Abnormal diagnosis method and system for an oil-free dry vacuum unit

Abstract

The present application provides an abnormal diagnosis method and system for an oil-free dry vacuum unit, which respectively detect abnormal operation data for normal instantaneous disturbance learning examples and abnormal instantaneous disturbance learning examples, and correspondingly obtain a first detection result and a second detection result. According to the prior labels of normal learning examples, the prior labels of abnormal learning examples, the first detection result, and the second detection result, the error value of the first error determination function is determined; similarly, the first error value of the second error determination function is determined; by training the abnormal operation diagnosis network of the unit simultaneously based on continuous disturbance information learning examples and instantaneous disturbance information learning examples, the cardinality of the training data of the network can be increased, overfitting of the network can be prevented, and moreover, the effect of the network for detecting multi-dimensional abnormal operation data can be increased. Coordinating based on the error values of the two error determination functions to optimize the network parameters can increase the accuracy and reliability of the network for abnormal diagnosis.

Description

Technical Field

[0001] This application relates to the technical field of data processing, and in particular, to an abnormal diagnosis method and system for an oil-free dry vacuum unit. Background Art

[0002] In the industrial production process, as a key device, the stable operation of an oil-free dry vacuum unit is crucial for ensuring production efficiency and product quality. However, due to the complexity and variability of the unit's operating environment, the unit is prone to being affected by various internal and external factors during operation, resulting in abnormal operating states. If these abnormal states cannot be detected and handled in a timely manner, they will seriously threaten the stability and safety of the unit and even trigger major accidents. Traditional methods for diagnosing unit abnormalities mainly rely on manual inspections and empirical judgments. This method is not only inefficient but also easily affected by human factors, leading to misdiagnosis and missed diagnosis. With the rapid development of machine learning and artificial intelligence technologies, using these advanced technologies to build an automated and intelligent unit abnormal operation diagnosis system has become a new trend. Existing machine learning-based abnormal diagnosis systems often face problems of insufficient data and overfitting during the training process. On the one hand, due to the scarcity and difficulty of collecting actual operation data, the quantity and quality of training samples often fail to meet the requirements. On the other hand, due to the complexity and diversity of the unit's operating states, it is difficult for a single type of training data to comprehensively cover all possible abnormal situations. Summary of the Invention

[0003] In view of this, this application provides an abnormal diagnosis method and system for an oil-free dry vacuum unit. The technical solution of the embodiments of this application is implemented as follows:

[0004] On the one hand, an embodiment of the present application provides an abnormal diagnosis method for an oil-free dry vacuum unit, and the method includes: respectively detecting abnormal operation data of a normal instantaneous disturbance learning sample and an abnormal instantaneous disturbance learning sample based on a unit abnormal operation diagnosis network, and correspondingly obtaining a first detection result and a second detection result; wherein, the normal instantaneous disturbance learning sample is obtained by integrating instantaneous disturbance information into a normal learning sample, and the abnormal instantaneous disturbance learning sample is obtained by integrating instantaneous disturbance information into an abnormal learning sample corresponding to the normal learning sample; determining an error value of a first error determination function according to a prior label of the normal learning sample, a prior label of the abnormal learning sample, the first detection result, and the second detection result; respectively detecting abnormal operation data of a normal continuous disturbance learning sample and an abnormal continuous disturbance learning sample based on the unit abnormal operation diagnosis network, and correspondingly obtaining a third detection result and a fourth detection result; wherein, the normal continuous disturbance learning sample is obtained by integrating continuous disturbance information into the normal learning sample, and the abnormal continuous disturbance learning sample is obtained by integrating continuous disturbance information into the abnormal learning sample; determining a first error value of a second error determination function according to a prior label of the normal continuous disturbance learning sample, a prior label of the abnormal continuous disturbance learning sample, the third detection result, and the fourth detection result; optimizing network parameters of the unit abnormal operation diagnosis network according to the error value of the first error determination function and the first error value of the second error determination function.

[0005] In some embodiments, the method further includes: respectively detecting abnormal operation data of a normal learning sample and an abnormal learning sample corresponding to the normal learning sample based on the unit abnormal operation diagnosis network, and correspondingly obtaining a fifth detection result and a sixth detection result; determining a first error value of a third error determination function according to a prior label of the normal learning sample, a prior label of the abnormal learning sample, the fifth detection result, and the sixth detection result; the optimizing network parameters of the unit abnormal operation diagnosis network according to the error value of the first error determination function and the first error value of the second error determination function includes: optimizing network parameters of the unit abnormal operation diagnosis network according to the error value of the first error determination function, the first error value of the second error determination function, and the first error value of the third error determination function.

[0006] In some embodiments, optimizing the network parameters of the unit abnormal operation diagnosis network based on the error value of the first error determination function, the first error value of the second error determination function, and the first error value of the third error determination function includes: respectively obtaining the importance coefficients of the first error determination function, the second error determination function, and the third error determination function; obtaining the error value of the total error determination function based on the error value and importance coefficient of the first error determination function, the first error value and importance coefficient of the second error determination function, and the first error value and importance coefficient of the third error determination function; optimizing the network parameters of the unit abnormal operation diagnosis network based on the error value of the total error determination function; Before the unit abnormal operation diagnosis network respectively performs abnormal operation data detection on the normal instantaneous disturbance learning samples and the abnormal instantaneous disturbance learning samples, the method further includes: swapping the prior labels of the normal learning samples and the prior labels of the abnormal learning samples to obtain the false prior labels of the normal learning samples and the false prior labels of the abnormal learning samples; determining the second error value of the third error determination function based on the false prior labels of the normal learning samples, the false prior labels of the abnormal learning samples, the fifth detection result, and the sixth detection result; generating instantaneous disturbance information based on the second error value of the third error determination function; respectively integrating the instantaneous disturbance information into the normal learning samples and the abnormal learning samples to correspondingly obtain the normal instantaneous disturbance learning samples and the abnormal instantaneous disturbance learning samples.

[0007] In some embodiments, generating instantaneous disturbance information based on the second error value of the third error determination function includes: determining the rate of change of the third error determination function with respect to the network parameters based on the second error value of the third error determination function to obtain the optimal descent direction and its rate of the unit abnormal operation diagnosis network; determining the optimal descent direction and its rate of the unit abnormal operation diagnosis network as the instantaneous disturbance information; The step of respectively integrating the instantaneous disturbance information into the normal learning samples and the abnormal learning samples to correspondingly obtain the normal instantaneous disturbance learning samples and the abnormal instantaneous disturbance learning samples includes: determining the importance coefficient of the instantaneous disturbance information, and performing a weighting operation on the instantaneous disturbance information according to the importance coefficient to obtain weighted instantaneous disturbance information; respectively integrating the weighted instantaneous disturbance information into the normal learning samples and the abnormal learning samples to obtain the normal instantaneous disturbance learning samples and the abnormal instantaneous disturbance learning samples.

[0008] In some embodiments, the abnormal operation data detection is performed based on a classification component in the unit abnormal operation diagnosis network, and the unit abnormal operation diagnosis network further includes a synthesis component; before the abnormal operation data detection is respectively performed on the normal continuous disturbance learning sample and the abnormal continuous disturbance learning sample based on the unit abnormal operation diagnosis network, the method further includes: based on the synthesis component in the unit abnormal operation diagnosis network, generating continuous disturbance information according to the loaded random disturbance; and integrating the continuous disturbance information into the normal learning sample and the abnormal learning sample respectively, to correspondingly obtain the normal continuous disturbance learning sample and the abnormal continuous disturbance learning sample.

[0009] In some embodiments, the integrating the continuous disturbance information into the normal learning sample and the abnormal learning sample respectively, to correspondingly obtain the normal continuous disturbance learning sample and the abnormal continuous disturbance learning sample includes: determining an importance coefficient of the continuous disturbance information, and performing a weighting operation on the continuous disturbance information according to the importance coefficient to obtain importance coefficient continuous disturbance information; and integrating the importance coefficient continuous disturbance information into the normal learning sample and the abnormal learning sample respectively, to obtain the normal continuous disturbance learning sample and the abnormal continuous disturbance learning sample.

[0010] In some embodiments, after optimizing the network parameters of the unit abnormal operation diagnosis network according to the error value of the first error determination function and the first error value of the second error determination function, the method further includes: respectively performing abnormal operation data detection on the normal continuous disturbance learning sample and the abnormal continuous disturbance learning sample based on the optimized unit abnormal operation diagnosis network, to correspondingly obtain a seventh detection result and an eighth detection result; determining a second error value of the second error determination function according to the prior label of the normal learning sample, the prior label of the abnormal learning sample, the seventh detection result, and the eighth detection result; and optimizing the network parameters of the optimized unit abnormal operation diagnosis network according to the error value of the first error determination function and the second error value of the second error determination function.

[0011] In some embodiments, before the abnormal operation data detection is respectively performed on the normal instantaneous disturbance learning sample and the abnormal instantaneous disturbance learning sample based on the unit abnormal operation diagnosis network, the method further includes: obtaining initial normal learning samples of several dimensions, and obtaining initial abnormal learning samples of several dimensions corresponding to the initial normal learning samples; merging the initial normal learning samples of several dimensions to obtain the normal learning sample; and merging the initial abnormal learning samples of several dimensions to obtain the abnormal learning sample.

[0012] In some embodiments, after determining the error value of the first error determination function and the first error value of the second error determination function, and optimizing the network parameters of the abnormal operation diagnosis network of the unit, the method further includes: when the abnormal operation diagnosis network of the unit reaches a convergence state, stopping the optimization of the network parameters of the abnormal operation diagnosis network of the unit to obtain a target abnormal operation diagnosis network of the unit; acquiring a sensor monitoring data set of the unit to be diagnosed; and detecting abnormal operation data through the target abnormal operation diagnosis network of the unit for the sensor monitoring data set of the unit to be diagnosed, so as to obtain a detection result indicating whether the sensor monitoring data set of the unit to be diagnosed includes abnormal operation data.

[0013] In a second aspect, the present application provides a computer system, including a memory and a processor, where the memory stores a computer program that can be run on the processor, and when the processor executes the program, the steps in the above method are implemented.

[0014] The beneficial effects of the present application at least include: the present application respectively performs abnormal operation data detection on normal instantaneous disturbance learning examples and abnormal instantaneous disturbance learning examples, and correspondingly obtains a first detection result and a second detection result. According to the prior labels of normal learning examples, the prior labels of abnormal learning examples, the first detection result, and the second detection result, the error value of the first error determination function is determined; respectively performs abnormal operation data detection on normal continuous disturbance learning examples and abnormal continuous disturbance learning examples, and correspondingly obtains a third detection result and a fourth detection result; according to the prior labels of normal continuous disturbance learning examples, the prior labels of abnormal continuous disturbance learning examples, the third detection result, and the fourth detection result, determines the first error value of the second error determination function; by training the abnormal operation diagnosis network of the unit based on both continuous disturbance information learning examples and instantaneous disturbance information learning examples, the cardinality of the training data of the network can be increased, overfitting of the network can be prevented, and the effect of the network for detecting multi-dimensional abnormal operation data can be increased; according to the error value of the first error determination function and the first error value of the second error determination function, the network parameters of the abnormal operation diagnosis network of the unit are optimized, and the optimization of the network parameters is realized by collaborating based on the error values of the two error determination functions, which can increase the accuracy and reliability of the network for abnormal diagnosis. Description of the Drawings

[0015] The accompanying drawings here are incorporated into the specification and constitute a part of this specification. These drawings show embodiments consistent with the present application and are used together with the specification to explain the technical solutions of the present application.

