration signal samples input to the prior feature extractor to simulate the linear change of rotational speed; the transformed signal is input to the prior feature extractor to train an auxiliary rotational speed estimation network to predict the resampling scaling factor based on the features output by the prior feature extractor; and the pre-training of the prior feature extractor is completed by minimizing the prediction error of the scaling factor.

[0008] In a possible implementation, a fault classifier and a rotational speed discriminator are constructed, and the prior feature extractor is jointly trained and optimized using a second vibration sample dataset through an adversarial domain adaptation method to obtain a target fault classifier and a target feature extractor. The following processing is performed: the fault classifier and the rotational speed discriminator are configured using a multilayer perceptron respectively; based on the adversarial domain adaptation method, using the second vibration sample dataset as training data, the first training objective and the second training objective are alternately optimized using a gradient descent algorithm: Step 1: Fix the parameters of the prior feature extractor and the fault classifier, and update the parameters of the rotational speed discriminator; Step 2: Fix the parameters of the rotational speed discriminator, and jointly update the parameters of the prior feature extractor and the fault classifier to complete the joint adversarial optimization of the first training objective and the second training objective; Step 1 and Step 2 are iteratively executed, and the joint adversarial loss function is used for convergence judgment to obtain a target fault classifier and a target feature extractor whose joint adversarial loss meets a preset loss threshold.

[0009] In a possible implementation, the following processing is performed: the expression for the joint adversarial loss function is: ;in, For the joint adversarial loss function; The cross-entropy loss of the fault classifier, This refers to the cross-entropy loss of the speed discriminator; This is a positive hyperparameter used to balance the weights of the fault classification task and the rotation speed determination task in gradient updates.

[0010] In a possible implementation, the following processing is performed: the λ value is determined through a grid search, the grid search steps of which include: randomly generating multiple candidate λ values ​​in ascending order within the range of 0 to 1; for each candidate λ value, alternating training and testing using the gradient descent algorithm, recording the decrease in fault classification accuracy and the decrease in rotational speed discrimination accuracy corresponding to each candidate λ value; selecting candidate λ values ​​whose decrease in fault classification accuracy is less than a first preset decrease value and whose decrease in rotational speed discrimination accuracy exceeds a second preset decrease value as the λ value of the joint adversarial loss function; wherein, the second preset decrease value is greater than the first preset decrease value.

[0011] In a possible implementation, the following processing is performed: positive hyperparameters are determined by a dynamic scheduler, specifically including: using the λ value determined by grid search as the initial λ value; alternately optimizing the first training objective and the second training objective using a gradient descent algorithm, and collecting the decrease magnitude of real-time fault classification accuracy and the decrease magnitude of real-time speed discrimination accuracy during the training process; inputting the decrease magnitude of real-time fault classification accuracy and the decrease magnitude of real-time speed discrimination accuracy into the dynamic scheduler for dynamic adjustment of the λ value; wherein, the dynamic scheduler includes an increase threshold and a decrease threshold, the decrease threshold being the adjustment step size for decreasing when the decrease magnitude of real-time fault classification accuracy is greater than the first preset decrease magnitude, and the increase threshold being the adjustment step size for increasing when the decrease magnitude of real-time speed discrimination accuracy is less than the second preset decrease magnitude.

[0012] In a possible implementation, after obtaining the target fault classifier and the target feature extractor, the following processing is performed: using the target feature extractor, outputting pure fault features independent of rotational speed; reconstructing the pure fault features based on the vibration signal of the fault-related periodic impact component to generate a reconstructed enhanced fault impact signal; and adding the enhanced fault impact signal to the equipment operation detection information.

[0013] In a possible implementation, the pure fault features are reconstructed using vibration signals based on fault-related periodic impact components to generate a reconstructed enhanced fault impact signal. The following processes are then performed: a decoder-generator network is constructed; pure fault feature samples and reconstructed signal samples are collected; the decoder-generator network is trained using the pure fault feature samples and reconstructed signal samples to obtain a signal reconstruction network in which the signal reconstruction loss, periodic consistency loss, and fault classification loss all satisfy preset loss thresholds; the pure fault features are then reconstructed using the signal reconstruction network to generate the enhanced fault impact signal.

[0014] This application also provides an intelligent detection system for the operating performance of industrial equipment, comprising: a prior feature extractor pre-training module, used to collect a first vibration sample dataset of the industrial equipment to be inspected, apply random resampling transformation to the vibration signal samples in the first vibration sample dataset to simulate rotational speed changes, and pre-train the prior feature extractor; a prior feature extractor optimization module, used to construct a fault classifier and a rotational speed discriminator, and jointly train and optimize the prior feature extractor using a second vibration sample dataset through an adversarial domain adaptation method to obtain a target fault classifier and a target feature extractor; wherein, the first training objective of the adversarial domain adaptation is to optimize the prior feature extractor and the fault classifier to maximize the fault classification accuracy, and the second training objective is to optimize the prior feature extractor to minimize the rotational speed classification accuracy of the rotational speed discriminator; and an equipment operation detection module, used to collect real-time vibration signals of the industrial equipment to be inspected, input them into the target feature extractor for feature extraction, and then input them into the target fault classifier to obtain equipment operation detection information.

[0015] The proposed intelligent detection method and system for industrial equipment operation performance first collects a first vibration sample dataset of the industrial equipment to be inspected. Random resampling transformation is applied to the vibration signal samples in the first vibration sample dataset to simulate rotational speed changes. A prior feature extractor is pre-trained. Then, a fault classifier and a rotational speed discriminator are constructed. A second vibration sample dataset is used to jointly train and optimize the prior feature extractor through adversarial domain adaptation, resulting in a target fault classifier and a target feature extractor. The first training objective of adversarial domain adaptation is to optimize the prior feature extractor and fault classifier to maximize fault classification accuracy. The second training objective is to optimize the prior feature extractor to minimize the rotational speed classification accuracy of the rotational speed discriminator. Finally, real-time vibration signals of the industrial equipment to be inspected are collected and input into the target feature extractor for feature extraction, and then input into the target fault classifier to obtain equipment operation detection information. This achieves the technical effect of improving feature extraction accuracy and model generalization ability, thereby improving the reliability of equipment operation detection results. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0017] Figure 1This is a flowchart illustrating an intelligent detection method for the operating performance of industrial equipment provided in an embodiment of this application.

[0018] Figure 2 This is a schematic diagram of the structure of an intelligent detection system for the operating performance of industrial equipment provided in an embodiment of this application.

[0019] Figure labeling: Prior feature extractor pre-training module 10, prior feature extractor optimization module 20, equipment operation detection module 30. Detailed Implementation

[0020] The above description is merely an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below.

