Thermal pipeline network condition diagnosis method, device, terminal device and storage medium

Abstract

The present invention discloses a method, device, terminal device and storage medium for diagnosing the condition of a heat pipe network. The method includes: inputting random noise data with heat pipe network condition category labels into an ACGAN generator to obtain pseudo-samples, inputting the obtained real samples and pseudo-samples of heat pipe network images into an ACGAN discriminator for true / false discrimination training and classification training, and inputting them into an SVM model for classification training. Integrate the classification loss functions of the ACGAN discriminator and the SVM model into a unified loss function. According to the unified loss function, update the parameters of the ACGAN discriminator, gradually reduce the influence of the SVM loss function on the unified loss function. After the unified loss function is not affected by the SVM loss function, use the ACGAN for independent classification training. According to the trained ACGAN discriminator, obtain a heat pipe network condition diagnosis model, and input the collected heat pipe network images into the heat pipe network condition diagnosis model to obtain a condition diagnosis result. It can diagnose the condition of the heat pipe network more accurately and with higher efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of heat pipe network fault diagnosis, and in particular to a method, device, terminal device and storage medium for diagnosing the condition of a heat pipe network. Background Technique

[0002] Currently, the fault diagnosis technologies applied to heat pipe networks mainly include traditional machine learning methods and deep neural network methods. Traditional machine learning methods include algorithms such as support vector machines, random forests, and multi-layer perceptrons. Neural networks mainly include convolutional neural networks such as VGGNet, GoogLeNet, and ResNet. By training a learner with a balanced number of normal samples and fault samples, the learner can learn the characteristics of fault samples and normal samples. After obtaining on-site data, the learner can correctly classify the data to achieve fault diagnosis. The number of fault samples in the heat pipe network is much lower than the number of normal samples, making it difficult for traditional machine learning methods and deep neural network methods to obtain sufficient fault training samples and extract more accurate fault sample characteristics. Generative Adversarial Networks (GANs) provide a new idea for solving the problem of unbalanced data sets. GANs can generate a large number of pseudo-samples similar to real samples to expand the data set. ACGAN (Auxiliary Classifier Generative Adversarial Networks) is an improvement of GAN and can accurately classify the input data type while expanding the data set.

[0003] However, ACGAN still has difficulties in the early training of unbalanced samples. In the early stage of ACGAN training, the discriminator parameters in ACGAN need to be optimized, so the classification effect on faults is poor. The traditional support vector machine SVM (Support Vector Machine) has a significant classification effect on data sets with a small number of samples, but has a poor classification effect on data sets with a large number of samples and low efficiency. Summary of the Invention

[0004] In order to at least partially solve the technical problems existing in the prior art, the inventors made the present invention and provided a method, device, terminal device and storage medium for diagnosing the condition of a heat pipe network through specific embodiments.

[0005] In a first aspect, an embodiment of the present invention provides a method for training a heat pipe network condition diagnosis model, including the following steps:

[0006] Input the random noise data with the tags of the status categories of the heat pipe network into the ACGAN generator to obtain the pseudo-samples with the status category tags. Input the real samples of the heat pipe network images that have been obtained and the pseudo-samples with the status category tags newly generated by the ACGAN generator into the ACGAN discriminator for true / false discrimination training and classification training, and input them into the SVM model for classification training. Integrate the classification loss functions of the ACGAN discriminator and the SVM model into a unified loss function. Update the parameters of the ACGAN generator according to the true / false discrimination situation of the ACGAN, update the parameters of the ACGAN discriminator according to the unified loss function, and update the parameters of the SVM model according to the classification loss function of the SVM model;

[0007] After iterating the above steps multiple times, every certain number of iterations, reduce the influence of the SVM loss function on the unified loss function until the unified loss function is not affected by the SVM loss function;

[0008] After the unified loss function is not affected by the SVM loss function, use the ACGAN to independently conduct classification training, and obtain the heat pipe network status diagnosis model according to the trained ACGAN discriminator.

[0009] Optionally, before inputting the random noise data with the tags of the status categories of the heat pipe network into the ACGAN generator, the following steps are included:

[0010] Obtain the real samples of the heat pipe network images;

[0011] Set the status category tags according to the status categories of the real samples of the heat pipe network images;

[0012] Generate the random noise data with the status category tags;

[0013] Construct the SVM model according to the status category;

[0014] Construct the ACGAN model.

[0015] Optionally, the obtaining of the real samples of the heat pipe network images includes the following steps:

[0016] Collect several heat pipe network image samples under different status categories respectively, and the status categories include: normal, wall leakage, pipe bending deformation, and pipe insulation layer damage;

[0017] Perform at least one of the data augmentation processes on the collected heat pipe network image samples under different status categories to obtain the heat pipe network image samples after the data augmentation process. The data augmentation process includes image rotation, cropping, homogeneous splicing, affine transformation, adding verification noise, adding light and shadow, adjusting contrast, and adjusting saturation;

[0018] Randomly combine the heat pipe network image samples after data augmentation processing to obtain the heat pipe network image samples after random combination;

[0019] Mark the collected heat pipe network image samples under different condition categories, the heat pipe network image samples after the data augmentation processing, and the heat pipe network image samples after the random combination as the true heat pipe network image samples.

[0020] Optionally, constructing the SVM model according to the condition category includes the following steps:

[0021] Construct N - 1 SVM sub - models according to the number N of the condition categories, where N is a positive integer.

[0022] Optionally, constructing the ACGAN model includes the following steps:

[0023] Use a de - convolutional neural network to establish an ACGAN generator;

[0024] Use a convolutional neural network to establish an ACGAN discriminator.

[0025] Optionally, generating the random noise data with the condition category labels includes the following steps:

[0026] Collect multiple random noise data, add one of the condition category labels to each random noise data, and each of the condition category labels is added to the multiple random noise data.

[0027] Optionally, integrating the classification loss functions of the ACGAN discriminator and the SVM model into a unified loss function includes the following steps:

[0028] Determine the classification loss function L C ;

[0029] Determine the classification loss function L SVM ;

[0030] According to the classification loss functions L C and L SVM , determine that the expression of the unified loss function Loss is:

[0031] Loss=(1 - λ)L SVM +λL C

[0032] In the formula, λ is the confidence proportion of the classification result of the ACGAN discriminator, and an arbitrary number greater than 0 and less than 1 is set as the initial value of λ.