[0016] Figure 1 It is a schematic flowchart of the implementation of an abnormal diagnosis method for an oil-free dry vacuum unit provided by an embodiment of the present application.

[0017] Figure 2 Schematic diagram of the hardware entity of a computer system provided by an embodiment of the present application. Detailed implementation manners

[0018] In order to make the objectives, technical solutions, and advantages of the present application clearer, the technical solutions of the present application will be further elaborated in detail below with reference to the accompanying drawings and embodiments. The described embodiments should not be construed as limitations on the present application. All other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present application.

[0019] An embodiment of the present application provides an abnormal diagnosis method for an oil-free dry vacuum unit, and this method can be executed by a processor of a computer system. Among them, the computer system may refer to devices with data processing capabilities such as servers, laptops, tablets, desktop computers, etc.

[0020] Figure 1 Schematic diagram of the implementation process of an abnormal diagnosis method for an oil-free dry vacuum unit provided by an embodiment of the present application, as Figure 1 shown, this method includes:

[0021] Step S100: Based on the abnormal operation diagnosis network of the unit, detect the abnormal operation data of the normal instantaneous disturbance learning example and the abnormal instantaneous disturbance learning example respectively, and obtain the first detection result and the second detection result correspondingly; among them, the normal instantaneous disturbance learning example is obtained by integrating instantaneous disturbance information into the normal learning example, and the abnormal instantaneous disturbance learning example is obtained by integrating instantaneous disturbance information into the abnormal learning example corresponding to the normal learning example.

[0022] In the training stage of the abnormal operation diagnosis network of the unit, the computer system first constructs an abnormal operation diagnosis network of the unit. This network is, for example, a complex neural network model, such as a combination of a convolutional neural network (CNN) or a recurrent neural network (RNN) and a long short-term memory unit (LSTM). The specific selection depends on the characteristics of the sensing data and the complexity of the diagnosis task. This network is designed to identify abnormal patterns in the unit operation data.

[0023] To enhance the generalization ability and robustness of this network, the computer system executes step S100. First, the computer system prepares a batch of normal learning examples, which are composed of sensing data generated by the oil-free dry vacuum unit in the normal operation state. These sensing data may include various parameters such as pressure, temperature, flow rate, vibration, etc. The values of each parameter at different time points constitute a multi-dimensional feature vector. For example, at a certain time point t, the feature vector can be [pressure value_t, temperature value_t, flow rate value_t, vibration amplitude_t].

[0024] Next, to generate normal transient perturbation learning examples, the computer system incorporates transient perturbation information based on these normal learning examples. The transient perturbation information is the tiny noise randomly generated during a single forward or backward propagation process, such as Gaussian noise. These noises are added to each element of the original feature vector, and the generated new feature vector may become [pressure value_t + noise_pressure, temperature value_t + noise_temperature, flow value_t + noise_flow, vibration amplitude_t + noise_vibration]. Such an operation simulates the data fluctuations caused by measurement errors or minor environmental changes during actual operation.

[0025] Similarly, to generate abnormal transient perturbation learning examples, the computer system first creates corresponding abnormal learning examples based on the normal learning examples. This is usually achieved by artificially modifying some data in the normal learning examples to simulate abnormal states, such as setting the pressure value or temperature value at a certain time point to an abnormal value beyond the normal range. Then, in the same way as the generation of normal transient perturbation learning examples, these abnormal learning examples are also added with transient perturbation information.

[0026] After the generation of the learning examples is completed, the computer system inputs these two types of transient perturbation learning examples into the unit abnormal operation diagnosis network for the detection of abnormal operation data. The network processes these feature vectors through its internal weight and bias parameters and outputs the probability that each example belongs to abnormal or normal. For normal transient perturbation learning examples, it is expected that the output of the network is close to the probability of the normal category; while for abnormal transient perturbation learning examples, it is expected that the output is close to the probability of the abnormal category.

[0027] To evaluate the performance of the network, the computer system compares the output of the network (i.e., the first detection result and the second detection result) with the prior label of the example (i.e., the true category) and calculates the error value of the loss function (such as cross-entropy loss). The calculation formula of the loss function is as follows, for example:

[0028]

[0029] where N is the total number of examples, y i is the true category of the i-th example (0 or 1, representing normal or abnormal), is the predicted probability of the network for the i-th example. This error value is then used for the optimization of the network in subsequent steps. Through step S100, the computer system not only trains the unit abnormal operation diagnosis network to identify normal and abnormal data, but also enhances the network's robustness to local minor changes in the training data by introducing transient perturbation information, which helps to reduce overfitting and improve the accuracy and reliability of the diagnosis.

[0030] Step S200: Determine the error value of the first error determination function based on the prior labels of the normal learning examples, the prior labels of the abnormal learning examples, the first detection result, and the second detection result.

[0031] During the process of training the abnormal operation diagnosis network of the unit, the computer system has completed Step S100, that is, it has detected the abnormal operation data of the normal instantaneous disturbance learning examples and the abnormal instantaneous disturbance learning examples, and obtained the first detection result and the second detection result respectively. These detection results reflect the prediction probabilities of the network for whether the current input example belongs to the normal or abnormal category. In Step S200, the computer system will use these detection results and the prior labels of the examples to calculate the error value of the first error determination function. Here, the first error determination function, in the context of machine learning, usually refers to the loss function, which is used to quantify the difference between the model's prediction result and the true result.

[0032] For the abnormal diagnosis task of the oil-free dry vacuum unit, the choice of the loss function may depend on the specific network structure and task requirements. A commonly used loss function is the cross-entropy loss, which is particularly suitable for dealing with classification problems.

[0033] Suppose there is a batch of normal learning examples, and their prior labels (true categories) are all 0 (indicating normal). In Step S100, the abnormal operation diagnosis network of the unit makes predictions on these examples and outputs the prediction probabilities that each example belongs to the abnormal category (i.e., ). Ideally, these prediction probabilities should be close to 0, indicating that the network correctly identifies these examples as normal.

[0034] Similarly, for the abnormal learning examples, their prior labels are all 1 (indicating abnormal). In Step S100, the network also outputs the prediction probabilities that these examples belong to the abnormal category. Here, the prediction probabilities should be close to 1, indicating that the network correctly identifies these examples as abnormal.

[0035] The computer system substitutes the prediction probability of each example and its prior label into the cross-entropy loss function, calculates the loss value of each example, and then averages the loss values of all examples to obtain the final error value of the first error determination function. This error value reflects the deviation degree between the network's prediction result and the actual result.

[0036] For example, if there is a normal learning example with a prior label of 0, but the network predicts that the probability of it belonging to the abnormal category is 0.3 (i.e., ) Then, according to the cross-entropy loss function, the contribution of this sample to the total error is: -[0·log(0.3)+(1 - 0)·log(1 - 0.3)] = -log(0.7).

[0037] By calculating the losses of all samples and summing and averaging them, the computer system obtains the error value of the first error determination function, which will be used in the subsequent steps to optimize the network to reduce the prediction error and improve the diagnostic accuracy.

[0038] Step S300: Based on the unit abnormal operation diagnosis network, detect the abnormal operation data of the normal continuous disturbance learning samples and the abnormal continuous disturbance learning samples respectively, and obtain the third detection result and the fourth detection result correspondingly; wherein, the normal continuous disturbance learning samples are obtained by integrating continuous disturbance information into the normal learning samples, and the abnormal continuous disturbance learning samples are obtained by integrating continuous disturbance information into the abnormal learning samples.

[0039] During the training process, the computer system has constructed the unit abnormal operation diagnosis network and prepared the normal learning samples and the abnormal learning samples as training data. To further challenge the network and improve its adaptability to complex and long-term changes, step S300 introduces continuous disturbance information. The continuous disturbance information is not simple random noise, but is generated by a synthesis component. This synthesis component itself can be the generator in the GAN network. It takes a random disturbance vector as input and processes it through a multi-layer neural network to gradually generate high-quality noise data. These noise data are designed to be able to affect the feature distribution of the samples in a long-term and continuous manner, simulating the complex changes in the unit operation data in the real world.

[0040] Specifically, the computer system first generates continuous disturbance information for the normal learning samples and the abnormal learning samples respectively through the synthesis component. Taking the normal learning samples as an example, each sample is originally a multi-dimensional feature vector, containing sensing data of various parameters such as pressure, temperature, flow rate, and vibration. The continuous disturbance information, as another group of feature vectors, is superimposed on the original feature vector in the same dimension to generate the normal continuous disturbance learning samples. In this process, the original feature vector may become [pressure value_t, temperature value_t, flow rate value_t, vibration amplitude_t], and the continuous disturbance information vector may be [disturbance_pressure_t, disturbance_temperature_t, disturbance_flow_rate_t, disturbance_vibration_t]. The new feature vector after superposition is [pressure value_t + disturbance_pressure_t, temperature value_t + disturbance_temperature_t, flow rate value_t + disturbance_flow_rate_t, vibration amplitude_t + disturbance_vibration_t].

[0041] Similarly, the abnormal learning samples are also given corresponding continuous disturbance information to generate the abnormal continuous disturbance learning samples.

[0042] Subsequently, the computer system inputs these learning examples with continuous perturbation information into the abnormal operation diagnosis network of the unit to detect abnormal operation data. Through its internal complex calculation process (which may include various operations such as convolution, pooling, activation, and fully connected), the network processes the input feature vectors and outputs the probability that each example belongs to the abnormal category. These outputs respectively correspond to the third detection result (for normal continuous perturbation learning examples) and the fourth detection result (for abnormal continuous perturbation learning examples).

[0043] Different from the instantaneous perturbation in step S100, the continuous perturbation aims to challenge the network through long-term and complex feature changes, forcing it to learn the essential features closely related to the prior label (i.e., the true category), rather than just the superficial and easily captured features of the training samples. This training method helps to improve the robustness and accuracy of the network in practical applications because it requires the network to be able to identify abnormal patterns that remain stable even in a continuously changing environment.

[0044] In practical applications, the computer system can update the weight and bias parameters of the network according to the differences between the third detection result and the fourth detection result and the prior labels of the examples. This update is usually achieved through the backpropagation algorithm, where the gradient of the loss function (such as cross-entropy loss) is used to guide the adjustment direction of the network parameters. By repeatedly iterating this process, the computer system can gradually optimize the performance of the abnormal operation diagnosis network of the unit.

[0045] Step S400: Determine the first error value of the second error determination function based on the prior label of the normal continuous perturbation learning example, the prior label of the abnormal continuous perturbation learning example, the third detection result, and the fourth detection result.

[0046] In step S400, the computer system uses the learning examples enhanced with continuous perturbation information (i.e., normal continuous perturbation learning examples and abnormal continuous perturbation learning examples) and the detection results (the third detection result and the fourth detection result) obtained by these examples through the abnormal operation diagnosis network of the unit to calculate the first error value of the second error determination function.

[0047] In step S300, the computer system has added continuous perturbation information to the normal learning examples and abnormal learning examples respectively through the synthesis component, generating normal continuous perturbation learning examples and abnormal continuous perturbation learning examples. These continuous perturbation information aims to simulate more complex and long-term changes in the real world by changing the feature distribution of the samples, thereby forcing the network to learn more robust feature representations.