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application will be provided in conjunction with the accompanying drawings. The described embodiments should not be considered as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] In the following description, references to "some embodiments" describe a subset of all possible embodiments. However, it is understood that "some embodiments" can be the same or different subsets of all possible embodiments and can be combined with each other without conflict. The terms "first" and "second" are used merely to distinguish similar objects and do not represent a specific ordering of objects. The terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules not explicitly listed or inherent to these processes, methods, products, or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only.

[0023] This application provides an intelligent detection method for the operating performance of industrial equipment, such as... Figure 1 As shown, the method includes:

[0024] Step S100: Collect the first vibration sample dataset of the industrial equipment to be inspected, apply random resampling transformation to the vibration signal samples in the first vibration sample dataset to simulate rotational speed changes, and pre-train the prior feature extractor.

[0025] Specifically, vibration sensors, such as piezoelectric accelerometers, are installed in critical parts of the industrial equipment under inspection, such as motor bearings and gearboxes, which are prone to vibration and closely related to the equipment's operating performance. The sensors can detect vibrations generated during equipment operation in real time and convert them into electrical signals. A data acquisition card converts the analog electrical signals output by the sensors into digital signals, samples them at a certain sampling frequency, and stores the acquired digital signals in chronological order to form the first vibration sample dataset.

[0026] The first vibration sample dataset is randomly resampled. First, the resampling parameters are determined, including the resampling ratio range. Then, a resampling ratio is randomly generated within this range. Next, each vibration signal sample in the first vibration sample dataset is resampled according to the generated resampling ratio. For example, if the resampling ratio is 0.9, the number of sampling points in the original vibration signal sample is reduced to 0.9 times the original value, simulating a decrease in equipment rotation speed; if the resampling ratio is 1.1, the number of sampling points is increased, simulating an increase in equipment rotation speed.

[0027] A neural network model for a prior feature extractor is constructed using a deep learning framework. Vibration signal samples, after random resampling, are used as input data and fed into the neural network model. The model is trained using backpropagation and an optimizer. By continuously adjusting the model's parameters, it is enabled to extract useful features from the input vibration signal samples. During training, a loss function is used to measure the difference between the model's output and the expected features, and the model parameters are updated based on the value of the loss function until the model converges, completing the pre-training of the prior feature extractor.

[0028] In one possible implementation, a first vibration sample dataset of the industrial equipment to be inspected is collected. A random resampling transformation is applied to the vibration signal samples in the first vibration sample dataset to simulate rotational speed changes. A prior feature extractor is pre-trained. Step S100 further includes step S110, where the first vibration sample dataset contains multiple sets of vibration signal samples collected at different rotational speeds and labeled with fault types. Specifically, the equipment is operated under different preset rotational speed conditions. For example, for a large centrifugal pump, its rotational speed is set to multiple different values ​​such as 1000 r / min, 1500 r / min, and 2000 r / min. Under each rotational speed condition, the equipment is allowed to run continuously for a period of time, and vibration signals are collected using vibration sensors installed on key parts of the equipment. Simultaneously with the collection of vibration signals, the vibration signal samples collected under each rotational speed condition are labeled with the corresponding fault type based on the actual operating state of the equipment or through artificially simulated faults. For example, if impurities are artificially added to the pump impeller at a certain speed to simulate impeller blockage, all vibration signal samples collected at that speed are labeled as "impeller blockage" fault type; if the equipment operates normally without faults, it is labeled as "normal". Vibration signal samples collected at different speeds and labeled with fault types are then compiled together to form the first vibration sample dataset.

[0029] Step S120: Construct the model architecture of the prior feature extractor. Using the first vibration sample dataset, pre-train the prior feature extractor model architecture based on a self-supervised learning task, enabling the prior feature extractor to perceive changes in rotational speed from vibration signals. Specifically, a convolutional neural network is constructed as the model architecture of the prior feature extractor. This convolutional neural network model includes multiple convolutional layers, pooling layers, and fully connected layers. The convolutional layers are used to extract local features from the vibration signals, for example, by setting different sized convolutional kernels to capture vibration patterns at different scales. The pooling layers are used to downsample the feature maps output by the convolutional layers, reducing feature dimensions and improving the computational efficiency of the model. The fully connected layers integrate and transform the features output by the pooling layers, outputting the final feature vector.

[0030] The self-supervised learning task is to predict the rotational speed of vibration signals. Vibration signal samples from a first vibration sample dataset are used as input, and the corresponding rotational speed values ​​are used as the target output. During training, the vibration signal samples are input into a pre-constructed convolutional neural network model. The model outputs the predicted rotational speed values, and the error between the predicted and actual rotational speed values ​​is calculated as the loss function. An optimizer is used to train the convolutional neural network model based on the loss function. By continuously adjusting the model's parameters, the predicted rotational speed values ​​output by the model gradually approach the actual rotational speed values, thus enabling the prior feature extractor to perceive changes in rotational speed from the vibration signals. During training, a batch training approach can be used, dividing the first vibration sample dataset into multiple mini-batches and using a mini-batch of data for training each time to improve training efficiency and stability.

[0031] This implementation uses a pre-trained prior feature extractor, which already has a certain ability to perceive changes in rotational speed. In subsequent adversarial training, it can focus more on learning features that are useful for fault classification and robust to rotational speed, thereby improving the accuracy and robustness of feature extraction.

[0032] In one possible implementation, a model architecture based on a prior feature extractor is pre-trained using a self-supervised learning task. Step S120 further includes step S121, applying a random resampling transformation to the vibration signal samples input to the prior feature extractor to simulate linear changes in rotational speed. Specifically, in the analysis of vibration signals from industrial equipment, changes in rotational speed cause changes in the frequency components of the vibration signal. Resampling is a technique that changes the signal sampling rate. By randomly changing the sampling rate of the vibration signal, it simulates the changes in the frequency components of the vibration signal similar to linear changes in rotational speed. For example, if the sampling rate of the original vibration signal is f... s When performing a resampling transformation, a resampling scaling factor r is randomly generated, and the new sampling rate becomes f. s new =r×f s The original vibration signal is resampled according to the new sampling rate to obtain the transformed vibration signal.

[0033] Step S122 involves inputting the transformed signal into the prior feature extractor to train an auxiliary speed estimation network. This network predicts the resampling scaling factor based on the features output by the prior feature extractor. Specifically, the auxiliary speed estimation network can be a fully connected neural network that receives the feature vector output by the prior feature extractor as input and outputs a predicted resampling scaling factor through calculations by multiple fully connected layers. The transformed vibration signal obtained in step S121 is input into the prior feature extractor to obtain a feature vector. This feature vector is then input into the auxiliary speed estimation network to obtain the predicted resampling scaling factor. Simultaneously, given the actual applied resampling scaling factor, the predicted value is compared with the actual value.