[0033] Optionally, every certain number of iterations, the influence of the SVM loss function on the unified loss function is reduced until the unified loss function is not affected by the SVM loss function, including the following steps:

[0034] Every certain number of iterations, λ in the Loss expression of the unified loss function is increased by a certain increment until λ increases to 1.

[0035] Optionally, after the unified loss function is not affected by the SVM loss function, ACGAN is used to independently perform classification training, and according to the trained ACGAN discriminator, a heat pipe network condition diagnosis model is obtained, including the following steps:

[0036] After λ increases to 1, the obtained true samples of heat pipe network images and the pseudo-samples with condition category labels newly generated by the ACGAN generator are input into the ACGAN discriminator for classification training, and according to the unified loss function, the parameters of the ACGAN discriminator are updated;

[0037] After the training termination condition is met, the update of the parameters of the ACGAN discriminator is stopped, and according to the trained ACGAN discriminator, a heat pipe network condition diagnosis model is obtained.

[0038] Optionally, obtaining the heat pipe network condition diagnosis model according to the trained ACGAN discriminator includes the following steps:

[0039] The trained ACGAN discriminator is stripped from the ACGAN model, the output for discriminating true and false of the trained ACGAN discriminator is cancelled, and only the classification output is retained to obtain the heat pipe network condition diagnosis model.

[0040] In a second aspect, an embodiment of the present invention provides a heat pipe network condition diagnosis method, including the following steps:

[0041] The collected heat pipe network images are input into the heat pipe network condition diagnosis model obtained by using the foregoing method, and the condition diagnosis result output by the heat pipe network condition diagnosis model is obtained.

[0042] In a third aspect, an embodiment of the present invention provides a heat pipe network condition diagnosis model training device, including:

[0043] An auxiliary classification module, configured to input random noise data with heat pipe network condition category labels into the ACGAN generator to obtain pseudo-samples with condition category labels, input the obtained true samples of heat pipe network images and the pseudo-samples with condition category labels newly generated by the ACGAN generator into the ACGAN discriminator for true / false discrimination training and classification training, and input them into the SVM model for classification training;

[0044] A loss function integration module, which is used to integrate the classification loss functions of the ACGAN discriminator and the SVM model into a unified loss function;

[0045] A parameter update module, which is used to update the parameters of the ACGAN generator according to the true or false discrimination of the ACGAN, update the parameters of the ACGAN discriminator according to the unified loss function, and update the parameters of the SVM model according to the classification loss function of the SVM model; after iterating the above steps multiple times, every certain number of iterations, the influence of the SVM loss function on the unified loss function is reduced until the unified loss function is not affected by the SVM loss function;

[0046] An independent classification module, which is used to perform independent classification training using the ACGAN after the unified loss function is not affected by the SVM loss function, and obtain a diagnosis model for the condition of the heat pipe network according to the trained ACGAN discriminator.

[0047] Optionally, it further includes:

[0048] A classification preparation module, which is used to obtain true samples of heat pipe network images; set condition category labels according to the condition categories of the true samples of heat pipe network images; generate random noise data containing the condition category labels; construct an SVM model according to the condition categories; construct an ACGAN model.

[0049] Optionally, the loss function integration module includes:

[0050] A discriminator loss function determination unit, which is used to determine the classification loss function L of the ACGAN discriminator C ;

[0051] An SVM loss function determination unit, which is used to determine the classification loss function L of the SVM model SVM ;

[0052] A loss function integration unit, which is used to determine the expression of the unified loss function Loss according to the classification loss functions L C and L SVM as:

[0053] Loss = (1 - λ)L SVM + λL C

[0054] In the formula, λ is the confidence ratio of the classification result of the ACGAN discriminator, and any number greater than 0 and less than 1 is set as the initial value of λ.

[0055] Optionally, the classification preparation module includes:

[0056] A true sample acquisition unit is configured to respectively collect a plurality of thermal pipeline network image samples under different condition categories, where the condition categories include: normal, pipe wall leakage, pipe bending deformation, and pipe insulation layer damage; perform at least one of data augmentation processing on the collected thermal pipeline network image samples under different condition categories to obtain thermal pipeline network image samples after data augmentation processing, where the data augmentation processing includes image rotation, cropping, homogeneous splicing, affine transformation, adding verification noise, adding light and shadow, adjusting contrast, and adjusting saturation; randomly combine the thermal pipeline network image samples after data augmentation processing to obtain randomly combined thermal pipeline network image samples; and form the true samples of the thermal pipeline network images with the collected thermal pipeline network image samples under different condition categories, the thermal pipeline network image samples after data augmentation processing, and the randomly combined thermal pipeline network image samples.

[0057] A condition category label setting unit is configured to set condition category labels according to the condition categories of the true samples of the thermal pipeline network images.

[0058] An SVM model construction unit is configured to construct N - 1 SVM sub-models according to the number N of the condition categories, where N is a positive integer.

[0059] An ACGAN model construction unit is configured to establish an ACGAN generator using a deconvolution neural network; and establish an ACGAN discriminator using a convolutional neural network.

[0060] A noise data acquisition unit is configured to collect a plurality of random noise data, add one of the condition category labels to each random noise data, and add each of the condition category labels to the plurality of random noise data.

[0061] In a fourth aspect, an embodiment of the present invention provides a thermal pipeline network condition diagnosis device, including:

[0062] An image acquisition module is configured to acquire thermal pipeline network images.

[0063] A condition diagnosis module is configured to input the acquired thermal pipeline network images into the thermal pipeline network condition diagnosis model obtained by using the foregoing method to obtain a condition diagnosis result output by the thermal pipeline network condition diagnosis model.

[0064] Based on the same inventive concept, an embodiment of the present invention further provides a terminal device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, where when the processor executes the computer program, the foregoing thermal pipeline network condition diagnosis model training method is implemented.

[0065] Based on the same inventive concept, an embodiment of the present invention further provides a computer storage medium, in which computer-executable instructions are stored, and when the computer-executable instructions are executed, the foregoing method for training a thermal pipeline network condition diagnosis model is implemented.

[0066] Based on the same inventive concept, an embodiment of the present invention further provides a terminal device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the computer program, the foregoing method for diagnosing the condition of a thermal pipeline network is implemented.

[0067] Based on the same inventive concept, an embodiment of the present invention further provides a computer storage medium, in which computer-executable instructions are stored, and when the computer-executable instructions are executed, the foregoing method for diagnosing the condition of a thermal pipeline network is implemented.