[0048] In step S400, the computer system inputs these learning examples with continuous perturbation information into the already trained unit abnormal operation diagnosis network for detecting abnormal operation data. Based on its internal complex structure and parameters, the network processes these input feature vectors and outputs the prediction probabilities that each example belongs to the abnormal category, namely the third detection result and the fourth detection result.

[0049] To evaluate the performance of the network under continuous perturbation conditions, the computer system compares these prediction results with the prior labels (i.e., the true categories) of the examples. Here, the prior label of a normal continuous perturbation learning example should be the normal category (e.g., represented by 0), while the prior label of an abnormal continuous perturbation learning example should be the abnormal category (e.g., represented by 1).

[0050] Then, the computer system uses an error determination function (which can be called the second error determination function) to calculate the difference between the prediction result and the prior label. This function is usually a loss function used to quantify the error degree of the network prediction.

[0051] The computer system calculates the loss values of normal continuous perturbation learning examples and abnormal continuous perturbation learning examples respectively, and sums and averages them to obtain the first error value of the second error determination function. This error value reflects the detection ability of the network for abnormal operation data under continuous perturbation conditions.

[0052] By comparing this error value with a preset threshold or a previous error value, the computer system can evaluate whether the performance of the network has improved or whether it is necessary to further adjust the parameters and structure of the network. If the error value is high, it indicates that the network performs poorly under continuous perturbation conditions, and it may be necessary to increase the number of training iterations, adjust the learning rate, change the network structure, or try other optimization strategies to improve the performance.

[0053] Step S500: Optimize the network parameters of the unit abnormal operation diagnosis network according to the error value of the first error determination function and the first error value of the second error determination function.

[0054] In step S500, specifically, the computer system collects two error values: one is the error value of the first error determination function based on the normal instantaneous perturbation learning examples and abnormal instantaneous perturbation learning examples calculated through step S200; the other is the first error value of the second error determination function based on the normal continuous perturbation learning examples and abnormal continuous perturbation learning examples calculated through step S400. These two error values respectively reflect the performance of the network under instantaneous perturbation and continuous perturbation conditions. To optimize the network, the computer system can use the Backpropagation Algorithm to update the network's parameters such as weights and biases. The core idea of the backpropagation algorithm is to calculate the gradient of the loss function with respect to the network parameters and update the parameters in the opposite direction of the gradient to minimize the loss function. In this process, the two error values generated by instantaneous perturbation and continuous perturbation will jointly act on the network optimization process.

[0055] In practical applications, since the two error values may have different magnitudes and importance, the computer system can assign different importance coefficients, i.e., weights, to them to give appropriate attention during the optimization process. These weights can be adjusted according to the actual situation to ensure that the network can perform well under different perturbation conditions. The optimization process can be iterated until a certain stopping criterion is met. The stopping criterion may include the error value dropping below a preset threshold, the number of iterations reaching a preset upper limit, or the network performance not significantly improving in consecutive iterations. Through the optimization process of step S500, the computer system can continuously improve the performance of the abnormal operation diagnosis network of the unit, enabling it to better adapt to various complex situations during the operation of the oil-free dry vacuum unit, including instantaneous perturbation and continuous perturbation, etc., so as to achieve more accurate abnormal diagnosis.

[0056] As an implementation manner, the method further includes:

[0057] Step S600: Based on the abnormal operation diagnosis network of the unit, respectively detect the abnormal operation data of the normal learning examples and the abnormal learning examples corresponding to the normal learning examples, and obtain the fifth detection result and the sixth detection result correspondingly;

[0058] Step S700: Determine the first error value of the third error determination function according to the prior label of the normal learning example, the prior label of the abnormal learning example, the fifth detection result, and the sixth detection result;

[0059] The step S500, according to the error value of the first error determination function and the first error value of the second error determination function, optimizing the network parameters of the abnormal operation diagnosis network of the unit may include:

[0060] Step S510: Optimize the network parameters of the unit abnormal operation diagnosis network according to the error value of the first error determination function, the first error value of the second error determination function, and the first error value of the third error determination function.

[0061] In step S600, the computer system extracts normal learning examples and abnormal learning examples from the training dataset. The normal learning examples contain the sensing data collected when the unit is in normal operation, and these data reflect the normal ranges of various parameters of the unit (such as pressure, temperature, flow rate, vibration, etc.). For example, the feature vector of a normal learning example may be represented as [0.8 MPa, 25 °C, 50 L / min, 0.05 mm / s], corresponding to pressure, temperature, flow rate, and vibration speed respectively. The abnormal learning examples are obtained by modifying the normal learning examples to simulate the sensing data of the unit in abnormal operation. For example, by modifying the pressure value in the above normal learning example to an abnormal value (such as 1.5 MPa), the feature vector of the abnormal learning example [1.5 MPa, 25 °C, 50 L / min, 0.1 mm / s] is obtained.

[0062] Next, the computer system inputs these original learning examples into the unit abnormal operation diagnosis network that has been preliminarily trained. Through its internal complex calculation logic (which may include convolutional layers, pooling layers, fully connected layers, etc.), the network extracts features and makes classification judgments for each input example, and outputs the predicted probability that each example belongs to the abnormal category.

[0063] For the normal learning examples, the network should output a lower abnormal probability (close to 0), indicating that it is judged to be in a normal state; for the abnormal learning examples, the network should output a higher abnormal probability (close to 1), indicating that it is judged to be in an abnormal state. These outputs correspond to the fifth detection result and the sixth detection result in step S600 respectively.

[0064] After obtaining the detection results of the original learning examples, the computer system evaluates the differences between these results and the actual categories (i.e., prior labels) to further guide the optimization of the network. This task is completed through step S700, where a third error determination function is introduced to calculate the error value.

[0065] The specific calculation process is as follows. First, the computer system compares the prior labels of the normal learning examples and the abnormal learning examples with the corresponding detection results (the fifth detection result and the sixth detection result). The prior labels are known, the prior label of the normal learning example is 0 (indicating normal), and the prior label of the abnormal learning example is 1 (indicating abnormal). Then, use the third error determination function (which can also be a cross-entropy loss function or other suitable classification loss functions) to calculate the error value.

[0066] After obtaining the error value of the first error determination function (based on instantaneous perturbation learning examples), the first error value of the second error determination function (based on continuous perturbation learning examples), and the first error value of the third error determination function (based on original learning examples), the computer system executes step S510 to optimize the network parameters of the unit abnormal operation diagnosis network by synthesizing these error values.

[0067] The specific optimization process can be as follows: First, the computer system assigns an importance coefficient (Weight) to each error determination function to reflect their relative importance in the optimization process. These importance coefficients can be adjusted according to actual needs to balance the performance of the network under different conditions. For example, if it is considered that continuous perturbation has a greater impact on the robustness of the network, a higher importance coefficient can be assigned to the second error determination function.

[0068] Then, calculate the weighted total error value. The weighted total error value can be obtained by summing the product of each error value and its corresponding importance coefficient. Finally, use the backpropagation algorithm and gradient descent (or other optimization algorithms) to update the network weights, biases, and other parameters to minimize the weighted total error value. In each iteration, calculate the gradient of the weighted total error value with respect to each network parameter and update the parameter values according to the gradient direction. By repeatedly iterating this process, the computer system gradually adjusts the network parameters so that the network can exhibit good performance under instantaneous perturbation, continuous perturbation, and standard conditions.

[0069] By introducing steps S600 and S700, the computer system can consider the performance of the original learning examples without perturbation processing during the optimization process, thereby further enhancing the generalization ability and accuracy of the unit abnormal operation diagnosis network.

[0070] Step S510 guides the optimization direction of the network by synthesizing multiple error values to ensure that the network can maintain stable performance under different conditions. This method of multi-error comprehensive optimization helps to improve the reliability and practicality of the network in actual applications. In the embodiments of this application, this optimization method is particularly important because the operating state of the unit is complex and changeable and may be affected by various internal and external factors. By introducing instantaneous perturbation and continuous perturbation information and comprehensively considering the performance of the original learning examples, the computer system can train a more robust and accurate abnormal diagnosis network to provide strong guarantee for the stable operation of the unit.

[0071] As an implementation manner, step S510, according to the error value of the first error determination function, the first error value of the second error determination function, and the first error value of the third error determination function, to optimize the network parameters of the unit abnormal operation diagnosis network, may include:

[0072] Step S511: Obtain the importance coefficient of the first error determination function, the importance coefficient of the second error determination function, and the importance coefficient of the third error determination function respectively;

[0073] Step S512: Based on the error value and importance coefficient of the first error determination function, the first error value and importance coefficient of the second error determination function, and the first error value and importance coefficient of the third error determination function, obtain the error value of the total error determination function;

[0074] Step S513: Optimize the network parameters of the unit abnormal operation diagnosis network based on the error value of the total error determination function;

[0075] In the embodiments of the present application, in order to improve the performance of the unit abnormal operation diagnosis network, the computer system adopts a multi-error weighted optimization strategy. The core of this strategy is to comprehensively consider the error values from different error determination functions and guide the network optimization process by assigning appropriate importance coefficients (weights) to each error value. In step S511, the computer system determines an importance coefficient for the first error determination function (based on instantaneous perturbation learning examples), the second error determination function (based on continuous perturbation learning examples), and the third error determination function (based on original learning examples) respectively. These importance coefficients reflect the relative importance of different error values in the optimization process.

[0076] For the specific setting of the importance coefficient, it can be determined based on methods such as expert experience, historical data analysis, or cross-validation. For example, if historical data shows that continuous perturbation has a more significant impact on unit abnormalities, the importance coefficient of the second error determination function can be set relatively high. On the contrary, if the instantaneous perturbation is frequent but has a small impact, the importance coefficient of the corresponding first error determination function can be appropriately reduced. The importance coefficient of the third error determination function depends on the representativeness of the original data in the training set.

[0077] After determining the importance coefficients of each error determination function, the computer system then multiplies these coefficients by the corresponding error values and sums the results to obtain the error value of the total error determination function. This total error value will be used as the basis for subsequent optimization steps.

[0078] This total error value combines information from different error determination functions and reflects the overall performance of the network under different conditions.

[0079] After obtaining the total error value, the computer system will use this value to guide the optimization process of the unit abnormal operation diagnosis network. Specifically, it is to update the network weights and biases and other parameters through the backpropagation algorithm and gradient descent (or other optimization algorithms) to minimize the total error value.

[0080] The optimization process may include, for example: First, the computer system calculates the gradient of the total error value with respect to each network parameter. This gradient indicates how the parameter should be adjusted to reduce the total error value. In a neural network, the gradient of each parameter is calculated by backpropagating the error signal layer by layer. After obtaining the gradient, the computer system updates the network parameters according to the rules of an optimization algorithm (such as gradient descent). Specifically, each parameter is moved a small step along the direction of its negative gradient (the step size is controlled by the learning rate) in the hope of reducing the total error value. The above parameter update process will be repeated for all training examples within one training cycle until a certain stopping criterion is met (such as the error value drops below a preset threshold, the number of iterations reaches the upper limit, etc.). In each iteration, the network adjusts its parameters based on the current error value and gradient information to gradually approach the optimal solution.

