Object evaluation method and device, computer equipment, readable storage medium and program product

Abstract

The invention relates to an object evaluation method and device, computer equipment, a computer readable storage medium and a computer program product. The method comprises the following steps: acquiring an original sample set of a target object and a random noise vector corresponding to each fault type; for each fault type, performing remodeling operation on a random noise vector corresponding to the fault type according to a target sample generation model to obtain a target multi-dimensional tensor corresponding to the random noise vector, and performing feature amplification on the target multi-dimensional tensor to obtain a generated sample of a first target type corresponding to the original sample set; adding the generated sample to the original sample set to obtain a target sample set; the target sample set is used for training the initial evaluation model. By adopting the method, the accuracy of judging the fault type of the air conditioning system can be improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to an object evaluation method, apparatus, computer equipment, computer-readable storage medium, and computer program product. Background Art

[0002] Data center air conditioning is an important device to ensure the temperature and humidity parameters in the data center room. The data center air conditioning system needs to ensure stable operation. With the development of artificial intelligence technology, equipment anomaly detection and fault identification based on artificial intelligence algorithms can be applied to fault assessment and diagnosis of air conditioning systems.

[0003] In traditional technology, deep learning models are generally used to evaluate air-conditioning system faults. The deep learning model can be SVM (Support Vector Machine), decision tree, random forest, deep neural network, etc. The operating parameter data of the air conditioner is analyzed and processed according to the deep learning classification model to identify whether there is a problem with the operation of the current air-conditioning system. If there is a problem, the fault type of the current air-conditioning system is determined.

[0004] However, in traditional technologies, since failures are rare events, the accuracy of distinguishing the failure type of the air-conditioning system is low. Summary of the invention

[0005] Based on this, it is necessary to provide an object evaluation method, apparatus, computer device, computer-readable storage medium and computer program product to address the above technical issues.

[0006] In a first aspect, the present application provides an object evaluation method, comprising:

[0007] Obtaining an original sample set of the target object and a random noise vector corresponding to each fault type;

[0008] For each of the fault types, reshape the random noise vector corresponding to the fault type according to the target sample generation model to obtain a target multidimensional tensor corresponding to the random noise vector, and perform feature amplification on the target multidimensional tensor to obtain a generated sample of the first target type corresponding to the original sample set;

[0009] The generated samples are added to the original sample set to obtain a target sample set; the target sample set is used to train the initial evaluation model.

[0010] In one embodiment, obtaining an original sample set of the target object includes:

[0011] Acquire operation and maintenance data of the target object in different states, and perform standardization processing on the operation and maintenance data to obtain a first sample set;

[0012] The first sample set is divided in a sliding time pane form to obtain first sample data in a plurality of time series, and the state of the target object in each of the time series is used as a sample label to obtain an original sample set.

[0013] In one embodiment, the target sample generation model includes a generator and a discriminator. Before obtaining the original sample set of the target object and the random noise vector corresponding to each fault type, the method further includes:

[0014] Generate an initial generated sample according to the generator, and analyze the initial generated sample based on the discriminator to obtain a first loss of the generator in the current round;

[0015] Performing a weight update on the generator of the current iteration round based on the first loss to obtain the updated generator of the current iteration round;

[0016] Based on the discriminator, each original sample in the original sample set and each of the initially generated samples are discriminated, and a second loss of the discriminator of the current round is determined according to the discrimination result, the first sample label of the original sample and the second sample label of the initially generated sample;

[0017] Performing a weight update on the discriminator of the current iteration round based on the second loss to obtain the updated discriminator of the current iteration round;

[0018] Execute the step of generating an initial generated sample according to the generator, and analyzing the initial generated sample based on the discriminator to obtain the first loss of the generator in the current round, until the initial sample generation model meets the preset training conditions, and obtains the target sample generation model.

[0019] In one embodiment, the discriminating each original sample and each initially generated sample in the original sample set based on the discriminator, and determining the second loss of the discriminator in the current round according to the discrimination result, the first sample label of the original sample and the second sample label of the initially generated sample, comprises:

[0020] Based on the discriminator, each original sample in the original sample set and each initially generated sample are discriminated to obtain a first discrimination result corresponding to the original sample and a second discrimination result corresponding to the initially generated sample;

[0021] Obtain a first initial loss based on the first discrimination result and the first sample label of the original sample, and obtain a second initial loss based on the second discrimination result and the second sample label of the initially generated sample;

[0022] A second loss of the discriminator of the current round is obtained according to the first initial loss and the second initial loss.