[0068] The beneficial effects of the above technical solutions provided by the embodiments of the present invention at least include:

[0069] An embodiment of the present invention provides a method for training a thermal pipeline network condition diagnosis model. Through an ACGAN generator, a large number of simulated samples with sample features identical to real fault samples are supplemented, solving the problem of imbalance in the number of fault samples and normal samples, which helps to improve the training effect of the ACGAN discriminator, more accurately extract the features of fault samples, and thus obtain a condition diagnosis model with more accurate condition diagnosis, improving the condition diagnosis accuracy rate; and combining ACGAN with SVM, using the characteristics of the SVM model to have accurate and efficient classification results when the sample size is small, and using the classification results of SVM to assist the classification training of ACGAN in the early stage of ACGAN training, making up for the deficiency of the classification effect of the ACGAN discriminator when the parameters need to be optimized in the early stage. As the number of samples increases, the parameters of the ACGAN discriminator are gradually optimized, and the classification effect becomes better and better. While SVM has a poor classification effect and low efficiency when the sample size is large, so the ACGAN discriminator starts independent training after updating the parameters multiple times, thus combining the advantages of the two, making the condition diagnosis model obtained by training more accurate and efficient in diagnosing the condition of the thermal pipeline network.

[0070] An embodiment of the present invention provides a method for diagnosing the condition of a thermal pipeline network. Since the thermal pipeline network condition diagnosis model used obtains a large number of pseudo-samples for training by simulating real samples of thermal pipeline network images, more accurate features of fault samples are extracted, and the advantages of accurate classification and high efficiency of the SVM model are combined in the early training, so the diagnosis of the condition of the thermal pipeline network by the thermal pipeline network condition diagnosis model will be more accurate and efficient.

[0071] Other features and advantages of the present invention will be set forth in the following description, and in part will be obvious from the description, or may be learned by practice of the present invention. The objectives and other advantages of the present invention may be realized and attained by the structure particularly pointed out in the written description, claims, as well as the drawings.

[0072] The technical solutions of the present invention will be further described in detail below with reference to the drawings and embodiments. Description of the Drawings

[0073] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, but do not constitute a limitation to the present invention. In the drawings:

[0074] Figure 1 It is a flowchart of a method for training a diagnosis model of a thermal pipeline network condition in an embodiment of the present invention;

[0075] Figure 2 It is a flowchart of a method for diagnosing a thermal pipeline network condition in an embodiment of the present invention;

[0076] Figure 3 It is a schematic diagram of the training and diagnosis process of a thermal pipeline network condition diagnosis model in an embodiment of the present invention;

[0077] Figure 4 It is a block diagram of a device for training a diagnosis model of a thermal pipeline network condition in an embodiment of the present invention;

[0078] Figure 5 It is a block diagram of a device for diagnosing a thermal pipeline network condition in an embodiment of the present invention;

[0079] Figure 6 It is a schematic diagram of the structure of a terminal device in an embodiment of the present invention. Detailed Embodiments

[0080] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the drawings. Although the exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be fully conveyed to those skilled in the art.

[0081] To solve the problems existing in the prior art, an embodiment of the present invention provides a method, device, terminal device and storage medium for diagnosing the condition of a thermal pipeline network.

[0082] Embodiment 1

[0083] Embodiment 1 of the present invention provides a method for training a diagnosis model of a thermal pipeline network condition, and its process is as Figure 1As shown in the figure, it includes the following steps:

[0084] Step S101: Input the random noise data containing the thermal pipeline network condition category label into the ACGAN generator to obtain the pseudo-samples containing the condition category label. Input the real samples of the thermal pipeline network images that have been obtained and the pseudo-samples containing the condition category label newly generated by the ACGAN generator into the ACGAN discriminator for true / false discrimination training and classification training, and input them into the SVM model for classification training. Integrate the classification loss functions of the ACGAN discriminator and the SVM model into a unified loss function. Update the parameters of the ACGAN generator according to the true / false discrimination situation of the ACGAN, update the parameters of the ACGAN discriminator according to the unified loss function, and update the parameters of the SVM model according to the SVM model classification loss function.

[0085] Optionally, before inputting the random noise data containing the thermal pipeline network condition category label into the ACGAN generator, it includes the following steps:

[0086] Obtain the real samples of the thermal pipeline network images;

[0087] Set the condition category label according to the condition category of the real samples of the thermal pipeline network images;

[0088] Generate the random noise data containing the condition category label;

[0089] Construct an SVM model according to the condition category;

[0090] Construct an ACGAN model.

[0091] Optionally, obtaining the real samples of the thermal pipeline network images includes the following steps:

[0092] Collect several thermal pipeline network image samples under different condition categories respectively. The condition categories include: normal, pipe wall leakage, pipe bending deformation, and pipe insulation layer damage;

[0093] Perform at least one of the data augmentation processes on the collected thermal pipeline network image samples under different condition categories to obtain the thermal pipeline network image samples after data augmentation. The data augmentation process includes image rotation, cropping, homogeneous splicing, affine transformation, adding verification noise, adding light and shadow, adjusting contrast, and adjusting saturation;

[0094] Randomly combine the thermal pipeline network image samples after data augmentation to obtain the randomly combined thermal pipeline network image samples;

[0095] Mark the collected thermal pipeline network image samples under different condition categories, the thermal pipeline network image samples after data augmentation, and the randomly combined thermal pipeline network image samples as the real samples of the thermal pipeline network images.

[0096] Optionally, according to the status category, an SVM model is constructed, including the following steps:

[0097] According to the number N of the status categories, N - 1 SVM sub-models are constructed, where N is a positive integer.

[0098] Optionally, an ACGAN model is constructed, including the following steps:

[0099] An ACGAN generator is established using a deconvolution neural network; an ACGAN discriminator is established using a convolutional neural network.

[0100] Optionally, random noise data containing the status category labels is generated, including the following steps:

[0101] Multiple random noise data are collected, and each random noise data is added with one of the status category labels, and each of the status category labels is added to the multiple random noise data.

[0102] Optionally, the classification loss functions of the ACGAN discriminator and the SVM model are integrated into a unified loss function, including the following steps:

[0103] Determine the classification loss function L of the ACGAN discriminator C ;

[0104] Determine the classification loss function L of the SVM model SVM ;

[0105] According to the classification loss functions L C and L SVM , determine that the expression of the unified loss function Loss is:

[0106] Loss = (1 - λ)L SVM + λL C

[0107] In the formula, λ is the confidence proportion of the classification result of the ACGAN discriminator, and an arbitrary number greater than 0 and less than 1 is set as the initial value of λ. The value of λ is between 0 and 1, including 0 and 1.