[0081] In the embodiments of the present application, the optimization process may involve a large amount of sensing data and high-dimensional feature vectors. For example, each sensing data point may contain multiple parameters (such as pressure, temperature, flow rate, etc.), and these data points may change dynamically over time. Therefore, during the optimization process, the computer system processes a series of high-dimensional and dynamic data sets. To cope with this complexity, the computer system can adopt advanced neural network structures (such as convolutional neural network CNN or recurrent neural network RNN) to construct a diagnostic network for abnormal operation of the unit. These network structures can effectively extract useful features from high-dimensional data and approximate complex input-output relationships through multi-layer non-linear transformations.

[0082] During the optimization process, the computer continuously adjusts the weight and bias parameters of each layer of the network according to the total error value, in the hope that the network can make accurate judgments for different types of disturbances (instantaneous disturbances and continuous disturbances) and the original data. Through repeated iterative optimization, a diagnostic network for abnormal operation of the unit with stable performance and high accuracy is finally obtained. This network will be deployed in the actual oil-free dry vacuum unit to monitor and diagnose the operating state of the unit in real time, ensuring the safe and reliable operation of the unit.

[0083] As an implementation, before the method respectively detects abnormal operation data for the normal instantaneous disturbance learning examples and the abnormal instantaneous disturbance learning examples based on the diagnostic network for abnormal operation of the unit in step S100, the method may further include:

[0084] Step S101: Swap the prior labels of the normal learning examples and the prior labels of the abnormal learning examples to obtain the false prior labels of the normal learning examples and the false prior labels of the abnormal learning examples;

[0085] Step S102: Determine the second error value of the third error determination function according to the pseudo-prior labels of the normal learning examples, the pseudo-prior labels of the abnormal learning examples, the fifth detection result, and the sixth detection result;

[0086] Step S103: Generate instantaneous perturbation information according to the second error value of the third error determination function;

[0087] Step S104: Incorporate the instantaneous perturbation information into the normal learning examples and the abnormal learning examples respectively to obtain the normal instantaneous perturbation learning examples and the abnormal instantaneous perturbation learning examples correspondingly.

[0088] In the embodiments of the present application, in order to improve the robustness and generalization ability of the abnormal operation diagnosis network of the unit, before the computer system executes step S100, a series of pre-steps will be taken to generate more effective training samples. These steps aim to enhance the network's adaptability to complex perturbations by introducing a combined strategy of a special data processing technique - Label Smoothing and Adversarial Training.

[0089] For example, in the actual operation of an oil-free dry vacuum unit, the unit state can be divided into normal and abnormal. In order to train a diagnosis network that can accurately identify these two states, the computer system first needs to collect a large amount of normal operation data and abnormal operation data as training samples. These data are divided into normal learning examples and abnormal learning examples, and are respectively given prior labels of 0 (indicating normal) and 1 (indicating abnormal).

[0090] However, in traditional supervised learning, hard labels (i.e., 0 and 1) can cause the network to predict the category of each sample too confidently, thus reducing its generalization ability for uncertain samples. To alleviate this problem, in step S101, the prior labels of the normal learning examples and the abnormal learning examples are swapped.

[0091] Specifically, assume there are N normal learning examples with prior label y normal = [0, 0, …, 0] (a zero vector of length N), and M abnormal learning examples with prior label y abnormal = [1, 1, …, 1] (a one vector of length M). In step S101, the computer system swaps the elements of these two vectors to obtain pseudo-prior label vectors y fake,normal = [1, 1, …, 1] (a one vector of length N) and y fake,abnormal = [0, 0, …, 0] (a zero vector of length M).

[0092] In this way, the network will be forced to learn more subtle feature differences to improve its robustness to label noise.

[0093] In step S102, the computer system uses the unit abnormal operation diagnosis network to detect the normal learning examples and abnormal learning examples (assuming this process has been completed in step S600 here), and obtains the fifth detection result (for normal learning examples) and the sixth detection result (for abnormal learning examples). These detection results reflect the performance of the network under the correct labels.

[0094] Subsequently, the computer system uses the false prior labels (y fake,normal and y fake,abnormal ) and the corresponding detection results to calculate the second error value of the third error determination function. Here, the cross-entropy loss function can still be used as the error determination function, but the labels used at this time are the swapped false labels. Example of the calculation formula:

[0095]

[0096] Since the false prior labels are all 1 or all 0, some terms in the above formula will be simplified to 0 (for example, when y fake,normal,i = 1, 1 - y fake,normal,i = 0, so this term is always 0). However, this simplification does not affect the calculation of the error value because the purpose is to evaluate the overall performance of the network under the wrong labels.

[0097] In step S103, the computer system uses the second error value of the third error determination function (i.e., the error value under the false labels) to generate instantaneous perturbation information. This is a strategy of adversarial training, aiming to enhance the robustness of the network by introducing samples that conflict with the current network decision boundary.

[0098] Specifically, the computer system can guide the generation of instantaneous perturbation information according to the magnitude and direction of the second error value. A common method is to use gradient ascent (opposite to gradient descent) to find the perturbation direction that can maximize the error value. However, in practical applications, directly maximizing the error value can lead to the generation of overly extreme perturbation samples, which may be very different from the real-world distribution. Therefore, a more practical method is to use the Fast Gradient Sign Method (FGSM) or its variants to generate instantaneous perturbations. These methods are based on the gradient information of the current network to quickly construct adversarial samples without the need for a complete optimization process.

[0099] Taking FGSM as an example: Assume that the loss function of the network for the input x and label y is L(x, y), and the parameters of the current network are θ. For a certain normal learning example xnormal and its false prior label y fake,normal = 1, FGSM can be used to generate adversarial perturbation η:

[0100]

[0101] where ∈ is a hyperparameter that controls the perturbation magnitude, is the gradient of the loss function with respect to the input x normal . Adding the generated perturbation η to the original sample, the instantaneous perturbation learning example can be obtained:

[0102] Similarly, instantaneous perturbation information can also be generated for abnormal learning examples.

[0103] In step S104, the computer system incorporates the instantaneous perturbation information generated in step S103 into the normal learning examples and abnormal learning examples respectively to generate normal instantaneous perturbation learning examples and abnormal instantaneous perturbation learning examples. These perturbed samples will be used in subsequent training steps to enhance the network's sensitivity to minor changes and improve its generalization ability.

[0104] Taking the pressure sensor data of an oil-free dry vacuum unit as an example, assume that the feature vector of a certain normal learning example is represented as [0.8 MPa, 25 °C, 50 L / min, 0.05 mm / s], where the first element represents the pressure value. If the component of the instantaneous perturbation information generated by FGSM in the pressure dimension is +0.05 MPa, the new sample obtained after incorporating this perturbation into the original sample is [0.85 MPa, 25 °C, 50 L / min, 0.05 mm / s].

[0105] Similarly, the same operation can be performed for abnormal learning examples. Through this method, the computer system can generate a large number of instantaneous perturbation learning examples for subsequent network training processes.

[0106] Through the above steps S101 to S104, the computer system not only utilizes label smoothing technology to enhance the network's robustness to label noise, but also combines an adversarial training strategy to introduce instantaneous perturbation information, thereby generating richer and more diverse training samples. These samples will help improve the performance of the unit abnormal operation diagnosis network in complex operating environments, enabling it to more accurately identify the abnormal states of the unit. In the practical application of an oil-free dry vacuum unit, this preprocessing data technology will greatly improve the reliability and accuracy of the diagnosis system.

[0107] As an implementation manner, in step S103, according to the second error value of the third error determination function, generating instantaneous perturbation information may specifically include:

[0108] Step S1031: Determine the rate of change of the third error determination function with respect to the network parameters based on the second error value of the third error determination function, and obtain the optimal descent direction and its rate of the unit abnormal operation diagnosis network;

[0109] Step S1032: Determine the optimal descent direction and its rate of the unit abnormal operation diagnosis network as the instantaneous perturbation information.

[0110] In the embodiments of the present application, in order to further enhance the robustness and generalization ability of the unit abnormal operation diagnosis network, the computer system adopts an innovative strategy to generate instantaneous perturbation information. The core of this strategy is to use the second error value of the third error determination function (i.e., the error value under the false prior label) to guide the generation of instantaneous perturbations, which is specifically implemented by calculating the gradient information of the network parameters.

[0111] In step S1031, the computer system uses the second error value calculated by the third error determination function (still assuming the cross-entropy loss function as an example here) to guide the subsequent operations. This error value reflects the performance of the network under the false prior label, that is, the prediction deviation of the network when receiving the wrong label.

[0112] For example, assume there is a set of normal learning examples and abnormal learning examples, and their true prior labels are 0 (normal) and 1 (abnormal) respectively. In step S101, these labels are artificially swapped to generate false prior labels. Subsequently, in step S102, the network system makes predictions based on these false labels and calculates the corresponding error values. Now, the computer system uses this error value to further analyze the behavior of the network. Specifically, it needs to calculate the gradient of the error value with respect to the network parameters (such as weights and biases), that is, how the error value changes with the tiny changes of the parameters. This gradient information will reveal which parameters contribute the most to the error value and how these parameters should be adjusted to reduce the error.

[0113] In a neural network, the gradient is usually calculated by the backpropagation algorithm. Backpropagation is an effective algorithm for calculating the gradient of the loss function with respect to the network parameters. In this process, the error signal propagates backward from the output layer to the input layer, and each layer calculates the gradient of this layer based on the error signal of the previous layer and its own parameters.

[0114] Specifically, for the cross-entropy loss function, its gradient with respect to the network output can be expressed as:

[0115]

[0116] where \(L\) is the loss function, \(y\) is the true label (in this scenario, it is the false prior label), is the predicted output of the network. However, this gradient is with respect to the network output and needs to be further propagated through each layer of the network to compute the gradient with respect to the network's parameters.

[0117] During the propagation process, the gradient of each layer is transformed according to the activation function and weights of that layer. Finally, the gradient of the loss function with respect to each parameter of the network can be obtained, that is, and where \(w_i\) and \(b_j\) represent the weight and bias parameters in the network respectively.

[0118] Once the gradient information with respect to each parameter is available, the optimal descent direction of the loss function in the network parameter space can be determined. This direction points to the direction in which the loss function value decreases most rapidly, that is, the negative direction of the gradient. Specifically, if one wants to reduce the error value by adjusting the parameters, the parameters should be moved along the negative direction of the gradient. Therefore, in step S1031, the computer system is actually computing the gradient of the loss function with respect to the network parameters and determining the direction and rate (i.e., the magnitude of the gradient) along which the error value can be most rapidly reduced by descending along this gradient.

[0119] After determining the optimal descent direction and its rate, the computer system does not directly adjust the network parameters to reduce the error, but uses this gradient information as a basis for generating instantaneous perturbation information.

[0120] The innovation of this step S1032 lies in that it breaks the conventional pattern in the traditional training process. In traditional supervised learning, the network parameters are usually updated according to the gradient information to optimize the model performance. But here, the computer system regards the gradient information as a source of perturbation for generating instantaneous perturbation samples that can challenge the current network decisions.