[0034] Step S123: Pre-training of the prior feature extractor is completed by minimizing the prediction error of the scaling factor. Specifically, to measure the difference between the predicted resampling scaling factor and the actual scaling factor, a loss function, such as the mean squared error loss function, is defined. An optimization algorithm is used to update the parameters of the prior feature extractor and the auxiliary rotational speed estimation network based on the loss function. During training, the transformed vibration signal is iteratively input into the model, the loss function value is calculated, and then the parameters are updated according to the optimization algorithm, causing the loss function value to gradually decrease. A training stopping condition is set, for example, when the loss function value no longer decreases significantly after several consecutive iterations, or when a preset maximum number of iterations is reached, the pre-training is considered complete. At this point, the prior feature extractor possesses a certain ability to perceive rotational speed-related features from the vibration signal.

[0035] This implementation introduces randomness into the rotational speed variation through random resampling transformation. This makes the features learned by the prior feature extractor during pre-training not only sensitive to specific rotational speed variation patterns but also robust to different types and degrees of rotational speed variation. In real industrial environments, equipment rotational speed variations can be complex and irregular. This pre-trained prior feature extractor can better adapt to actual working conditions, extracting stable features unaffected by rotational speed variations, thereby improving the reliability and accuracy of the entire system.

[0036] Step S200: Construct a fault classifier and a rotational speed discriminator, and use the second vibration sample dataset to jointly train and optimize the prior feature extractor through adversarial domain adaptation to obtain the target fault classifier and the target feature extractor; wherein, the first training objective of adversarial domain adaptation is to optimize the prior feature extractor and the fault classifier to maximize the fault classification accuracy, and the second training objective is to optimize the prior feature extractor to minimize the rotational speed classification accuracy of the rotational speed discriminator.

[0037] Specifically, a deep learning framework is used to construct neural network models for both the fault classifier and the speed discriminator. The fault classifier can employ a multilayer perceptron or a more complex deep neural network structure. Its input is the features extracted by the prior feature extractor, and its output is the probability distribution of the possible fault types of the equipment. The speed discriminator also uses a neural network structure, such as a fully connected neural network. Its input is also the features extracted by the prior feature extractor, and its output is the probability of the equipment's speed belonging to a specific category. This helps optimize the prior feature extractor and eliminates the influence of speed on the features.

[0038] Adversarial domain adaptation (ADO) is a machine learning training method that simultaneously considers two opposing training objectives, enabling the model to learn more generalizable features across data in different speed domains, thereby improving model performance. The ADO training process involves two training objectives. The first objective is to maximize fault classification accuracy. A second vibration sample dataset is input into a prior feature extractor to extract features, which are then input into a fault classifier for classification. The loss function for fault classification is calculated, and backpropagation and an optimizer are used to update the parameters of both the prior feature extractor and the fault classifier to reduce the loss function value and improve fault classification accuracy. The second objective is to minimize the speed classification accuracy of the speed discriminator. Features extracted by the prior feature extractor are input into the speed discriminator for speed classification, and the loss function for speed classification is calculated. However, when updating parameters, only the parameters of the prior feature extractor are updated. A gradient inversion layer technique is used to invert the gradient of the speed classification loss function before backpropagation, causing the prior feature extractor to update parameters in a direction that reduces speed classification accuracy, thus eliminating the influence of speed on the features. Through continuous iterative training, until both training objectives reach a good balance, the target fault classifier and target feature extractor are obtained.

[0039] The second vibration sample dataset is similar to the first vibration sample dataset. It is also a collection of device vibration signal data collected by sensors and is used for adversarial domain adaptation training to optimize the prior feature extractor.

[0040] In one possible implementation, a fault classifier and a rotational speed discriminator are constructed, and the prior feature extractor is jointly trained and optimized using a second vibration sample dataset through an adversarial domain adaptation method to obtain a target fault classifier and a target feature extractor. Step S200 further includes step S210, configuring the fault classifier and the rotational speed discriminator respectively using a multilayer perceptron. Specifically, a multilayer perceptron is a feedforward artificial neural network model consisting of an input layer, multiple hidden layers, and an output layer. For the fault classifier, the input layer receives feature vectors extracted from the vibration signal by the prior feature extractor; the hidden layer contains multiple neurons and uses an activation function to introduce nonlinearity to enhance the model's ability to extract complex features; the number of neurons in the output layer is the same as the number of fault categories, with each neuron corresponding to a fault type. The output is converted into a probability distribution using a softmax function, representing the probability that the input signal belongs to each type of fault.

[0041] The multilayer perceptron structure of the speed discriminator is similar to that of a fault classifier. The input layer also receives feature vectors from the prior feature extractor; the hidden layers also employ a multilayer structure and use activation functions such as ReLU; the output layer has only one neuron, outputting a value between 0 and 1, representing the probability that the input feature comes from different speed domains. For example, during training, if there are two speed domains, a high speed domain and a low speed domain, an output value close to 1 indicates that the feature comes from the high speed domain; a value close to 0 indicates that the feature comes from the low speed domain.

[0042] Step S220: Based on the adversarial domain adaptation method, using the second vibration sample dataset as training data, the gradient descent algorithm is used to alternately optimize the first training objective and the second training objective: Step 1: Fix the parameters of the prior feature extractor and the fault classifier, and update the parameters of the rotational speed discriminator; Step 2: Fix the parameters of the rotational speed discriminator, and jointly update the parameters of the prior feature extractor and the fault classifier to complete the joint adversarial optimization of the first training objective and the second training objective. Specifically, in Step 1, the objective is to maximize the accuracy of rotational speed discrimination. The signals in the second vibration sample dataset are input into the prior feature extractor to obtain feature vectors, and then these feature vectors are input into the rotational speed discriminator. The rotational speed discriminator determines which rotational speed domain the input feature vectors come from based on the input feature vectors and outputs the corresponding probability values. The gradient descent algorithm is used to update the parameters of the rotational speed discriminator. By calculating the gradient of the loss function with respect to the parameters of the rotational speed discriminator, the parameters are adjusted along the direction of gradient descent, so that the rotational speed discriminator can more accurately determine which rotational speed domain the input features come from. For example, if the current speed discriminator makes an incorrect judgment about a certain feature vector, the loss function will calculate a large gradient, prompting the parameters of the speed discriminator to be adjusted in the direction that can make the correct judgment.