[0108] Step S102: After iterating step S101 multiple times, every certain number of iterations, the influence of the SVM loss function on the unified loss function is reduced until the unified loss function is not affected by the SVM loss function;

[0109] Optionally, every certain number of iterations, the influence of the SVM loss function on the unified loss function is reduced until the unified loss function is not affected by the SVM loss function, including the following steps:

[0110] Every certain number of iterations, increase λ in the Loss expression of the unified loss function by a certain increment until λ increases to 1.

[0111] Step S103: After the unified loss function is not affected by the SVM loss function, use ACGAN to independently perform classification training. According to the trained ACGAN discriminator, obtain the diagnosis model for the condition of the heat pipe network.

[0112] Optionally, after the unified loss function is not affected by the SVM loss function, use ACGAN to independently perform classification training. According to the trained ACGAN discriminator, obtain the diagnosis model for the condition of the heat pipe network, including the following steps:

[0113] After λ increases to 1, input the obtained real samples of heat pipe network images and the pseudo-samples with condition category labels newly generated by the ACGAN generator into the ACGAN discriminator for classification training. According to the unified loss function, update the parameters of the ACGAN discriminator;

[0114] After meeting the training termination condition, stop updating the parameters of the ACGAN discriminator. According to the trained ACGAN discriminator, obtain the diagnosis model for the condition of the heat pipe network.

[0115] Optionally, according to the trained ACGAN discriminator, obtain the diagnosis model for the condition of the heat pipe network, including the following steps:

[0116] Strip the trained ACGAN discriminator from the ACGAN model, cancel the output of the trained ACGAN discriminator for discriminating true and false, and only retain the classification output to obtain the diagnosis model for the condition of the heat pipe network.

[0117] For example, by mounting a high-definition camera and an infrared thermal imager on a robotic dog, the robotic dog walks in the heat pipe network tunnel for data collection, and a total of 20 normal samples and 20 samples of various faults such as pipe wall leakage, pipe bending deformation, and pipe insulation layer damage are collected.

[0118] Perform data augmentation on the small amount of collected samples. The data augmentation methods used include but are not limited to image rotation, cropping, splicing of samples of the same category, affine transformation, adding verification noise, adding light and shadow, adjusting contrast, and adjusting saturation. In addition, the above multiple augmentation methods can be randomly combined to better achieve data diversity. After the augmentation process, each type of data is increased to 200, and these 200 samples are used as the real samples of heat pipe network images.

[0119] Generate one-hot encoding according to the number of status category labels. One-hot encoding is a process of converting categorical variables into a form that is easy for machine learning algorithms to utilize. There are a total of 4 categories, namely normal samples and various fault samples, so the corresponding one-hot encodings are [1, 0, 0, 0], [0, 1, 0, 0], [0, 0, 1, 0], and [0, 0, 0, 1] respectively.

[0120] Randomly sample 110 times from a normal distribution with a mean of 1 and a standard deviation of 0 as the initial noise data, and cover the first 4 bits of the above one-hot encoding on the random noise as the status category label of the noise.

[0121] The ACGAN generator G takes random noise containing status category labels as input.

[0122] As a representative of traditional machine learning, the support vector machine often exhibits excellent characteristics in small-sample classification tasks. Therefore, in the early stage when the training samples are insufficient, an SVM classifier can be trained with small samples to assist the ACGAN training. The SVM classifies samples by mapping the data to a high-dimensional space and finding a maximum margin hyperplane in the high-dimensional space. In the present invention, a first-level classification SVM is first used to divide the samples into two major categories. The first major category includes normal samples and the first type of fault samples, and the second major category includes the second and third types of fault samples. Two second-level classification SVMs are respectively used to reclassify the two categories of samples after classification. Therefore, a total of three SVMs need to be trained.

[0123] The Generative Adversarial Networks (GAN) consists of a generator G and a discriminator D, and reaches the Nash equilibrium through the game training of the two. The generator takes random noise as input, and its training goal is to learn the data distribution law of the real sample X real so as to generate a pseudo-sample X fake that is similar to the real sample and can "pass off as real"; the discriminator takes the real sample and the pseudo-sample generated by the generator as input, and its training goal is to accurately judge whether the input is a real sample or a pseudo-sample. Ideally, the trained generator can generate pseudo-samples that are exactly the same as the real samples, while the discriminator cannot distinguish the authenticity of the samples, and the output discrimination probability is always 0.5. The objective function of GAN is:

[0124] L GAN = arg min G max D {E(lnD(X real )) + E(ln(1 - D(X fake )))}

[0125] where X fake= G(z), where z is random noise. Since GAN takes random noise as input without any constraints, the training process is relatively divergent, and the mechanism of adversarial training between the generator and the discriminator increases the risk of network collapse during training. In addition, GAN is effective in the field of data augmentation but is difficult to handle classification tasks.

[0126] ACGAN is an improved network based on GAN. ACGAN is also composed of a generator G and a discriminator D. The generator takes random noise z as input, but compared with GAN, it adds a class label c to guide the generator to generate pseudo-samples of different classes; the discriminator takes real samples of different classes and pseudo-samples generated by the generator as input, and its training objective is improved compared with GAN. The discriminator of ACGAN not only has to discriminate whether the input data is real but also has to judge the class to which the input data belongs.

[0127] The training objective of the generator is to generate pseudo-samples of different classes according to the sample labels. That is, when the generated data passes through the discriminator, it is expected that the output probability of the discriminator is as large as possible and as close to 1 as possible. Therefore, the objective function of the generator can be expressed as:

[0128] L G-S = max E(ln(D(G(z))))

[0129] The training objective of the discriminator is to accurately discriminate the authenticity of the data. That is, when the input is real data, it is expected that the discrimination probability is as large as possible, and when the input is pseudo-data generated by the generator, it is expected that the output probability is as small as possible. Therefore, the objective function of the discriminator can be expressed as:

[0130] L D-S = max E(ln(D(x)))+min E(ln(D(G(z))))

[0131] Another training objective of the discriminator is to accurately classify the samples. That is, when real data and pseudo-data pass through the discriminator, the discriminator has a high classification accuracy. The classification loss of the discriminator can be expressed as:

[0132] L C = max E(ln(D(x)))+max E(ln(D(G(z))))

[0133] To make the objective functions consistent, the above objective functions are integrated into the log-likelihood function L of the correct data input source S and the log-likelihood function L of the correct class C :

[0134] L S = E(ln(P(S = real|X real )))+E(ln(P(S = fake|Xfake )))

[0135] = E(ln(D S (x))) + E(ln(D S (G(z))))

[0136] L C = E(ln(P(C = c|X real ))) + E(ln(P(C = c|X fake )))

[0137] = E(ln(D C (x))) + E(ln(D C (G(z))))

[0138] Among them, S represents true or false judgment, and C represents classification.