[0121] Specifically, the computer system can construct an instantaneous perturbation vector based on the magnitude and direction of the gradient information. This vector has the same dimension as the input sample, and each of its elements is adjusted according to the gradient information of the corresponding dimension. The adjustment method can be a simple linear transformation (for example, multiplying the gradient value by a small constant factor as the perturbation amount), or a more complex non - linear transformation (depending on the specific application scenario and design choices). However, in actual operation, it should be noted that directly using the gradient information as a perturbation may not always be feasible or effective. Because the gradient information is for the network parameters, rather than directly for the input sample. Therefore, a more reasonable approach is to indirectly map the gradient information onto the input sample space to generate meaningful perturbations. A possible method is to use the generator part in Generative Adversarial Networks (GANs) to generate perturbation samples based on the gradient information. The generator can learn how to generate perturbation samples that can maximize the error value according to the input random noise and gradient information. These samples are then used as instantaneous perturbation learning examples to train the abnormal operation diagnosis network of the unit.

[0122] To simplify the explanation and maintain coherence with the previous text, assume a more direct method: scale the gradient information proportionally to the corresponding dimension of the input sample to generate an instantaneous perturbation. Although this method is simple, it can still effectively enhance the robustness of the network in some cases.

[0123] Taking the pressure sensor data of an oil - free dry vacuum unit as an example. Suppose the pressure characteristic value of a normal learning example is 0.8 MPa, and the gradient value calculated on this characteristic dimension is - 0.1 MPa / unit error (here, "unit error" is a hypothetical unit used to quantify the impact of the error value on the characteristic value). A small constant factor (such as 0.01) can be selected to scale the gradient value into a reasonable perturbation range, resulting in an instantaneous perturbation amount of - 0.001 MPa. Adding this perturbation amount to the original sample, the new sample with instantaneous perturbation can be obtained: 0.8 - 0.001 = 0.799 MPa.

[0124] Similarly, corresponding instantaneous perturbations can be generated for each characteristic dimension of the input sample, and these perturbations can be fused into the original sample to generate a complete instantaneous perturbation learning example.

[0125] Through the implementation of the above step S103, the computer system successfully uses the second error value of the third error determination function to calculate the gradient information of the network parameters, and uses this gradient information as the basis for generating instantaneous perturbation information. This method not only breaks the conventional pattern in the traditional training process, but also provides a new idea for enhancing the network robustness. In the embodiments of the present application, this instantaneous perturbation generation strategy based on gradient information will help improve the performance of the diagnostic network, enabling it to better handle various complex situations in actual operation.

[0126] As an implementation manner, in step S104, the instantaneous perturbation information is respectively incorporated into the normal learning examples and the abnormal learning examples to obtain the normal instantaneous perturbation learning examples and the abnormal instantaneous perturbation learning examples, which may specifically include:

[0127] Step S1041: Determine the importance coefficient of the instantaneous perturbation information, and perform a weighting operation on the instantaneous perturbation information according to the importance coefficient to obtain weighted instantaneous perturbation information;

[0128] Step S1042: Incorporate the weighted instantaneous perturbation information into the normal learning examples and the abnormal learning examples respectively to obtain the normal instantaneous perturbation learning examples and the abnormal instantaneous perturbation learning examples.

[0129] In step S1041, the computer system evaluates the importance of the instantaneous perturbation information and assigns an importance coefficient to each perturbation information accordingly. This coefficient reflects the influence degree of the perturbation information on the training process and is the key to balancing the perturbation intensity and the network stability.

[0130] For example, assume that during the operation of an oil-free dry vacuum unit, the readings of the pressure sensor are affected by a short-term voltage fluctuation, generating instantaneous perturbations. Although these perturbations have a short duration, they may have a significant impact on the performance of the diagnostic network. To simulate this actual situation, the computer system generates a series of instantaneous perturbation information and applies them to the training data. However, not all instantaneous perturbations have the same importance. For example, some perturbations may only slightly deviate from the normal value, while others may completely exceed the normal range. Therefore, before incorporating the perturbation information into the learning examples, the computer system determines its importance coefficient according to factors such as the intensity and frequency of the perturbation.

[0131] The determination of the importance coefficient can be based on various factors, including but not limited to the amplitude, frequency, duration of the perturbation, and the operating state of the unit when the perturbation occurs, etc. In practical applications, the computer system can adopt a heuristic method or a rule-based algorithm to automatically calculate these coefficients.

[0132] For example, a threshold can be set to distinguish between "important perturbations" and "minor perturbations". For perturbations with amplitudes exceeding this threshold, a higher importance coefficient is assigned; while for perturbations with lower amplitudes, a lower importance coefficient is assigned. In addition, the frequency and duration of the perturbations can also be considered, and higher importance is given to perturbations that occur frequently or have a longer duration.

[0133] Suppose the value of a certain instantaneous perturbation information in the pressure dimension is Δp, and its corresponding importance coefficient is α. The weighting operation is to multiply the perturbation value by its importance coefficient to obtain the weighted instantaneous perturbation information:

[0134] Δpweighted = α·Δp;

[0135] In this way, through the weighting operation, the computer system can adjust the intensity according to the importance of the perturbation information, ensuring that it can have an appropriate impact when integrating into the learning examples.

[0136] After determining the weighted instantaneous perturbation information, the computer system then effectively integrates this perturbation information into the normal learning examples and abnormal learning examples. This process needs to be carefully handled to ensure that the generated instantaneous perturbation learning examples can challenge the decision boundary of the network without completely destroying the authenticity and representativeness of the data.

[0137] The integration process usually involves directly adding the weighted instantaneous perturbation information to the corresponding features of the original learning examples. Taking the sensing data of an oil-free dry vacuum unit as an example, each learning example is a multi-dimensional feature vector, containing the readings of multiple sensors at different time points.

[0138] Suppose the feature vector of a certain normal learning example at time t is x t = [p t , T t , q t , …], where p t represents the pressure value at time t, T t represents the temperature value, q t represents other sensor readings, etc. Now, the computer system has generated the weighted instantaneous perturbation information Δpweighted for the pressure dimension.

[0139] The process of integrating the perturbation information into the learning example is to add the perturbation value to the original feature value:

[0140]

[0141] Then, replace the corresponding value in the original feature vector with the updated pressure value to obtain the new instantaneous perturbation learning example:

[0142]

[0143] Similarly, weighted instantaneous perturbation information can also be generated for abnormal learning examples and incorporated into the corresponding feature vectors.

[0144] Through the implementation manner of the above step S104, the computer system can effectively incorporate the instantaneous perturbation information into normal learning examples and abnormal learning examples to generate more challenging training data. This process not only enhances the robustness and generalization ability of the abnormal operation diagnosis network of the unit, but also provides rich and diverse training samples for subsequent optimization steps. In the embodiments of the present application, this method of incorporating weighted instantaneous perturbation information will help improve the accuracy and reliability of the diagnosis system.

[0145] As an implementation manner, the abnormal operation data detection is performed based on the classification component in the abnormal operation diagnosis network of the unit, and the abnormal operation diagnosis network of the unit further includes a synthesis component. Based on this, before performing the abnormal operation data detection on the normal continuous perturbation learning example and the abnormal continuous perturbation learning example respectively based on the abnormal operation diagnosis network of the unit in the step S300, the method may further include:

[0146] Step S301: Based on the synthesis component in the abnormal operation diagnosis network of the unit, generate continuous perturbation information according to the loaded random perturbation;

[0147] Step S302: Incorporate the continuous perturbation information into the normal learning example and the abnormal learning example respectively to obtain the normal continuous perturbation learning example and the abnormal continuous perturbation learning example correspondingly.

[0148] During the actual operation of an oil-free dry vacuum unit, the sensing data of the unit can be affected by various continuous disturbances, such as equipment aging, environmental temperature changes, power supply voltage fluctuations, etc. These disturbances usually do not disappear in a short period of time but will continuously affect the operating state of the unit. To simulate this continuous disturbance effect, the computer system uses a synthetic component (Generator) in the abnormal operation diagnosis network of the unit to generate corresponding continuous disturbance information. The synthetic component can be a complex neural network model specifically designed to generate high-quality noise data. This model can be based on the architecture of a generative adversarial network (GAN), where a generator network is responsible for generating disturbance information, and another discriminator network is responsible for evaluating the quality of the generated disturbances. By training these two networks to compete with each other, the synthetic component can learn how to generate noise data that is increasingly close to real disturbances. In practical applications, the synthetic component may adopt variational autoencoders (VAEs), generative adversarial networks (GANs), or their variants (such as WGAN, DCGAN, etc.) as its core algorithms. These algorithms optimize specific loss functions (such as adversarial loss, reconstruction loss, etc.) to make the generated disturbance information both diverse and conform to certain statistical laws.

[0149] In step S301, the computer system first loads a set of random disturbance vectors as inputs into the synthetic component. These random disturbance vectors can be vectors sampled from a standard normal distribution, and each element is a random number. Then, the synthetic component processes these random disturbance vectors through its internal network structure to generate a series of continuous disturbance information. This continuous disturbance information can be multi-dimensional, and each dimension corresponds to a feature in the unit's sensing data. For example, if the unit's sensing data includes four features: pressure, temperature, flow rate, and vibration, then the generated continuous disturbance information will also be a four-dimensional vector, such as [Δp, ΔT, Δq, Δv], where Δp, ΔT, Δq, and Δv represent the disturbance amounts in the pressure, temperature, flow rate, and vibration features respectively.

[0150] Suppose that at a certain point in time, the continuous disturbance information generated by the synthetic component is [0.01 MPa, 0.5 °C, 1 L / min, 0.005 mm / s]. This set of disturbance information will be used in subsequent steps to simulate the operating state of the unit under continuous disturbance conditions.

[0151] In step S302, the computer system incorporates the continuous disturbance information generated in step S301 into the normal learning examples and abnormal learning examples respectively. The incorporation process is usually achieved through simple addition operations, that is, adding each element of the disturbance information to the value of the corresponding feature.

[0152] Taking a normal learning example as an illustration, assume that the feature vector of a certain normal learning example at a certain time point is [0.8 MPa, 25 °C, 50 L / min, 0.05 mm / s]. According to the continuous perturbation information generated in step S301, the computer system adds this perturbation information to the feature vector to obtain a new feature vector:

[0153]

[0154] This new feature vector is the representation of the normal continuous perturbation learning example at this time point. Similarly, the same perturbation information can also be added to the abnormal learning example to obtain an abnormal continuous perturbation learning example.

[0155] After completing steps S301 and S302, the computer system obtains normal continuous perturbation learning examples and abnormal continuous perturbation learning examples. Next, in step S300, these examples will be input into the unit abnormal operation diagnosis network for abnormal operation data detection.

[0156] The detection process is mainly executed by the classification component in the unit abnormal operation diagnosis network. The classification component can be a neural network model of types such as a multi-layer perceptron (MLP), a convolutional neural network (CNN), or a recurrent neural network (RNN). This component processes the input feature vector layer by layer through its internal multi-layer network structure and finally outputs a scalar value representing the probability of abnormality.

[0157] Taking the multi-layer perceptron as an example, the classification component may include multiple fully connected layers and an output layer. Each fully connected layer contains a certain number of neurons and introduces non-linearity through activation functions (such as ReLU, Sigmoid, etc.). The input feature vector is first processed by the first fully connected layer and then sequentially passes through the subsequent fully connected layers until it reaches the output layer. The output layer usually contains one neuron, and its output value is processed by the Sigmoid activation function to obtain a probability value between 0 and 1, representing the probability that the input example belongs to the abnormal class.