[0043] In step two, the goal is to simultaneously minimize the fault classification loss and maximize the speed discrimination loss. The fault classification loss measures the accuracy of the fault classifier in classifying fault categories, while the speed discrimination loss measures the accuracy of the speed discriminator in determining which speed domain a feature originates from. Minimizing the fault classification loss allows the fault classifier to better classify faults; maximizing the speed discrimination loss forces the feature extractor to adjust its parameters to generate features that can "deceive" the speed discriminator, making it impossible for the speed discriminator to accurately determine the speed based on the features. The gradient descent algorithm is also used to jointly update the parameters of the prior feature extractor and the fault classifier. The gradient of the joint loss function with respect to the parameters of these two models is calculated, and then the parameters are adjusted along the direction of gradient descent. For example, if the features generated by the prior feature extractor are still easily identified by the speed discriminator in terms of speed domain, the joint loss function will calculate a larger gradient, prompting the parameters of the prior feature extractor to be adjusted towards generating features that are more difficult for the speed discriminator to identify. Simultaneously, the parameters of the fault classifier will also be adjusted accordingly to maintain good fault classification performance.

[0044] Step S230: Iteratively execute steps one and two, using the joint adversarial loss function for convergence judgment, to obtain the target fault classifier and target feature extractor whose joint adversarial loss meets the preset loss threshold. Specifically, parameters are updated alternately in the order of steps one and two, continuously iterating this process. Each iteration makes the model make certain progress in fault classification and adversarial speed discrimination. After each iteration, the value of the joint adversarial loss function is calculated. The joint adversarial loss function is a weighted sum of the fault classification loss and the speed discrimination loss, used to measure the overall performance of the entire model on the two training objectives. When the value of the joint adversarial loss function no longer decreases significantly after several consecutive iterations, or reaches the preset maximum number of iterations, the model is considered to have converged. At this time, the obtained fault classifier and prior feature extractor are the target fault classifier and target feature extractor.

[0045] This implementation employs an adversarial domain adaptation approach, where the prior feature extractor continuously adjusts its parameters during training to generate features insensitive to changes in rotational speed. This enables the target fault classifier to more accurately identify fault types when faced with vibration signals at different rotational speeds. For example, in industrial equipment, the equipment's rotational speed fluctuates due to factors such as load variations. The fault classifier trained in this adversarial manner can maintain high classification accuracy under different rotational speed conditions, improving the model's reliability in practical applications.

[0046] In one possible implementation, step S230 further includes: the expression for the joint adversarial loss function is: ;in, For the joint adversarial loss function; The cross-entropy loss of the fault classifier, This refers to the cross-entropy loss of the speed discriminator; This is a positive hyperparameter used to balance the weights of the fault classification task and the rotation speed determination task in gradient updates.

[0047] Specifically, L c This refers to the cross-entropy loss of the fault classifier. Cross-entropy is a loss function used to measure the difference between two probability distributions. In fault classification tasks, the fault classifier outputs a probability distribution for the input feature vector, representing the probability that the feature belongs to different fault categories. The true fault category label can also be viewed as a probability distribution. c Cross-entropy is used to calculate the difference between the probability distribution output by the fault classifier and the probability distribution of the true labels. Its value reflects the accuracy of the fault classifier in judging the fault category under the current parameters. The smaller the value, the more accurate the classification.

[0048] L d This refers to the cross-entropy loss of the speed discriminator. Similarly, in the speed discrimination task, the speed discriminator outputs a probability distribution for the input feature vector, representing the probability that the feature comes from different speed domains. Likewise, the true speed domain label is also a probability distribution. d It is used to measure the difference between the probability distribution output by the speed discriminator and the probability distribution of the true speed domain labels. Its value reflects the accuracy of the speed discriminator in judging which speed domain the feature comes from.

[0049] λ is a positive hyperparameter used to balance the weights of fault classification and rotational speed discrimination tasks in gradient updates. Since the two tasks have different objectives—fault classification aims to accurately identify fault types, while rotational speed discrimination aims to accurately determine which rotational speed domain a feature originates from—the value of λ needs to be adjusted during joint training to coordinate the learning processes of the two tasks. If λ is too large, the model will focus more on adversarial rotational speed discrimination, potentially affecting the performance of fault classification; if λ is too small, the rotational speed discrimination task may not effectively encourage the feature extractor to generate rotational speed-insensitive features, thus affecting the model's generalization ability across different rotational speeds.

[0050] In one possible implementation, step S230 further includes: determining the λ value through a grid search, wherein the grid search step includes: randomly generating multiple candidate λ values ​​in ascending order within the range of 0 to 1; for each candidate λ value, alternately training and testing using a gradient descent algorithm, and recording the decrease in fault classification accuracy and the decrease in rotational speed discrimination accuracy corresponding to each candidate λ value; selecting candidate λ values ​​whose decrease in fault classification accuracy is less than a first preset decrease value and whose decrease in rotational speed discrimination accuracy exceeds a second preset decrease value as the λ value of the joint adversarial loss function; wherein the second preset decrease value is greater than the first preset decrease value.

[0051] Specifically, multiple candidate λ values ​​are randomly generated in ascending order within the range of 0 to 1. For example, a random number generation function in a programming language can be used to uniformly and randomly select a certain number of values ​​between 0 and 1 as candidate λ values. For example, the generated candidate λ values ​​might be 0.05, 0.12, 0.23, 0.37, etc.

[0052] For each generated candidate λ value, gradient descent algorithm is used for alternating training and testing. The specific process is similar to the alternating optimization process in step S220. First, the parameters of the prior feature extractor and fault classifier are fixed, and the parameters of the speed discriminator are updated to maximize the speed discrimination accuracy. Then, the parameters of the speed discriminator are fixed, and the parameters of the prior feature extractor and fault classifier are jointly updated to complete the joint adversarial optimization of the first and second training objectives. This process will iterate multiple times until the model reaches a certain convergence condition. During the alternating training and testing process, the decrease in fault classification accuracy and speed discrimination accuracy corresponding to each candidate λ value are recorded. The decrease in fault classification accuracy refers to the degree of decrease in fault classification accuracy under the current candidate λ value relative to the fault classification accuracy without introducing adversarial training (i.e., when λ=0). The decrease in speed discrimination accuracy refers to the degree of decrease in speed discrimination accuracy under the current candidate λ value relative to the speed discrimination accuracy when only the speed discriminator is trained, i.e., without considering the influence of the speed discrimination part in the adversarial loss on the feature extractor.