[0139] The objective function of the generator G is to maximize L C -L S , and the objective function of the discriminator D is to maximize L C +L S .

[0140] ACGAN guides the generator and the discriminator to play a game through max-minimization, so that the generator can generate various pseudo-samples that are realistic enough, and the discriminator can accurately judge the data source and accurately classify the sample categories.

[0141] The present invention uses a transposed convolutional neural network as the generator of ACGAN. Transposed convolution, also known as deconvolution, is the inverse process of the general convolution operation that gradually increases the number of feature maps and decreases the specification. Through the deconvolution operation, the specification of the feature map becomes larger and the number becomes smaller. The generator uses a large number of deconvolution operations. When the input sample specification is small, it continuously integrates information under the guidance of the loss function, enabling the generator to generate detailed and realistic pseudo-samples. After the deconvolution operation, to prevent gradient explosion, the batch normalization operation BatchNormalization is added.

[0142] The present invention uses a convolutional neural network as the discriminator of the ACGAN. A convolutional neural network is a feedforward neural network, which is a supervised model trained end-to-end. The convolutional layer is mainly used for feature extraction. The convolutional kernel adopts the form of local connection and weight sharing, which greatly reduces the model parameters, and also reduces the network complexity and the risk of overfitting. After the feature extraction is completed, a fully connected layer is used to refit the image features, so as to reduce the loss of image feature information. The classifier receives its output value and completes the classification of the data according to actual needs. According to different classification tasks, the present invention parallelly adopts two fully connected layers with different parameters. One is used to integrate the data source information, and the output is the probability that the data is true. The other is used to integrate the data category information, and the output is the probability of the category to which the data belongs. After each convolution operation, in order to prevent gradient explosion, the batch normalization operation BatchNormalization is added. At the same time, in order to prevent the model from relying too much on individual neurons, a Dropout layer is added after each round of activation function in the model.

[0143] The training process of the ACGAN is separate and alternative iterative training, that is, the generator and the discriminator are alternately trained.

[0144] The discriminator D has two goals. One is to try its best to distinguish whether the input is real data or data generated by the generator. The other is to classify various types of data, that is, to distinguish whether it is normal data or a certain fault.

[0145] A generation network is obtained by randomly configuring the weights of each node of the generator. At this time, when a random signal and a class label are input, a fake sample data set will be obtained. Since the model is not optimized at this time and the generation network G is still at a disadvantage, the generated samples do not learn any rules of the real samples, and this sample set is easily recognized as fake data by the discriminant network. Next, the real samples and the generated pseudo-samples are used to complete the training of the discriminator D. The labels of the true and false sample sets are artificially defined. All class labels of the true sample set are set to 1, while all class labels of the pseudo-sample set are set to 0. And the class labels are set for all samples respectively. The true and false data sets are sent into the discriminant network D for training. The training goal of the discriminator is not only to distinguish between true and false samples, but also to distinguish the data categories.

[0146] The role of the generator G is to generate various samples as realistic as possible.

[0147] When training the generation network, it is necessary to jointly train with the discriminant network to achieve the training purpose. Therefore, the training of the generation network is actually the training of the concatenation of the generation-discriminant network. Change the labels of the fake samples to 1, that is, consider these fake samples as real samples during the training of the generation network, so as to achieve the purpose of confusing the discriminator and make the generated fake samples gradually approach the real samples. In addition, when training the generator, it is necessary to fix the parameters of the discriminator, that is, do not let its weights be updated. The discriminator D is only responsible for transmitting the error, so as to guide the generator to complete the parameter update.

[0148] According to the situation of the ACGAN to distinguish true and false, update the parameters of the ACGAN generator. After updating the generation parameters, generate new fake samples for the previous random noise z according to the current new generation network, and the generated fake samples are closer to the real samples at this time.

[0149] When training the ACGAN, perform multi-classification training of the SVM using the same data set. Since the SVM has outstanding classification effects on small sample data sets, and in the early stage of ACGAN training, the parameters of the discriminator D need to be optimized, so the classification effect on faults is poor. At this time, combine the classification results of the SVM with the classification results of the ACGAN to enhance the classification confidence, and use the combined classification results as a unified loss to feedback to the ACGAN to guide the parameter update. At this time, the ACGAN classification loss function is:

[0150] Loss=(1-λ)L SVM +λL C

[0151] Among them, λ is the confidence proportion of the classification result of the discriminator in the ACGAN, and an arbitrary number greater than 0 and less than 1 is set as the initial value of λ. The value of λ is between 0 and 1, including 0 and 1.

[0152] According to the unified loss function, update the parameters of the ACGAN discriminator, and update the parameters of the SVM model according to the SVM model classification loss function.

[0153] Input the random noise data containing the labels of the thermal pipeline network status categories into the ACGAN generator with updated parameters to obtain more realistic fake samples containing status category labels. Input the obtained real samples of the thermal pipeline network images and the more realistic fake samples containing status category labels newly generated by the ACGAN generator into the ACGAN discriminator with updated parameters for true and false discrimination training and classification training, and input them into the SVM model with updated parameters for classification training.

[0154] After iterating the above steps multiple times, the set composed of the pseudo-samples generated by the generator of the trained ACGAN and the real samples of the heat pipe network images will be expanded to 1 times the original. Retraining will be carried out on the basis of the previous training of SVM and ACGAN using the new dataset. As the parameters of ACGAN are continuously optimized, λ is increased once every certain number of iteration batches. For example, the value of λ is increased by 0.1 every certain number of iteration batches to improve the confidence proportion of the ACGAN classification result. At the same time, the newly output pseudo-samples of the generator are used for training. This process continues until λ is 1, and the coefficient of L SVM in the unified loss function is 0. At this time, only the classification loss function L C of the ACGAN discriminator is used to update the parameters of the ACGAN discriminator, and the SVM model exits the training, and the ACGAN starts to be trained alone.