[0158] The abnormality probability value output by the classification component will be used in the subsequent training and optimization processes. Specifically, the computer system will calculate the error value of the loss function (such as cross-entropy loss) based on these probability values and the prior labels (i.e., the true classes) of the examples. Then, the weights and bias parameters of the network are updated through optimization methods such as backpropagation algorithm and gradient descent to minimize the error value and improve the abnormal detection performance of the network.

[0159] Suppose the output of the classification component for a normal continuous perturbation learning example is 0.05 (indicating an abnormal probability of 5%), and the output for an abnormal continuous perturbation learning example is 0.95 (indicating an abnormal probability of 95%). If the prior labels of these two examples are 0 (normal) and 1 (abnormal) respectively, the performance of the classification component is satisfactory. However, if the output of the classification component for the normal example is too high (such as close to or exceeding 0.5), it indicates that the network may have made a misjudgment or needs further optimization.

[0160] Through the implementation analysis of the above steps S301, S302, and S300, the computer system uses the synthesis component to generate continuous perturbation information, integrates it into the learning examples to generate more challenging training data. Then, the classification component in the unit abnormal operation diagnosis network is used to detect the abnormal operation data of these examples, and the performance of the network is optimized according to the detection results. This process not only enhances the robustness of the network to continuous perturbations but also improves its abnormal detection ability in complex environments, providing a strong guarantee for the stable operation of the oil-free dry vacuum unit.

[0161] As an implementation method, in step S302, integrating the continuous perturbation information into the normal learning example and the abnormal learning example respectively to obtain the normal continuous perturbation learning example and the abnormal continuous perturbation learning example may specifically include:

[0162] Step S3021: Determine the importance coefficient of the continuous perturbation information, and perform a weighting operation on the continuous perturbation information according to the importance coefficient to obtain the importance coefficient continuous perturbation information;

[0163] Step S3022: Integrate the importance coefficient continuous perturbation information into the normal learning example and the abnormal learning example respectively to obtain the normal continuous perturbation learning example and the abnormal continuous perturbation learning example.

[0164] In step S302, the computer system not only integrates the continuous perturbation information into the normal learning example and the abnormal learning example but also performs a weighting operation according to the importance of the perturbation information to ensure that the generated continuous perturbation learning examples can more effectively train the unit abnormal operation diagnosis network.

[0165] For example, in the actual operation of the oil-free dry vacuum unit, the influence degrees of different types of continuous perturbations on the unit state are different. For example, a slight fluctuation in the power supply voltage may have a small impact on the unit performance, while a performance decline caused by equipment aging may have a significant impact on the unit state. Therefore, before integrating the continuous perturbation information into the learning examples, it is necessary to assign different weights according to the importance of the perturbation information.

[0166] The determination of the importance coefficient can be based on various factors, including but not limited to the type, amplitude, frequency, duration of the disturbance, and the actual impact of the disturbance on the unit performance, etc. In practical applications, the computer system may adopt one of the following methods to determine the importance coefficient:

[0167] Evaluation based on expert experience: Domain experts subjectively evaluate different types of continuous disturbances according to their own knowledge and experience, and assign corresponding importance coefficients to them. This method is simple and easy to implement, but there may be certain subjectivity.

[0168] Evaluation based on data analysis: The computer system analyzes historical operation data to evaluate the impact degree of different types of disturbances on the unit state, and determines the importance coefficient accordingly. This method is more objective, but requires a large amount of historical data as support.

[0169] Evaluation based on machine learning models: Use machine learning models (such as random forest, gradient boosting tree, etc.) to classify or regressively predict the disturbance data, and determine the importance coefficient according to the output results of the model. This method can automatically learn the complex relationships in the data, but the model training and parameter tuning processes may be relatively complex.

[0170] Regardless of which method is adopted, the finally obtained importance coefficient should be a value between 0 and 1, which is used to represent the importance degree of the disturbance information. The larger the value, the more important the disturbance information; the smaller the value, the relatively lower the importance of the disturbance information.

[0171] After determining the importance coefficient, the computer system performs a weighting operation on the continuous disturbance information. The weighting operation is usually achieved by multiplying the disturbance information by the corresponding importance coefficient. Suppose the value of a certain continuous disturbance information in the pressure dimension is Δp, and its corresponding importance coefficient is αp, then the weighted disturbance value is:

[0172] Δpweighted = αp·Δp;

[0173] Similarly, the same weighting operation can be performed on the disturbance information in other dimensions. The weighted disturbance information will be used to generate continuous disturbance learning examples in the subsequent steps.

[0174] Suppose the values of a certain continuous disturbance information in the four dimensions of pressure, temperature, flow rate, and vibration are [0.01MPa, 0.5℃, 1L / min, 0.005mm / s] respectively, and the corresponding importance coefficients of these four dimensions are [αp, αT, αq, αv] = [0.8, 0.6, 0.4, 0.9]. After the weighting operation, the obtained weighted continuous disturbance information is

[0175]

[0176] This set of weighted continuous perturbation information will be used to generate continuous perturbation learning examples in subsequent steps.

[0177] In step S3022, the computer system incorporates the weighted continuous perturbation information into the normal learning examples and the abnormal learning examples respectively. The incorporation process is usually achieved through simple addition operations, that is, each element of the weighted perturbation information is added to the value of the corresponding feature.

[0178] Taking the normal learning example as an example, assume that the feature vector of a certain normal learning example at a certain time point is [p, T, q, v] = [0.8 MPa, 25 °C, 50 L / min, 0.05 mm / s]. According to the weighted continuous perturbation information obtained in step S3021, the computer system adds this perturbation information to the feature vector to obtain a new feature vector:

[0179]

[0180] This new feature vector is the representation of the normal continuous perturbation learning example at this time point. Similarly, the same weighted perturbation information can also be added to the abnormal learning example to obtain an abnormal continuous perturbation learning example.

[0181] Through the implementation method analysis of the above step S302, it can be seen how the computer system performs weighted operations according to the importance coefficient of the continuous perturbation information and effectively incorporates it into the normal learning examples and the abnormal learning examples to generate more challenging continuous perturbation learning examples. This process not only enhances the diversity and complexity of the training data, but also helps to improve the robustness and generalization ability of the unit abnormal operation diagnosis network to continuous perturbations. In the embodiments of the present application, this method of incorporating weighted continuous perturbation information will help to improve the accuracy and reliability of the diagnosis system.

[0182] As an implementation method, in step S500, after optimizing the network parameters of the unit abnormal operation diagnosis network according to the error value of the first error determination function and the first error value of the second error determination function, the method may further include:

[0183] Step S800: Based on the optimized unit abnormal operation diagnosis network, detect the abnormal operation data of the normal continuous perturbation learning example and the abnormal continuous perturbation learning example respectively, and obtain a seventh detection result and an eighth detection result correspondingly;

[0184] Step S900: Determine the second error value of the second error determination function according to the prior label of the normal learning example, the prior label of the abnormal learning example, the seventh detection result, and the eighth detection result;

[0185] Step S1000: Determine the error value of the first error determination function and the second error value of the second error determination function, and optimize the network parameters of the optimized unit abnormal operation diagnosis network.

[0186] During the abnormal diagnosis process of the oil-free dry vacuum unit, in order to continuously improve the performance and accuracy of the diagnosis network, the computer system adopts a series of complex optimization strategies. After the initial optimization step (such as step S500), in order to further consolidate and enhance the robustness and generalization ability of the network, a derivative implementation method is proposed, which tests and optimizes the optimized network by reusing the learning examples processed by continuous perturbation.

[0187] For example, after initially optimizing the unit abnormal operation diagnosis network (step S500), the computer system does not immediately stop the optimization process, but chooses to use the learning examples processed by continuous perturbation (i.e., normal continuous perturbation learning examples and abnormal continuous perturbation learning examples) to further evaluate and optimize the network. This is because continuous perturbation can simulate various long-term and complex change situations that the unit may encounter during actual operation. By detecting such examples, the performance of the network can be evaluated more comprehensively.

[0188] In step S800, the computer system uses the optimized unit abnormal operation diagnosis network as a detection tool to detect abnormal operation data of normal continuous perturbation learning examples and abnormal continuous perturbation learning examples respectively. These learning examples are generated in previous steps (such as step S302), and they contain the sensing data of the unit under continuous perturbation conditions.

[0189] Specifically, for each continuous perturbation learning example, the network receives its feature vector as input and processes it through an internal neural network structure (such as convolutional layer, pooling layer, fully connected layer, etc.), and finally outputs a scalar value representing the probability of abnormality. This probability value is the detection result of the example, which is used to represent the confidence level that the network believes the example belongs to the abnormal category.

[0190] For example, assume there is a normal continuous perturbation learning example, and its feature vector is: [p perturbed ,T perturbed ,q perturbed ,v perturbed , where each element contains continuous perturbation information. After the network detects this example, it outputs a probability value close to 0 (such as 0.02), indicating that the network has a high confidence that this example belongs to the normal category. On the contrary, for an abnormal continuous perturbation learning example, the network may output a probability value close to 1 (such as 0.98), indicating that the network has a high confidence that this example belongs to the abnormal category.

[0191] Through this step, the computer system obtains the seventh detection result (for the normal continuous perturbation learning example) and the eighth detection result (for the abnormal continuous perturbation learning example), which will be used in the subsequent optimization process.

[0192] After obtaining the seventh and eighth detection results, the computer system calculates the second error value of the second error determination function. This error value reflects the performance of the optimized network under continuous perturbation conditions and is an important indicator for evaluating the network's robustness and generalization ability.

[0193] Similar to the previous steps, the cross-entropy loss function can still be used as the error determination function here. For each continuous perturbation learning example, the computer system compares its detection result (i.e., the abnormal probability value) with the prior label (i.e., the true class label) and calculates the error value using the cross-entropy loss function. The specific calculation formula is as follows:

[0194]

[0195] where N is the number of continuous perturbation learning examples (there are N normal continuous perturbation examples and N abnormal continuous perturbation examples each), yi and yj are the prior labels (0 and 1) of the normal continuous perturbation example and the abnormal continuous perturbation example respectively, and are the detection results (i.e., the abnormal probability values) of the network for them respectively.

[0196] By calculating the above formula, the computer system obtains the second error value of the second error determination function, which will be used to guide the subsequent optimization process.

[0197] After obtaining the error value of the first error determination function (based on instantaneous perturbation learning examples) and the second error value of the second error determination function (based on continuous perturbation learning examples), the computer system will start the optimization process again to further improve the performance of the unit abnormal operation diagnosis network. This optimization process is similar to the previous step S500, but the difference is that it uses more diverse error information to guide the optimization direction. Specifically, the computer system performs a weighted sum of these two error values (different importance coefficients can be given) to obtain a comprehensive error value. Then, optimization methods such as the backpropagation algorithm and gradient descent are used to update the weight and bias parameters of the network to minimize this comprehensive error value.

[0198] During the weighted summation process, the computer system assigns appropriate importance coefficients to the two error values according to the actual situation. These coefficients can be determined by methods such as expert experience, historical data analysis, or cross-validation. For example, if it is considered that the impact of continuous disturbances on the abnormal operation of the unit is more significant, the importance coefficient of the second error value can be set relatively high; conversely, if it is considered that instantaneous disturbances are equally important or more important, the coefficient value can be adjusted accordingly.