[0053] From all candidate λ values, the candidate λ value whose decrease in fault classification accuracy is less than a first preset decrease value and whose decrease in rotational speed discrimination accuracy exceeds a second preset decrease value is selected as the λ value for the joint adversarial loss function. The second preset decrease value is greater than the first preset decrease value. For example, assuming the first preset decrease value is 5% and the second preset decrease value is 20%, if a candidate λ value corresponds to a 3% decrease in fault classification accuracy and a 25% decrease in rotational speed discrimination accuracy, then that candidate λ value meets the selection criteria. This is because in adversarial training, it is desirable to minimize the decrease in fault classification accuracy, i.e., to maintain the basic performance of the fault classification task, while significantly reducing rotational speed discrimination accuracy, i.e., to encourage the feature extractor to generate features insensitive to rotational speed, thus meeting the goals of joint adversarial training.

[0054] This implementation determines a suitable λ value through grid search, accurately balancing the weights of fault classification and rotational speed determination tasks in joint adversarial training. This allows the model to effectively reduce rotational speed determination accuracy without significantly decreasing fault classification accuracy, thus enabling the features generated by the prior feature extractor to focus more on fault information while exhibiting better robustness to rotational speed information.

[0055] In one possible implementation, step S230 further includes: determining the positive hyperparameter through a dynamic scheduler, specifically including: using the λ value determined by grid search as the initial λ value; alternately optimizing the first training objective and the second training objective using a gradient descent algorithm, and collecting the decrease amplitude of real-time fault classification accuracy and the decrease amplitude of real-time speed discrimination accuracy during the training process; inputting the decrease amplitude of real-time fault classification accuracy and the decrease amplitude of real-time speed discrimination accuracy into the dynamic scheduler for dynamic adjustment of the λ value; wherein, the dynamic scheduler includes an increase threshold and a decrease threshold, the decrease threshold being the adjustment step size for decreasing when the decrease amplitude of real-time fault classification accuracy is greater than the first preset decrease amplitude, and the increase threshold being the adjustment step size for increasing when the decrease amplitude of real-time speed discrimination accuracy is less than the second preset decrease amplitude.

[0056] Specifically, firstly, an initial λ value is determined according to the grid search method described above. During the grid search process, multiple candidate λ values ​​are generated within the range of 0 to 1. Based on the recording and selection criteria of the decrease in fault classification accuracy and the decrease in speed discrimination accuracy, the candidate λ value that meets the requirements is selected as the initial λ value. For example, after the grid search, the initial λ value is determined to be 0.3, and this value is used as the starting point for subsequent dynamic control.

[0057] During the training process, which alternately optimizes the first and second training objectives using the gradient descent algorithm, the decrease in fault classification accuracy and the decrease in rotational speed determination accuracy are collected in real time. Specifically, in each iteration of the alternating training, the decrease in fault classification accuracy relative to the initial state and the decrease in rotational speed determination accuracy are calculated.

[0058] The real-time data on the decrease in fault classification accuracy and the decrease in speed discrimination accuracy are input into the dynamic scheduler. The dynamic scheduler has two parameters: an increase threshold and a decrease threshold. When the decrease in real-time fault classification accuracy exceeds a first preset decrease threshold, the dynamic scheduler decreases the value of λ by the adjustment step size specified by the decrease threshold. For example, if the first preset decrease threshold is 5%, and the current real-time fault classification accuracy decrease is 7%, and the adjustment step size specified by the decrease threshold is 0.05, then the value of λ will decrease by 0.05, becoming 0.25. This is because the decrease in fault classification accuracy exceeds the preset tolerance range; decreasing the value of λ reduces the weight of the speed discrimination task in the joint adversarial loss function, thereby mitigating the impact of adversarial training on the performance of the fault classification task. When the decrease in real-time speed discrimination accuracy is less than a second preset decrease threshold, the dynamic scheduler increases the value of λ by the adjustment step size specified by the increase threshold. For example, if the second preset decrease value is 20%, and the current real-time speed discrimination accuracy decreases by 18%, and the adjustment step size specified by increasing the threshold is 0.03, then the λ value will increase by 0.03, becoming 0.33. This is because the decrease in speed discrimination accuracy did not reach the expected level. Increasing the λ value can enhance the weight of the speed discrimination task in the joint adversarial loss function, prompting the feature extractor to better eliminate the influence of speed information on features.

[0059] This implementation uses a dynamic scheduler that adjusts the λ value based on the accuracy drop during real-time training, enabling a more precise balance between fault classification and speed determination tasks. Compared to a fixed λ value, dynamic adjustment can respond promptly to changes in model performance during training, avoiding situations where a fixed λ value leads to excessive performance degradation in one task or failure to effectively achieve the objective of another task.

[0060] Step S300: Real-time vibration signal of the industrial equipment under inspection is collected and input into the target feature extractor for feature extraction, and then input into the target fault classifier to obtain equipment operation detection information.

[0061] Specifically, similar to the method of collecting the first vibration sample dataset in step S100, vibration sensors are installed on key parts of the industrial equipment to be inspected. Vibration signals during equipment operation are collected in real time via a data acquisition card and converted into digital signals for storage. The real-time collected vibration signals are input into a target feature extractor, which processes the input vibration signals and extracts feature vectors that reflect the equipment's operating status. The feature vectors extracted by the target feature extractor are input into a target fault classifier, which classifies the input features based on learned fault feature patterns and outputs the possible fault types of the equipment and their corresponding probabilities. Based on the classification results, it is determined whether the equipment's operating status is normal and what possible faults may exist.

[0062] In one possible implementation, after obtaining the target fault classifier and the target feature extractor, step S300 further includes step S310, whereby the target feature extractor outputs pure fault features independent of rotational speed. Specifically, the real-time acquired equipment vibration signal is input into the pre-trained target feature extractor. The target feature extractor processes the input vibration signal based on the parameters and feature extraction rules learned during its previous adversarial domain adaptation training. Through a multi-layer neural network structure, features are extracted and transformed from the vibration signal, filtering out information related to rotational speed, and finally outputting a pure fault feature vector independent of rotational speed.

[0063] Step S320 involves reconstructing the vibration signal of the pure fault features based on fault-related periodic impact components to generate an enhanced fault impact signal. Specifically, the fault-related information contained in the pure fault features is analyzed to determine the characteristic parameters of the fault-related periodic impact components, such as impact frequency and amplitude. Based on these characteristic parameters, a signal reconstruction method, such as a method based on sine wave superposition or pulse sequence generation, is used to generate a fault-related periodic impact signal, i.e., an enhanced fault impact signal. This signal highlights the fault-related periodic impact characteristics and is more significant and easily identifiable.