[0155] The real samples of the heat pipe network images that have been obtained and the pseudo-samples with status category labels newly generated by the ACGAN generator are input into the ACGAN discriminator for classification training. The parameters of the ACGAN discriminator are updated according to the classification loss function L C of the ACGAN discriminator. The real samples of the heat pipe network images that have been obtained and the pseudo-samples with status category labels newly generated by the ACGAN generator are input into the ACGAN discriminator with updated parameters for classification training. The above-mentioned separate training process is continuously repeated. After a certain number of iterations, the update of the network parameters is stopped. At this time, the data generated by the generator G is already quite real, and the discriminator cannot distinguish the authenticity of the input data but can accurately classify the status category.

[0156] After the training is completed, the discriminator D in the ACGAN is separated, and at the same time, the output of D cancels the judgment on the authenticity of the data and only retains the classification output. After the image is collected in real time, it is input into the current discriminator D, that is, the heat pipe network status diagnosis model, for status diagnosis.

[0157] In the above method of this embodiment, through the ACGAN generator, a large number of simulated samples with the same sample features as the real fault samples are supplemented, solving the problem of the imbalance in the number of fault samples and normal samples, which helps to improve the training effect of the ACGAN discriminator, more accurately extract the features of the fault samples, and thus obtain a condition diagnosis model with more accurate condition diagnosis, improving the accuracy of condition diagnosis; and by combining ACGAN with SVM, taking advantage of the accurate and efficient classification effect of the SVM model when the sample size is small, the classification results of SVM are used to assist the classification training of ACGAN in the early stage of ACGAN training, making up for the deficiency of the classification effect of the ACGAN discriminator when the parameters need to be optimized in the early stage. As the sample size increases, the parameters of the ACGAN discriminator are gradually optimized, and the classification effect becomes better and better. While SVM has a poor classification effect and low efficiency when the sample size is large, so the ACGAN discriminator starts independent training after updating the parameters multiple times, thus combining the advantages of the two, making the condition diagnosis model obtained by training more accurate and efficient in diagnosing the condition of the heat pipe network.

[0158] Embodiment 2

[0159] Embodiment 2 of the present invention provides a method for diagnosing the condition of a heat pipe network, and its process is as Figure 2 , including the following steps:

[0160] Step S201: Collect heat pipe network images;

[0161] Step S202: Input the collected heat pipe network images into the heat pipe network condition diagnosis model obtained by using the foregoing method, and obtain the condition diagnosis result output by the heat pipe network condition diagnosis model.

[0162] The processes of training and diagnosis are as Figure 3 shown in the figure. In the figure, G represents the ACGAN generator, which generates pseudo data, that is, pseudo samples, and D represents the ACGAN discriminator. The ACGAN discriminator discriminates and classifies the pseudo data and real data for true or false. The role of true or false discrimination is to feedback the discrimination result to the generator to help the generator generate more real pseudo data. In the early stage of training, the SVM model assists the ACGAN discriminator in classification. In the later stage, the ACGAN discriminator conducts classification training alone. The trained ACGAN discriminator is set in the online fault diagnosis system to diagnose the real-time condition of the heat pipe network and output the diagnosis category result.

[0163] In the above method of this embodiment, since the heat pipe network condition diagnosis model used is to obtain a large number of pseudo samples through simulating the real samples of heat pipe network images for training, more accurate features of the fault samples are extracted, and the advantages of accurate classification and high efficiency of the SVM model are combined in the early training, so the diagnosis of the heat pipe network condition by the heat pipe network condition diagnosis model will be more accurate and efficient.

[0164] Embodiment III

[0165] Embodiment III of the present invention provides a training device for a thermal pipeline network condition diagnosis model, and its structure is as Figure 4 shown, including:

[0166] The auxiliary classification module 101 is used to input random noise data with thermal pipeline network condition category labels into the ACGAN generator to obtain pseudo-samples with condition category labels, input the obtained real samples of thermal pipeline network images and the pseudo-samples with condition category labels newly generated by the ACGAN generator into the ACGAN discriminator for true / false discrimination training and classification training, and input them into the SVM model for classification training;

[0167] The loss function integration module 102 is used to integrate the classification loss functions of the ACGAN discriminator and the SVM model into a unified loss function;

[0168] The parameter update module 103 is used to update the parameters of the ACGAN generator according to the true / false discrimination situation of the ACGAN, update the parameters of the ACGAN discriminator according to the unified loss function, and update the parameters of the SVM model according to the SVM model classification loss function; after iterating the above steps multiple times, every certain number of iterations, reduce the influence of the SVM loss function on the unified loss function until the unified loss function is not affected by the SVM loss function;

[0169] The independent classification module 104 is used to perform independent classification training using the ACGAN after the unified loss function is not affected by the SVM loss function, and obtain a thermal pipeline network condition diagnosis model according to the trained ACGAN discriminator.

[0170] Optionally, it further includes:

[0171] The classification preparation module 100 is used to obtain real samples of thermal pipeline network images; set condition category labels according to the condition categories of the real samples of thermal pipeline network images; generate random noise data with the condition category labels; construct an SVM model according to the condition categories; construct an ACGAN model.

[0172] Optionally, the loss function integration module includes:

[0173] The discriminator loss function determination unit is used to determine the classification loss function L of the ACGAN discriminator C ;

[0174] The SVM loss function determination unit is used to determine the classification loss function L of the SVM model SVM ;

[0175] A loss function integration unit, which is used to determine the expression of the unified loss function Loss according to the classification loss functions L C and L SVM as:

[0176] Loss = (1 - λ)L SVM + λL C

[0177] where λ is the confidence proportion of the classification result of the ACGAN discriminator, and any number greater than 0 and less than 1 is set as the initial value of λ.

[0178] Optionally, the classification preparation module includes:

[0179] A true sample acquisition unit, which is used to collect a number of thermal pipeline network image samples under different condition categories respectively. The condition categories include: normal, pipe wall leakage, pipe bending deformation, and pipe insulation layer damage; perform at least one of the data enhancement processes on the collected thermal pipeline network image samples under different condition categories to obtain the thermal pipeline network image samples after data enhancement. The data enhancement process includes image rotation, cropping, homogeneous splicing, affine transformation, adding verification noise, adding light and shadow, adjusting contrast, and adjusting saturation; randomly combine the thermal pipeline network image samples after data enhancement to obtain the randomly combined thermal pipeline network image samples; combine the collected thermal pipeline network image samples under different condition categories, the thermal pipeline network image samples after data enhancement, and the randomly combined thermal pipeline network image samples to form the true samples of the thermal pipeline network images;

[0180] A condition category label setting unit, which is used to set condition category labels according to the condition categories of the true samples of the thermal pipeline network images;

[0181] An SVM model construction unit, which is used to construct N - 1 SVM sub - models according to the number of condition categories N, where N is a positive integer;

[0182] An ACGAN model construction unit, which is used to establish an ACGAN generator using a de - convolutional neural network; establish an ACGAN discriminator using a convolutional neural network;

[0183] A noise data acquisition unit, which is used to collect a plurality of random noise data, add a condition category label to each random noise data, and add each condition category label to the plurality of random noise data.