[0199] The optimization iteration process will be repeated within one or more training cycles until a certain stopping criterion is met (such as the error value drops below a preset threshold, the number of iterations reaches the upper limit, etc.). In each iteration, the computer system adjusts the network parameters based on the current error value and gradient information and gradually approaches the optimal solution.

[0200] To determine whether the optimization process converges, the computer system can monitor the change trend of the comprehensive error value. If the change amount of the error value is very small after several consecutive iterations (i.e., the gradient is close to zero), it can be considered that the network has converged to a local optimal solution or near the global optimal solution. At this time, the optimization process can be stopped and the current network parameters can be saved as the final model.

[0201] Suppose that after the first round of optimization (step S500), the error value of the first error determination function is 0.15, and the second error value of the second error determination function is 0.20. If the same importance coefficient (0.5 for each) is given to them, the comprehensive error value is (0.15 + 0.2) / 2 = 0.175. In the second round of optimization process (step S1000), the computer system tries to reduce this comprehensive error value by adjusting the network parameters. After multiple iterations, if the comprehensive error value drops below 0.10 and there is no significant change after several consecutive iterations, it can be considered that the network has converged and the optimization process is stopped.

[0202] Through each step (S800, S900, S1000) in the above-derived implementation manner, the computer system can test and optimize again after the initial optimization of the abnormal operation diagnosis network of the unit. This process not only uses more diverse error information to guide the optimization direction but also gradually approaches the optimal solution through multiple iterations. In the embodiments of the present application, this derived implementation manner helps to further improve the performance and accuracy of the diagnosis network, thereby ensuring the safe and reliable operation of the unit.

[0203] As an implementation manner, before the method respectively detects abnormal operation data for the normal instantaneous disturbance learning sample and the abnormal instantaneous disturbance learning sample based on the abnormal operation diagnosis network of the unit in step S100, the method may further include:

[0204] Step 100a: Obtain initial normal learning examples of several dimensions, and obtain initial abnormal learning examples of several dimensions corresponding to the initial normal learning examples;

[0205] Step 100b: Combine the initial normal learning examples of several dimensions to obtain the normal learning examples;

[0206] Step 100c: Combine the initial abnormal learning examples of several dimensions to obtain the abnormal learning examples.

[0207] In the process of abnormal diagnosis of an oil-free dry vacuum unit, in order to ensure that the abnormal operation diagnosis network of the unit can comprehensively and accurately identify the abnormal state of the unit, the computer system first prepares sufficient and diverse training data. These training data include not only the unit's sensing data under normal conditions but also data under simulated abnormal conditions. To further enrich the diversity of training data, the computer system adopts an implementation method. Before formally detecting abnormal operation data, it first obtains and combines initial learning examples of multiple dimensions.

[0208] In Step 100a, obtain initial normal learning examples of several dimensions, and obtain initial abnormal learning examples of several dimensions corresponding to the initial normal learning examples. For example, during the operation of an oil-free dry vacuum unit, various types of sensing data will be generated, such as pressure, temperature, flow rate, vibration, etc. These data reflect the real-time operation state of the unit and are important bases for abnormal diagnosis. However, due to the complexity of the unit structure and operation environment, different types of sensing data may have different sensitivities to abnormal states. Therefore, in order to more comprehensively capture the abnormal characteristics of the unit, the computer system obtains initial learning examples from multiple dimensions.

[0209] Specifically, the computer system can obtain initial learning examples in the following ways:

[0210] Real-time collect sensing data through various sensors installed on the unit. These sensors may include pressure sensors, temperature sensors, flow meters, vibration sensors, etc., which measure different parameters of the unit respectively.

[0211] In addition to real-time data, the computer system can also review and extract sensing data from the unit's historical operation records. These data may contain the operation state information of the unit under different working conditions and are of great significance for constructing a comprehensive training data set.

[0212] In some cases, in order to make up for the deficiency of actual data or enhance the diversity of data, the computer system can use physical models or simulation software to generate simulated sensing data. These simulated data can cover a wider range of working conditions and abnormal types.

[0213] When obtaining the initial learning examples, the computer system simultaneously obtains the data under normal conditions (initial normal learning examples) and the data under abnormal conditions (initial abnormal learning examples). For each type of sensing data (i.e., each dimension), it is necessary to obtain the initial learning examples under both normal and abnormal conditions separately.

[0214] Suppose the oil-free dry vacuum unit has three types of sensing data: pressure (P), temperature (T), and vibration (V). For each type of sensing data, the computer system has obtained a certain number of initial normal learning examples and initial abnormal learning examples. For example:

[0215] Pressure data (P):

[0216] Initial normal learning examples: P normal ={0.8MPa, 0.85MPa, 0.9MPa,...};

[0217] Initial abnormal learning examples: P abnormal ={1.2MPa, 1.3MPa, 1.1MPa,...};

[0218] Temperature data (T):

[0219] Initial normal learning examples: T normal ={25°C, 26°C, 24.5°C,...};

[0220] Initial abnormal learning examples: T abnormal ={30°C, 22°C, 29°C,...};

[0221] Vibration data (V):

[0222] Initial normal learning examples: V normal ={0.05mm / s, 0.04mm / s, 0.06mm / s,...};

[0223] Initial abnormal learning examples: V abnormal ={0.2mm / s, 0.15mm / s, 0.3mm / s,...};

[0224] These initial learning examples form the basis for data merging and processing in the subsequent steps.

[0225] After obtaining the initial normal learning examples of multiple dimensions, the computer system merges these examples into a unified data set, namely the normal learning examples. The merging process usually involves aligning each type of sensing data in chronological order or sample order and combining them into a multi-dimensional feature vector.

[0226] Specifically, for the sensing data at each time point (or each sample), the computer system extracts a value from each type of sensing data and combines these values into a feature vector. This process can be regarded as a "feature concatenation" process, which preserves the uniqueness of each type of sensing data while integrating them into a unified data structure.

[0227] For example, following the previous example, assume that at a certain time point t, the reading of the pressure sensor is 0.85 MPa, the reading of the temperature sensor is 26 °C, and the reading of the vibration sensor is 0.04 mm / s. The computer system combines these three values into a feature vector: [0.85 MPa, 26 °C, 0.04 mm / s].

[0228] Similarly, for the sensing data at other time points, the computer system will perform the same operation, and finally obtain a normal learning example dataset containing multiple feature vectors.

[0229] In step 100c, the initial abnormal learning examples of the several dimensions are combined to obtain the combination process of the abnormal learning examples (similar to step 100b). Similar to the process of combining the initial normal learning examples, the computer system also needs to combine the initial abnormal learning examples of multiple dimensions to obtain a unified abnormal learning example dataset. The combination process also involves the operation of feature concatenation, that is, combining the sensing data of each type into a multi-dimensional feature vector according to the same rule. Through the implementation of the above steps 100a, 100b, and 100c, the computer system successfully obtains and combines the initial learning examples of multiple dimensions, and obtains the normal learning examples and abnormal learning examples for the subsequent training process. This process not only enriches the diversity of the training data, but also provides comprehensive input features for the unit abnormal operation diagnosis network. In the subsequent steps, these learning examples will be used to generate instantaneous disturbance learning examples and continuous disturbance learning examples, and the abnormal operation data will be detected and optimized through the unit abnormal operation diagnosis network. This process will help improve the performance and accuracy of the diagnosis network, thereby ensuring the stable operation of the oil-free dry vacuum unit.

[0230] As an implementation manner, after optimizing the network parameters of the unit abnormal operation diagnosis network according to the error value of the first error determination function and the first error value of the second error determination function in step S500, the method may further include:

[0231] Step S1100: When the unit abnormal operation diagnosis network reaches the convergence state, stop optimizing the network parameters of the unit abnormal operation diagnosis network to obtain the target unit abnormal operation diagnosis network.

[0232] Application stage of the unit abnormal operation diagnosis network:

[0233] Step S10: Obtain the sensor monitoring data set of the unit to be diagnosed;

[0234] Step S20: Detect the abnormal operation data of the sensor monitoring data set of the unit to be diagnosed through the abnormal operation diagnosis network of the target unit, and obtain a detection result indicating whether the sensor monitoring data set of the unit to be diagnosed includes abnormal operation data.

[0235] In the abnormal diagnosis technology of oil-free dry vacuum units, the convergence of the optimization process and the performance of the diagnosis network in practical applications are important indicators for evaluating its performance.

[0236] During the process of training the abnormal operation diagnosis network of the unit, the computer system continuously adjusts the network parameters (such as weights and biases) to minimize the error value, thereby improving the prediction accuracy of the network. However, this optimization process cannot continue indefinitely because when the network performance reaches a certain level, further optimization may not bring significant performance improvement and may even lead to overfitting. Therefore, a reasonable stopping condition needs to be set to terminate the optimization process and obtain the final target diagnosis network.

[0237] Judging whether the abnormal operation diagnosis network of the unit has reached the convergence state usually involves the following considerations:

[0238] Change in error value: In consecutive multiple iterations, if the change in the error value (including the error value of the first error determination function and the first error value of the second error determination function) is very small and hardly decreases any further, then it can be considered that the network has approached the optimal solution and reached the convergence state. Specifically, a threshold value (such as 0.001 or a smaller number) can be set, and when the change in the error value is less than this threshold value, it is considered that the network has converged.

[0239] Gradient approaching zero: During the optimization process, calculating the gradient of the error value with respect to the network parameters is the key to adjusting the parameters. If the absolute value of the gradient is very small (approaching zero), then further adjusting the parameters will have a very limited effect on improving the error value. Therefore, the gradient approaching zero is also an important basis for judging network convergence.

[0240] Iteration number limit: In addition to the judgment based on the error value and gradient, an upper limit on the number of iterations can also be set to limit the progress of the optimization process. For example, set the maximum number of iterations to 1000 or more. When the number of iterations reaches 1000, regardless of whether the network has converged, stop the optimization process.

[0241] Suppose that during the training process, the computer system continuously adjusts the weights and biases of the abnormal operation diagnosis network of the unit through the backpropagation algorithm and the gradient descent method. After multiple iterations, the error value of the first error determination function gradually decreases from the initial 0.5 to 0.05, and in the recent 100 iterations, the change amount of the error value always remains within 0.0005. At the same time, the calculated gradient value is also very close to zero. At this time, the computer system determines that the network has converged and stops the optimization process to obtain the abnormal operation diagnosis network of the target unit.

[0242] After training the abnormal operation diagnosis network of the target unit, it can be applied to the abnormal diagnosis of the actual oil-free dry vacuum unit. First, it is necessary to obtain the sensing monitoring data set of the unit to be diagnosed. These data sets usually contain various sensing data generated during the operation of the unit, such as the values of parameters such as pressure, temperature, flow rate, and vibration. These data reflect the real-time operation state of the unit and are an important basis for abnormal diagnosis.

[0243] Suppose an oil-free dry vacuum unit is in operation, and the computer system collects sensing data in real time through sensors installed on the unit. After preprocessing (such as denoising, filtering, normalization, etc.), these data form the sensing monitoring data set to be diagnosed. This data set contains the operation records of the unit over a period of time and is the basis for subsequent abnormal diagnosis.

[0244] After obtaining the sensing monitoring data set to be diagnosed, the computer system inputs it into the abnormal operation diagnosis network of the target unit, processes and analyzes the input data through the neural network structure inside the network, and finally outputs the detection result indicating whether the operation state of the unit is abnormal. This detection result can be a probability value (indicating the probability of the unit having an abnormality) or a binary classification label (such as 0 for normal and 1 for abnormal).