[0064] Step S330: The enhanced fault impulse signal is added to the equipment operation detection information. Specifically, the equipment operation detection information includes fault diagnosis-related data such as fault type judgment and fault probability. The enhanced fault impulse signal strengthens the key fault-related information. The fault characteristic information represented by the enhanced fault impulse signal is added to the equipment operation detection information through data fusion. This implementation method, by adding the information of the enhanced fault impulse signal to the equipment operation detection information, makes the equipment operation detection information contain richer fault characteristic details.

[0065] In one possible implementation, the pure fault features are reconstructed into vibration signals based on fault-related periodic impact components to generate a reconstructed enhanced fault impact signal. Step S320 further includes step S321, constructing a decoder-generator network. Specifically, the decoder-generator network is a deep learning model structure used to transform feature representations in the latent space back to the original data space. In this application, it is used to map pure fault features from the feature space back to a vibration signal space with periodic impact components. This network consists of multiple fully connected layers, deconvolutional layers, etc.

[0066] Step S322 involves collecting clean fault feature samples and reconstructed signal samples. Specifically, clean fault feature samples are collected from the processed data. These samples are rotational speed-independent feature vectors obtained by processing a large number of collected equipment vibration signals using a target feature extractor. For example, clean fault features can be extracted from equipment vibration signals collected at different times and under different operating conditions to form a diverse sample set covering various possible fault scenarios. Reconstructed signal samples can be obtained in several ways. One method is to utilize existing fault simulation experimental data, artificially setting different fault types and degrees in the experiment, and collecting corresponding vibration signals as a reference for the reconstructed signal samples. Another method is to generate theoretically periodic impact signals based on the physical model of the fault, as part of the reconstructed signal samples.

[0067] Step S323: Using clean fault feature samples and reconstructed signal samples, the decoder-generator network is trained to obtain a signal reconstruction network where the signal reconstruction loss, periodicity consistency loss, and fault classification loss all meet preset loss thresholds. Specifically, the clean fault feature samples are used as input to the decoder-generator network, and the network outputs the reconstructed signal. Then, the signal reconstruction loss, periodicity consistency loss, and fault classification loss are calculated. The signal reconstruction loss measures the difference between the reconstructed signal output by the network and the corresponding reconstructed signal sample. It can be calculated using loss functions such as mean square error, for example, by calculating the average of the sum of squares of the amplitude differences between the reconstructed signal and the reconstructed signal sample at each time point. The periodicity consistency loss ensures that the periodicity characteristics of the reconstructed signal are consistent with expectations. It can be defined by analyzing the frequency components of the reconstructed signal and calculating the difference between them and the theoretical fault frequency. For example, Fourier transform can be used to convert the reconstructed signal to the frequency domain, the deviation between the actual frequency components and the theoretical fault frequency can be calculated, and this deviation can be used as part of the loss function. The fault classification loss ensures that the reconstructed signal reflects fault information. The reconstructed signal can be input into an auxiliary fault classifier to calculate the difference between the classification result and the actual fault type, which is used as the fault classification loss.

[0068] The optimizer updates the parameters of the decoder-generator network based on the calculated total loss function, which is the weighted sum of the three losses. Through continuous iterative training, the network parameters are adjusted so that the signal reconstruction loss, periodic consistency loss, and fault classification loss gradually decrease. When all three losses meet the preset loss thresholds, the network training is considered complete, and the signal reconstruction network is obtained.

[0069] Step S324: The pure fault features are reconstructed using the signal reconstruction network to generate the enhanced fault impact signal. Specifically, the pure fault features obtained in step S310 are input into the trained signal reconstruction network. The network converts the pure fault features into a vibration signal with a periodic impact component, i.e., the enhanced fault impact signal, based on the learned mapping relationship.

[0070] This application's embodiments employ a first vibration sample dataset of the industrial equipment under inspection. The vibration signal samples in this dataset undergo random resampling to simulate rotational speed changes, thereby pre-training a prior feature extractor and constructing a fault classifier and a rotational speed discriminator. Using a second vibration sample dataset, a joint training method based on adversarial domain adaptation is employed to optimize the prior feature extractor, achieving two training objectives: obtaining a target fault classifier and a target feature extractor. Real-time vibration signals from the industrial equipment under inspection are collected, first input into the target feature extractor to extract features, and then input into the target fault classifier to obtain equipment operation detection information. These technical means address the technical problems of inaccurate feature extraction and poor model generalization ability in existing industrial equipment operation performance detection methods when dealing with rotational speed changes and differences in data distribution under different operating conditions, leading to unreliable equipment operation detection results. This achieves the technical effect of improving feature extraction accuracy and model generalization ability, thereby improving the reliability of equipment operation detection results.

[0071] In the above text, refer to Figure 1 A detailed description of an intelligent detection method for the operating performance of industrial equipment according to embodiments of the present invention is provided. Next, reference will be made to... Figure 2 This invention describes an intelligent detection system for industrial equipment operating performance based on embodiments of the present invention.

[0072] The intelligent detection system for industrial equipment operating performance according to embodiments of the present invention addresses the technical problems of inaccurate feature extraction and poor model generalization ability in existing detection methods for industrial equipment operating performance when dealing with variations in rotational speed and differences in data distribution under different operating conditions, leading to unreliable equipment operation detection results. The system aims to improve the accuracy of feature extraction and the generalization ability of the model, thereby enhancing the reliability of equipment operation detection results. The intelligent detection system for industrial equipment operating performance includes: a prior feature extractor pre-training module 10, a prior feature extractor optimization module 20, and an equipment operation detection module 30.

[0073] The prior feature extractor pre-training module 10 is used to collect a first vibration sample dataset of the industrial equipment under inspection, apply random resampling transformation to the vibration signal samples in the first vibration sample dataset to simulate speed changes, and pre-train the prior feature extractor. The prior feature extractor optimization module 20 is used to construct a fault classifier and a speed discriminator, and jointly train and optimize the prior feature extractor using a second vibration sample dataset through adversarial domain adaptation to obtain a target fault classifier and a target feature extractor. The first training objective of adversarial domain adaptation is to optimize the prior feature extractor and the fault classifier to maximize the fault classification accuracy, and the second training objective is to optimize the prior feature extractor to minimize the speed classification accuracy of the speed discriminator. The equipment operation detection module 30 is used to collect the real-time vibration signal of the industrial equipment under inspection, input it into the target feature extractor for feature extraction, and then input it into the target fault classifier to obtain equipment operation detection information.