[0184] In the above device of this embodiment, through the ACGAN generator, a large number of simulated samples with the same sample features as the real fault samples are supplemented, solving the problem of the imbalance in the number of fault samples and normal samples, which helps to improve the training effect of the ACGAN discriminator, extract the features of the fault samples more accurately, and thus obtain a condition diagnosis model with more accurate condition diagnosis, improving the condition diagnosis accuracy. Moreover, by combining ACGAN with SVM and taking advantage of the accurate and efficient classification effect of the SVM model when the sample size is small, the classification result of SVM is used to assist the classification training of ACGAN in the early stage of ACGAN training, making up for the deficiency of the classification effect of the ACGAN discriminator when the parameters need to be optimized in the early stage. As the sample size increases, the parameters of the ACGAN discriminator are gradually optimized, and the classification effect becomes better and better. While when the sample size is large, the classification effect of SVM is poor and the efficiency is very low. Therefore, the ACGAN discriminator starts independent training after updating the parameters multiple times, thus combining the advantages of the two, making the condition diagnosis model obtained by training more accurate and efficient in diagnosing the condition of the heat pipe network.

[0185] Embodiment 4

[0186] Embodiment 4 of the present invention provides a heat pipe network condition diagnosis device, the structure of which is as Figure 5 shown, including:

[0187] An image acquisition module 201 for acquiring heat pipe network images;

[0188] A condition diagnosis module 202 for inputting the acquired heat pipe network images into the heat pipe network condition diagnosis model obtained by using the foregoing method to obtain the condition diagnosis result output by the heat pipe network condition diagnosis model.

[0189] In the above device of this embodiment, since the heat pipe network condition diagnosis model used is to obtain a large number of pseudo samples for training by simulating the real samples of heat pipe network images, more accurate features of the fault samples are extracted, and the advantages of accurate classification and high efficiency of the SVM model are combined in the early training, so the diagnosis of the heat pipe network condition by the heat pipe network condition diagnosis model will be more accurate and efficient.

[0190] Based on the same inventive concept, the embodiment of the present invention also provides a terminal device, the structure of which is as Figure 5 shown, including: a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the computer program, the foregoing heat pipe network condition diagnosis model training method is implemented.

[0191] Based on the same inventive concept, an embodiment of the present invention further provides a computer storage medium, in which computer-executable instructions are stored, and when the computer-executable instructions are executed, the foregoing method for training a thermal pipeline network condition diagnosis model is implemented.

[0192] Based on the same inventive concept, an embodiment of the present invention further provides a terminal device, the structure of which is as Figure 5 shown, and includes: a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the foregoing method for diagnosing the condition of a thermal pipeline network is implemented.

[0193] Based on the same inventive concept, an embodiment of the present invention further provides a computer storage medium, in which computer-executable instructions are stored, and when the computer-executable instructions are executed, the foregoing method for diagnosing the condition of a thermal pipeline network is implemented.

[0194] Regarding the device in the above embodiments, the specific manners in which each module performs operations have been described in detail in the embodiments of the related method, and will not be elaborated herein.

Claims

1. A method for training a diagnosis model of the condition of a heating pipeline network, characterized in that, Including the following steps: Input the random noise data with the thermal pipeline network condition category label into the ACGAN generator to obtain the pseudo samples with the condition category label. Input the obtained real samples of the thermal pipeline network images and the newly generated pseudo samples with the condition category label by the ACGAN generator into the ACGAN discriminator for true / false discrimination training and classification training, and input them into the SVM model for classification training. Integrate the classification loss functions of the ACGAN discriminator and the SVM model into a unified loss function. According to the true / false discrimination situation of the ACGAN, update the parameters of the ACGAN generator. According to the unified loss function, update the parameters of the ACGAN discriminator. According to the SVM model classification loss function, update the parameters of the SVM model; After iterating the above steps multiple times, every certain number of iterations, reduce the influence of the SVM loss function on the unified loss function until the unified loss function is not affected by the SVM loss function; After the unified loss function is not affected by the SVM loss function, use the ACGAN to independently conduct classification training, and obtain the thermal pipeline network condition diagnosis model according to the trained ACGAN discriminator; The integration of the classification loss functions of the ACGAN discriminator and the SVM model into a unified loss function includes the following steps: Determine the classification loss function L of the ACGAN discriminator C ; Determine the classification loss function L of the SVM model SVM ; According to the classification loss functions L C and L SVM , the expression of the unified loss function Loss is determined as follows: Loss=(1-λ)L SVM +λL C In the formula, λ is the confidence proportion of the classification result of the ACGAN discriminator, and any number greater than 0 and less than 1 is set as the initial value of λ.

2. The method according to claim 1, characterized in that, Before inputting the random noise data with the thermal pipeline network condition category label into the ACGAN generator, it includes the following steps: Obtain the real samples of the thermal pipeline network images; Set the condition category label according to the condition category of the real samples of the thermal pipeline network images; Generate the random noise data with the condition category label; Construct the SVM model according to the condition category; Construct the ACGAN model.

3. The method according to claim 2, wherein The obtaining of the real samples of the thermal pipeline network images includes the following steps: Collect several thermal pipeline network image samples under different condition categories respectively, and the condition categories include: normal, pipe wall leakage, pipe bending deformation, and pipe insulation layer damage; Perform at least one of the data augmentation processes on the collected thermal pipeline network image samples under different condition categories to obtain the thermal pipeline network image samples after data augmentation processing. The data augmentation processing includes image rotation, cropping, homogeneous splicing, affine transformation, adding verification noise, adding light and shadow, adjusting contrast, and adjusting saturation; Randomly combine the thermal pipeline network image samples after data augmentation processing to obtain the randomly combined thermal pipeline network image samples; Mark the collected thermal pipeline network image samples under different condition categories, the thermal pipeline network image samples after data augmentation processing, and the randomly combined thermal pipeline network image samples as the real samples of the thermal pipeline network images.