[0245] The computer system inputs the sensing monitoring data set to be diagnosed into the abnormal operation diagnosis network of the target unit in batches or at one time. After receiving the input data, the network first extracts the feature representation of the data through structures such as the convolutional layer and the pooling layer; then, maps these features to the output space through the fully connected layer; finally, outputs the abnormal probability value or the binary classification label through the Softmax function or the Sigmoid function. For example, for the sensing data at a certain time point, the network outputs an abnormal probability value close to 1 (such as 0.98), indicating that the unit is very likely to have an abnormality at that time point.

[0246] After obtaining the detection result, the computer system determines whether the unit is abnormal according to a preset threshold or rule. If the detection result exceeds the preset threshold (for example, the abnormal probability value is greater than 0.5), it is considered that the unit is in an abnormal state; otherwise, it is considered that the unit is in a normal state. According to the detection result, the computer system can further trigger an alarm mechanism, record an abnormal log, send notification information, etc., so as to take measures in time to handle the abnormal problems of the unit.

[0247] An embodiment of the present application further provides a computer system, including a memory and a processor. The memory stores a computer program that can run on the processor, and when the processor executes the program, it implements some or all of the steps in the above method. Figure 2 The following is a schematic diagram of the hardware entity of a computer system provided by an embodiment of the present application, as Figure 2 shown. The hardware entity of the computer system 1000 includes: a processor 1001 and a memory 1002. Among them, the memory 1002 stores a computer program that can run on the processor 1001, and when the processor 1001 executes the program, it implements the steps in the method of any one of the above embodiments. The memory 1002 stores a computer program that can run on the processor. The memory 1002 is configured to store instructions and applications executable by the processor 1001, and can also cache data to be processed or already processed by the processor 1001 and each module in the computer system 1000 (for example, image data, audio data, voice communication data, and video communication data), and can be implemented by flash memory (FLASH) or random access memory (Random Access Memory, RAM). When the processor 1001 executes the program, it implements the steps of the abnormal diagnosis method of the oil-free dry vacuum unit in any one of the above. The processor 1001 generally controls the overall operation of the computer system 1000.

[0248] It should be understood that the "embodiments of the present application" or "as an implementation manner" mentioned throughout the specification mean that specific features, structures, or characteristics related to the embodiments are included in at least one embodiment of the present application. Therefore, the "in one embodiment" or "in an embodiment" that appears throughout the specification does not necessarily refer to the same embodiment. In addition, these specific features, structures, or characteristics can be combined in one or more embodiments in any suitable manner. It should be understood that in various embodiments of the present application, the magnitude of the serial numbers of the above steps / processes does not mean the sequence of execution. The execution sequence of each step / process should be determined according to its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present application. The serial numbers of the embodiments of the present application above are only for description and do not represent the superiority or inferiority of the embodiments. It should be noted that in this article, the term "including", "comprising", or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article, or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or further includes elements inherent to such a process, method, article, or device. Without further limitation, an element defined by the statement "including a..." does not exclude the existence of another identical element in the process, method, article, or device including the element. In several embodiments provided by the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are only illustrative. For example, the division of the units is only a logical function division, and there can be other division methods in actual implementation. For example, multiple units or components can be combined, or can be integrated into another system, or some features can be ignored, or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed with each other can be through some interfaces. The indirect coupling or communication connection of devices or units can be electrical, mechanical, or other forms.

[0249] The units described above as separate components may or may not be physically separated. The components shown as units may or may not be physical units; they can be located in one place or distributed to multiple network units; some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0250] In addition, each functional unit in the embodiments of the present application can be all integrated in a processing unit, or each unit can be separately used as a unit, or two or more units can be integrated in one unit; the above integrated units can be implemented in the form of hardware, or in the form of a combination of hardware and software functional units.

[0251] Those of ordinary skill in the art will understand that all or part of the steps of implementing the above method embodiments can be accomplished by hardware associated with program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps including those of the above method embodiments. The aforementioned storage medium includes various media that can store program codes, such as removable storage devices, read-only memory (ROM), magnetic disks, or optical discs.

[0252] As described above, the above are only the implementation manners of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application.

Claims

1. A method for diagnosing abnormality of an oil-free dry vacuum unit, characterized in that: The method comprises: Based on the abnormal operation diagnosis network of the unit, abnormal operation data detection is performed on the normal learning samples and the abnormal learning samples corresponding to the normal learning samples, and the fifth detection result and the sixth detection result are obtained accordingly; wherein each learning sample is a multidimensional feature vector including pressure, temperature, flow and vibration; Swapping the priori mark of the normal learning sample and the priori mark of the abnormal learning sample to obtain a false priori mark of the normal learning sample and a false priori mark of the abnormal learning sample; Determining a second error value of a third error determination function according to the false a priori mark of the normal learning sample, the false a priori mark of the abnormal learning sample, the fifth detection result, and the sixth detection result; generating instantaneous disturbance information according to a second error value of the third error determination function; Integrating the instantaneous disturbance information into the normal learning sample and the abnormal learning sample respectively, and obtaining a normal instantaneous disturbance learning sample and an abnormal instantaneous disturbance learning sample accordingly; Based on the abnormal operation diagnosis network of the unit, abnormal operation data detection is performed on normal instantaneous disturbance learning samples and abnormal instantaneous disturbance learning samples respectively, and a first detection result and a second detection result are obtained correspondingly; the abnormal operation data detection is performed based on a classification component in the abnormal operation diagnosis network of the unit, and the abnormal operation diagnosis network of the unit also includes a synthesis component; Based on the synthetic components in the abnormal operation diagnosis network of the unit, continuous disturbance information is generated according to the loaded random disturbance; Integrating the continuous disturbance information into the normal learning sample and the abnormal learning sample respectively, and obtaining a normal continuous disturbance learning sample and an abnormal continuous disturbance learning sample accordingly; Determining an error value of a first error determination function according to the priori mark of the normal learning sample, the priori mark of the abnormal learning sample, the first detection result, and the second detection result; Based on the abnormal operation diagnosis network of the unit, abnormal operation data detection is performed on the normal continuous disturbance learning samples and the abnormal continuous disturbance learning samples respectively, and a third detection result and a fourth detection result are obtained correspondingly; Determining a first error value of a second error determination function according to the priori mark of the normal continuous disturbance learning sample, the priori mark of the abnormal continuous disturbance learning sample, the third detection result, and the fourth detection result; The network parameters of the abnormal operation diagnosis network of the unit are optimized according to the error value of the first error determination function and the first error value of the second error determination function.

2. The method according to claim 1, characterized in that: The method further comprises: Determining a first error value of a third error determination function according to the priori mark of the normal learning sample, the priori mark of the abnormal learning sample, the fifth detection result, and the sixth detection result; The optimizing the network parameter variables of the abnormal operation diagnosis network of the unit according to the error value of the first error determination function and the first error value of the second error determination function comprises: The network parameters of the abnormal operation diagnosis network of the unit are optimized according to the error value of the first error determination function, the first error value of the second error determination function and the first error value of the third error determination function.

3. The method according to claim 2, characterized in that The optimizing the network parameter variables of the abnormal operation diagnosis network of the unit according to the error value of the first error determination function, the first error value of the second error determination function and the first error value of the third error determination function comprises: Respectively obtaining a significance coefficient of the first error determination function, a significance coefficient of the second error determination function, and a significance coefficient of the third error determination function; Obtaining an error value of a total error determination function according to the error value and the importance coefficient of the first error determination function, the first error value and the importance coefficient of the second error determination function, and the first error value and the importance coefficient of the third error determination function; Based on the error value of the total error determination function, the network parameters of the unit abnormal operation diagnosis network are optimized.

4. The method according to claim 3, characterized in that The step of generating instantaneous disturbance information according to the second error value of the third error determination function comprises: Based on the second error value of the third error determination function, determining the rate of change of the third error determination function with respect to the network parameter, and obtaining the optimal descent direction and rate of the abnormal operation diagnosis network of the unit; Determining the optimal descent direction and rate of the abnormal operation diagnosis network of the unit as instantaneous disturbance information; The step of integrating the instantaneous disturbance information into the normal learning sample and the abnormal learning sample to obtain the normal instantaneous disturbance learning sample and the abnormal instantaneous disturbance learning sample respectively includes: Determining the importance coefficient of the instantaneous disturbance information, and performing a weighted operation on the instantaneous disturbance information according to the importance coefficient to obtain weighted instantaneous disturbance information; The weighted instantaneous disturbance information is respectively integrated into the normal learning sample and the abnormal learning sample to obtain the normal instantaneous disturbance learning sample and the abnormal instantaneous disturbance learning sample.

5. The method according to claim 1, characterized in that The step of integrating the continuous disturbance information into the normal learning sample and the abnormal learning sample respectively, and correspondingly obtaining the normal continuous disturbance learning sample and the abnormal continuous disturbance learning sample, comprises: Determining the importance coefficient of the continuous disturbance information, and performing a weighted operation on the continuous disturbance information according to the importance coefficient to obtain importance coefficient continuous disturbance information; The importance coefficient continuous disturbance information is respectively integrated into the normal learning sample and the abnormal learning sample to obtain the normal continuous disturbance learning sample and the abnormal continuous disturbance learning sample.

6. The method according to claim 1, characterized in that After optimizing the network parameter of the abnormal operation diagnosis network of the unit according to the error value of the first error determination function and the first error value of the second error determination function, the method further includes: Based on the optimized abnormal operation diagnosis network of the unit, abnormal operation data detection is performed on the normal continuous disturbance learning sample and the abnormal continuous disturbance learning sample respectively, and a seventh detection result and an eighth detection result are obtained correspondingly; Determining a second error value of the second error determination function according to the priori mark of the normal learning sample, the priori mark of the abnormal learning sample, the seventh detection result, and the eighth detection result; The network parameters of the optimized unit abnormal operation diagnosis network are optimized according to the error value of the first error determination function and the second error value of the second error determination function.

7. The method according to claim 1, characterized in that Before performing abnormal operation data detection on normal transient disturbance learning samples and abnormal transient disturbance learning samples based on the abnormal operation diagnosis network of the unit, the method further includes: Acquire initial normal learning samples of several dimensions, and acquire initial abnormal learning samples of several dimensions corresponding to the initial normal learning samples; Merging the initial normal learning samples of the several dimensions to obtain the normal learning samples; The initial abnormal learning samples of the several dimensions are combined to obtain the abnormal learning samples.

8. The method according to claim 1, characterized in that After optimizing the network parameters of the abnormal operation diagnosis network of the unit according to the error value of the first error determination function and the first error value of the second error determination function, the method further includes: When the abnormal operation diagnosis network of the unit reaches a convergence state, the optimization of the network parameters of the abnormal operation diagnosis network of the unit is stopped to obtain a target abnormal operation diagnosis network of the unit; Obtaining sensor monitoring data sets for the unit to be diagnosed; The abnormal operation data detection is performed on the sensor monitoring data set of the unit to be diagnosed through the target unit abnormal operation diagnosis network to obtain a detection result indicating whether the sensor monitoring data set of the unit to be diagnosed includes abnormal operation data.

9. A computer system comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, wherein: When the processor executes the program, the steps in the method according to any one of claims 1 to 8 are implemented.

Images