[0074] The detailed description of the specific configuration of the prior feature extractor pre-training module 10 is explained as follows: As mentioned above, a first vibration sample dataset of the industrial equipment to be inspected is collected, and a random resampling transformation is applied to the vibration signal samples in the first vibration sample dataset to simulate speed changes. The prior feature extractor is pre-trained. The prior feature extractor pre-training module 10 may further include: a first vibration sample dataset acquisition unit for collecting the first vibration sample dataset of the industrial equipment to be inspected, the first vibration sample dataset containing multiple sets of vibration signal samples with labeled fault types collected at different speeds; and a model architecture construction unit for constructing the model architecture of the prior feature extractor. Using the first vibration sample dataset, the prior feature extractor is pre-trained based on a self-supervised learning task and the model architecture of the prior feature extractor, so that the prior feature extractor has the ability to perceive speed changes from vibration signals.

[0075] Specifically, the model architecture based on the prior feature extractor is pre-trained based on the self-supervised learning task. The model architecture building unit may further include: a resampling transformation subunit for applying a random resampling transformation to the vibration signal sample input to the prior feature extractor to simulate the linear change of rotational speed; an auxiliary rotational speed estimation network training subunit for inputting the transformed signal to the prior feature extractor to train the auxiliary rotational speed estimation network to predict the resampling scaling factor based on the features output by the prior feature extractor; and a pre-training subunit for completing the pre-training of the prior feature extractor by minimizing the prediction error of the scaling factor.

[0076] The detailed description of the specific configuration of the prior feature extractor optimization module 20 is explained as follows: As mentioned above, a fault classifier and a rotational speed discriminator are constructed, and the prior feature extractor is jointly trained and optimized using the second vibration sample dataset through an adversarial domain adaptation method to obtain a target fault classifier and a target feature extractor. The prior feature extractor optimization module 20 may further include: a multilayer perceptron configuration unit for configuring the fault classifier and the rotational speed discriminator respectively using a multilayer perceptron; an alternating optimization unit for alternately optimizing the first training objective and the second training objective based on the adversarial domain adaptation method, using the second vibration sample dataset as training data, and employing a gradient descent algorithm: Step 1: Fix the parameters of the prior feature extractor and the fault classifier, and update the parameters of the rotational speed discriminator; Step 2: Fix the parameters of the rotational speed discriminator, and jointly update the parameters of the prior feature extractor and the fault classifier to complete the joint adversarial optimization of the first training objective and the second training objective; a convergence judgment unit for iteratively executing steps 1 and 2, using the joint adversarial loss function to perform convergence judgment, and obtaining a target fault classifier and a target feature extractor whose joint adversarial loss meets a preset loss threshold.

[0077] The convergence determination unit may further include: the expression for the joint adversarial loss function is: ;in, For the joint adversarial loss function; The cross-entropy loss of the fault classifier, This refers to the cross-entropy loss of the speed discriminator; This is a positive hyperparameter used to balance the weights of the fault classification task and the rotation speed determination task in gradient updates.

[0078] The convergence judgment unit may further include: determining the λ value through grid search, wherein the grid search steps include: randomly generating multiple candidate λ values ​​in ascending order within the range of 0 to 1; for each candidate λ value, alternately training and testing using the gradient descent algorithm, and recording the decrease in fault classification accuracy and the decrease in rotational speed discrimination accuracy corresponding to each candidate λ value; selecting candidate λ values ​​whose decrease in fault classification accuracy is less than a first preset decrease value and whose decrease in rotational speed discrimination accuracy exceeds a second preset decrease value as the λ value of the joint adversarial loss function; wherein the second preset decrease value is greater than the first preset decrease value.

[0079] The convergence judgment unit may further include: determining positive hyperparameters through a dynamic scheduler, specifically including: using the λ value determined by grid search as the initial λ value; alternately optimizing the first training objective and the second training objective using a gradient descent algorithm, and collecting the decrease magnitude of real-time fault classification accuracy and the decrease magnitude of real-time speed discrimination accuracy during the training process; inputting the decrease magnitude of real-time fault classification accuracy and the decrease magnitude of real-time speed discrimination accuracy into the dynamic scheduler for dynamic adjustment of the λ value; wherein the dynamic scheduler includes an increase threshold and a decrease threshold, the decrease threshold being the adjustment step size for decreasing when the decrease magnitude of real-time fault classification accuracy is greater than the first preset decrease magnitude, and the increase threshold being the adjustment step size for increasing when the decrease magnitude of real-time speed discrimination accuracy is less than the second preset decrease magnitude.

[0080] The detailed description of the specific configuration of the equipment operation detection module 30 is explained as follows: As mentioned above, after obtaining the target fault classifier and the target feature extractor, the equipment operation detection module 30 may further include: a pure fault feature output unit for using the target feature extractor to output pure fault features independent of rotational speed; a fault impact signal enhancement unit for reconstructing the pure fault features based on the periodic impact components related to the fault, generating a reconstructed enhanced fault impact signal; and an equipment operation detection information generation unit for adding the enhanced fault impact signal into the equipment operation detection information.

[0081] Specifically, the pure fault features are reconstructed using vibration signals based on fault-related periodic impact components to generate a reconstructed enhanced fault impact signal. The fault impact signal enhancement unit may further include: a decoder-generator network construction subunit for constructing a decoder-generator network; a sample acquisition subunit for acquiring pure fault feature samples and reconstructed signal samples; a signal reconstruction network acquisition subunit for training the decoder-generator network using the pure fault feature samples and reconstructed signal samples to obtain a signal reconstruction network whose signal reconstruction loss, periodic consistency loss, and fault classification loss all meet preset loss thresholds; and a signal reconstruction subunit for reconstructing the pure fault features using the signal reconstruction network to generate the enhanced fault impact signal.

[0082] The intelligent detection system for industrial equipment operation performance provided in this embodiment of the invention can execute the intelligent detection method for industrial equipment operation performance provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method execution.

[0083] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0084] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application. In some cases, the actions or steps described in this application can be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. An intelligent detection method for the operating performance of industrial equipment, characterized in that, include: Collect a first vibration sample dataset of the industrial equipment to be inspected, apply a random resampling transformation to the vibration signal samples in the first vibration sample dataset to simulate speed changes, and pre-train a prior feature extractor. A fault classifier and a rotation speed discriminator are constructed, and the prior feature extractor is jointly trained and optimized using a second vibration sample dataset through an adversarial domain adaptation method to obtain the target fault classifier and the target feature extractor. The first training objective of the adversarial domain adaptation is to optimize the prior feature extractor and the fault classifier to maximize the fault classification accuracy, and the second training objective is to optimize the prior feature extractor to minimize the rotational speed classification accuracy of the rotational speed discriminator. The real-time vibration signal of the industrial equipment under inspection is collected and input into the target feature extractor for feature extraction, and then input into the target fault classifier to obtain the equipment operation detection information.