4. The method according to claim 2, wherein The construction of the SVM model according to the condition category includes the following steps: Construct N - 1 SVM sub - models according to the number of condition categories N, where N is a positive integer.

5. The method according to claim 2, wherein The construction of the ACGAN model includes the following steps: Use the de - convolutional neural network to establish the ACGAN generator; Build an ACGAN discriminator using a convolutional neural network.

6. The method according to claim 2, wherein The generation of random noise data containing the status category label includes the following steps: Collect multiple random noise data, add one of the status category labels to each random noise data, and add each of the status category labels to the multiple random noise data.

7. The method according to claim 1, characterized in that, Each time a certain number of iterations is completed, reduce the influence of the SVM loss function on the unified loss function until the unified loss function is not affected by the SVM loss function, including the following steps: Each time a certain number of iterations is completed, increase λ in the Loss expression of the unified loss function by a certain increment until λ increases to 1.

8. The method according to claim 1, wherein After the unified loss function is not affected by the SVM loss function, use ACGAN to independently perform classification training, and obtain a heat pipe network status diagnosis model according to the trained ACGAN discriminator, including the following steps: After λ increases to 1, input the real samples of the heat pipe network images and the pseudo-samples containing the status category labels newly generated by the ACGAN generator into the ACGAN discriminator for classification training, and update the parameters of the ACGAN discriminator according to the unified loss function; After the training termination condition is met, stop updating the parameters of the ACGAN discriminator, and obtain a heat pipe network status diagnosis model according to the trained ACGAN discriminator.

9. The method according to claim 8, wherein The obtaining of the heat pipe network status diagnosis model according to the trained ACGAN discriminator includes the following steps: Detach the trained ACGAN discriminator from the ACGAN model, cancel the output of the trained ACGAN discriminator for distinguishing true and false, and only retain the classification output to obtain a heat pipe network status diagnosis model.

10. A method for diagnosing the condition of a heat pipe network, characterized in that, Include the following steps: Input the collected heat pipe network images into the heat pipe network status diagnosis model obtained by using any one of the methods in claims 1-9 to obtain the status diagnosis result output by the heat pipe network status diagnosis model.

11. A training device for a diagnosis model of a thermal pipeline network, characterized in that, Include: An auxiliary classification module for inputting random noise data containing heat pipe network status category labels into the ACGAN generator to obtain pseudo-samples containing status category labels, inputting the real samples of the heat pipe network images and the pseudo-samples containing status category labels newly generated by the ACGAN generator into the ACGAN discriminator for true / false discrimination training and classification training, and inputting them into the SVM model for classification training; A loss function integration module for integrating the classification loss functions of the ACGAN discriminator and the SVM model into a unified loss function; A parameter update module for updating the parameters of the ACGAN generator according to the true / false discrimination situation of the ACGAN, updating the parameters of the ACGAN discriminator according to the unified loss function, and updating the parameters of the SVM model according to the classification loss function of the SVM model; after iterating the above steps multiple times, each time a certain number of iterations is completed, reduce the influence of the SVM loss function on the unified loss function until the unified loss function is not affected by the SVM loss function; An independent classification module, which is used to perform independent classification training using ACGAN after the unified loss function is not affected by the SVM loss function, and obtain a heat pipe network condition diagnosis model according to the trained ACGAN discriminator; The loss function integration module includes: A discriminator loss function determination unit for determining a classification loss function L of the ACGAN discriminator C ; The SVM loss function determination unit is used to determine the classification loss function L of the SVM model SVM ; A loss function integration unit for determining, according to classification loss functions L C and L SVM , that the expression of the unified loss function Loss is: Loss=(1-λ)L SVM +λL C In the formula, λ is the confidence ratio of the classification result of the ACGAN discriminator, and an arbitrary number greater than 0 and less than 1 is set as the initial value of λ.

12. The device according to claim 11, characterized in that, It also includes: A classification preparation module, which is used to obtain true samples of heat pipe network images; Set status category labels according to the status categories of the true samples of heat pipe network images; Generate random noise data containing the status category labels; Construct an SVM model according to the status category; construct an ACGAN model.

13. The device according to claim 12, characterized in that The classification preparation module includes: A true sample acquisition unit, which is used to collect several heat pipe network image samples under different status categories respectively. The status categories include: normal, pipe wall leakage, pipe bending deformation, and pipe insulation layer damage; perform at least one of the data enhancement processes on the collected heat pipe network image samples under different status categories to obtain heat pipe network image samples after data enhancement processing. The data enhancement processing includes image rotation, cropping, homogeneous splicing, affine transformation, adding verification noise, adding light and shadow, adjusting contrast, and adjusting saturation; randomly combine the heat pipe network image samples after data enhancement processing to obtain randomly combined heat pipe network image samples; the collected heat pipe network image samples under different status categories, the heat pipe network image samples after data enhancement processing, and the randomly combined heat pipe network image samples constitute the true samples of heat pipe network images; A status category label setting unit, which is used to set status category labels according to the status categories of the true samples of heat pipe network images; An SVM model construction unit, which is used to construct N - 1 SVM sub-models according to the number of status categories N, where N is a positive integer; An ACGAN model construction unit, which is used to establish an ACGAN generator using a deconvolution neural network; establish an ACGAN discriminator using a convolutional neural network; A noise data acquisition unit, which is used to collect multiple random noise data, add a status category label to each random noise data, and add each status category label to the multiple random noise data.

14. A device for diagnosing the condition of a heat pipe network, characterized in that, It includes: An image acquisition module, which is used to acquire heat pipe network images; A condition diagnosis module, which is used to input the acquired heat pipe network images into the heat pipe network condition diagnosis model obtained by using any one of the methods in claims 1 - 9, and obtain the condition diagnosis result output by the heat pipe network condition diagnosis model.

15. A terminal device, characterized in that It includes: A memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the heat pipe network condition diagnosis model training method according to any one of claims 1 - 9.

16. A computer storage medium, characterized in that, Computer-executable instructions are stored in the computer storage medium, and when the computer-executable instructions are executed, they implement the heat pipe network condition diagnosis model training method according to any one of claims 1 - 9.

17. A terminal device, characterized in that, It includes: A memory, a processor, and a computer program stored on the memory and executable on the processor, wherein when the processor executes the computer program, the method for diagnosing the condition of a heat pipe network according to claim 10 is implemented.

18. A computer storage medium, characterized in that, Computer-executable instructions are stored in the computer storage medium, and when the computer-executable instructions are executed, the method for diagnosing the condition of a heat pipe network according to claim 10 is implemented.

Images