2. The intelligent detection method for industrial equipment operating performance as described in claim 1, characterized in that, A first vibration sample dataset of the industrial equipment to be inspected is collected. A random resampling transformation is applied to the vibration signal samples in the first vibration sample dataset to simulate rotational speed changes. A pre-trained prior feature extractor is used, including: The first vibration sample dataset contains multiple sets of vibration signal samples with labeled fault types, collected at different rotational speeds; A model architecture for a prior feature extractor is constructed. Using the first vibration sample dataset, the model architecture for the prior feature extractor is pre-trained based on a self-supervised learning task, enabling the prior feature extractor to sense changes in rotational speed from vibration signals.

3. The intelligent detection method for industrial equipment operating performance as described in claim 2, characterized in that, Pre-training is performed on a model architecture based on a prior feature extractor for a self-supervised learning task, including: A random resampling transformation is applied to the vibration signal samples input to the prior feature extractor to simulate the linear change in rotational speed; The transformed signal is input to the prior feature extractor to train the auxiliary rotation speed estimation network, so as to predict the resampling scale factor based on the features output by the prior feature extractor. The pre-training of the prior feature extractor is completed by minimizing the prediction error of the scaling factor.

4. The intelligent detection method for industrial equipment operating performance as described in claim 1, characterized in that, A fault classifier and a rotational speed discriminator are constructed, and the prior feature extractor is jointly trained and optimized using a second vibration sample dataset through an adversarial domain adaptation approach to obtain the target fault classifier and the target feature extractor, including: The fault classifier and the speed discriminator are configured using a multilayer perceptron, respectively. Based on the adversarial domain adaptation approach, using the second vibration sample dataset as training data, the gradient descent algorithm is employed to alternately optimize the first training objective and the second training objective: Step 1: Fix the parameters of the prior feature extractor and the fault classifier, and update the parameters of the rotation speed discriminator; Step 2: Fix the parameters of the speed discriminator, jointly update the parameters of the prior feature extractor and the fault classifier, and complete the joint adversarial optimization of the first training objective and the second training objective; Iteratively execute steps one and two, and use the joint adversarial loss function to determine convergence, thereby obtaining a target fault classifier and a target feature extractor whose joint adversarial loss meets the preset loss threshold.

5. The intelligent detection method for industrial equipment operating performance as described in claim 4, characterized in that, The expression for the joint adversarial loss function is: ； in, For joint adversarial loss function; The cross-entropy loss of the fault classifier, This refers to the cross-entropy loss of the speed discriminator; This is a positive hyperparameter used to balance the weights of the fault classification task and the rotation speed determination task in gradient updates.

6. The intelligent detection method for industrial equipment operating performance as described in claim 5, characterized in that, The value of λ is determined through a grid search, the steps of which include: Multiple candidate λ values ​​are randomly generated in ascending order within the range of 0 to 1; For each candidate λ value, the gradient descent algorithm is used for alternating training and testing, and the decrease in fault classification accuracy and the decrease in speed discrimination accuracy corresponding to each candidate λ value are recorded. Candidate λ values ​​with a decrease in fault classification accuracy less than a first preset decrease value and a decrease in speed discrimination accuracy exceeding a second preset decrease value are selected as λ values ​​for the joint adversarial loss function. Wherein, the second preset decrease value is greater than the first preset decrease value.

7. The intelligent detection method for industrial equipment operating performance as described in claim 6, characterized in that, Positive hyperparameters are determined by the dynamic scheduler and include: The λ value determined through grid search is used as the initial λ value; The gradient descent algorithm is used to alternately optimize the first training objective and the second training objective, and the decrease in real-time fault classification accuracy and the decrease in real-time rotation speed discrimination accuracy are collected during the training process. The decrease in real-time fault classification accuracy and the decrease in real-time speed discrimination accuracy are input into the dynamic scheduler for dynamic adjustment of the λ value. The dynamic scheduler includes an increasing threshold and a decreasing threshold. The decreasing threshold is the adjustment step size when the decrease in real-time fault classification accuracy is greater than the first preset decrease value, and the increasing threshold is the adjustment step size when the decrease in real-time speed discrimination accuracy is less than the second preset decrease value.

8. The intelligent detection method for industrial equipment operating performance as described in claim 1, characterized in that, After obtaining the target fault classifier and the target feature extractor, the following is also included: Using the target feature extractor, pure fault features independent of rotational speed are output; The pure fault characteristics are reconstructed using vibration signals based on fault-related periodic impact components to generate reconstructed enhanced fault impact signals. The enhanced fault impact signal is added to the equipment operation detection information.

9. The intelligent detection method for industrial equipment operating performance as described in claim 8, characterized in that, The pure fault characteristics are reconstructed using vibration signals based on fault-related periodic impact components to generate a reconstructed enhanced fault impact signal, including: Construct a decoder-generator network; Collect clean fault feature samples and reconstructed signal samples; Using clean fault feature samples and reconstructed signal samples, the decoder-generator network is trained to obtain a signal reconstruction network in which the signal reconstruction loss, periodic consistency loss, and fault classification loss all meet the preset loss thresholds. The pure fault features are reconstructed using the signal reconstruction network to generate the enhanced fault impulse signal.

10. An intelligent detection system for the operating performance of industrial equipment, characterized in that, The system is used to implement the intelligent detection method for industrial equipment operating performance as described in any one of claims 1-9, and the system comprises: The prior feature extractor pre-training module is used to collect the first vibration sample dataset of the industrial equipment to be inspected, apply random resampling transformation to the vibration signal samples in the first vibration sample dataset to simulate speed change, and pre-train the prior feature extractor. The prior feature extractor optimization module is used to construct a fault classifier and a rotational speed discriminator, and to jointly train and optimize the prior feature extractor using a second vibration sample dataset through an adversarial domain adaptation approach to obtain a target fault classifier and a target feature extractor. The first training objective of the adversarial domain adaptation is to optimize the prior feature extractor and the fault classifier to maximize the fault classification accuracy, and the second training objective is to optimize the prior feature extractor to minimize the rotational speed classification accuracy of the rotational speed discriminator. The equipment operation detection module is used to collect real-time vibration signals of the industrial equipment under inspection, input them into the target feature extractor for feature extraction, and then input them into the target fault classifier to obtain equipment operation detection information.

Images