[0023] In one embodiment, after the discriminator discriminates each original sample in the original sample set and each initially generated sample, and determines the second loss of the discriminator in the current round according to the discrimination result, the first sample label of the original sample and the second sample label of the initially generated sample, the method further includes:

[0024] If the first discrimination result of the generator for the original sample and the second discrimination result for the initial generated sample tend to a preset interval, it is determined that the initial sample generation model meets the preset training conditions, and the target sample generation model is obtained.

[0025] In one embodiment, adding the generated sample to the original sample set to obtain the target sample set includes:

[0026] Training a screening discriminator based on original samples of a second target type in the original sample set;

[0027] The generated samples are filtered and screened based on the screening discriminator to obtain target generated samples, and the target generated samples are added to the original sample set to obtain a target sample set.

[0028] In one embodiment, after adding the generated sample to the original sample set to obtain the target sample set, the method further includes:

[0029] Training the initial evaluation model based on the target sample set to obtain a trained target evaluation model;

[0030] The target object is evaluated according to the target evaluation model to obtain an evaluation result.

[0031] In a second aspect, the present application further provides an object evaluation device, comprising:

[0032] An acquisition module, used to acquire an original sample set of the target object and a random noise vector corresponding to each fault type;

[0033] A sample generation module is used to reshape the random noise vector corresponding to each fault type according to a target sample generation model to obtain a target multidimensional tensor corresponding to the random noise vector, and perform feature amplification on the target multidimensional tensor to obtain a generated sample of the first target type corresponding to the original sample set;

[0034] The sample construction module adds the generated samples to the original sample set to obtain a target sample set; the target sample set is used to train the initial evaluation model.

[0035] In one embodiment, the acquisition module is specifically used to acquire operation and maintenance data of the target object in different states, and perform standardization processing on the operation and maintenance data to obtain a first sample set;

[0036] The first sample set is divided in a sliding time pane form to obtain first sample data in a plurality of time series, and the state of the target object in each of the time series is used as a sample label to obtain an original sample set.

[0037] In one embodiment, the target sample generation model includes a generator and a discriminator, and the device further includes:

[0038] A generation module, configured to generate an initial generation sample according to the generator, and analyze the initial generation sample based on the discriminator to obtain a first loss of the generator in the current round;

[0039] A first updating module, configured to perform a weight update on the generator of the current iteration round based on the first loss to obtain the updated generator of the current iteration round;

[0040] A discriminator module, configured to discriminate each original sample in the original sample set and each initially generated sample based on the discriminator, and determine a second loss of the discriminator in a current round according to a discriminator result, a first sample label of the original sample, and a second sample label of the initially generated sample;

[0041] A second updating module, configured to perform a weight update on the discriminator of the current iteration round based on the second loss to obtain the updated discriminator of the current iteration round;

[0042] An iterative module is used to execute the step of generating an initial generated sample according to the generator, analyzing the initial generated sample based on the discriminator, and obtaining the first loss of the generator in the current round, until the initial sample generation model meets the preset training conditions and obtains the target sample generation model.

[0043] In one embodiment, the discrimination module is specifically used to discriminate each original sample and each initially generated sample in the original sample set based on the discriminator to obtain a first discrimination result corresponding to the original sample and a second discrimination result corresponding to the initially generated sample;

[0044] Obtain a first initial loss based on the first discrimination result and the first sample label of the original sample, and obtain a second initial loss based on the second discrimination result and the second sample label of the initially generated sample;

[0045] A second loss of the discriminator of the current round is obtained according to the first initial loss and the second initial loss.

[0046] In one embodiment, the device further comprises:

[0047] A determination module is used to determine that the initial sample generation model meets the preset training conditions and obtain the target sample generation model if the first discrimination result of the generator for the original sample and the second discrimination result for the initial generated sample tend to a preset interval.

[0048] In one of the embodiments, the sample construction module is specifically used to train a screening discriminator based on original samples of the second target type in the original sample set;

[0049] The generated samples are filtered and screened based on the screening discriminator to obtain target generated samples, and the target generated samples are added to the original sample set to obtain a target sample set.

[0050] In one embodiment, the device further comprises:

[0051] A training module, used to train the initial evaluation model based on the target sample set to obtain a trained target evaluation model;

[0052] The evaluation module is used to evaluate the target object according to the target evaluation model to obtain an evaluation result.

[0053] In a third aspect, the present application further provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0054] Obtaining an original sample set of the target object and a random noise vector corresponding to each fault type;

[0055] For each of the fault types, reshape the random noise vector corresponding to the fault type according to the target sample generation model to obtain a target multidimensional tensor corresponding to the random noise vector, and perform feature amplification on the target multidimensional tensor to obtain a generated sample of the first target type corresponding to the original sample set;

[0056] The generated samples are added to the original sample set to obtain a target sample set; the target sample set is used to train the initial evaluation model.

[0057] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the following steps are implemented:

[0058] Obtaining an original sample set of the target object and a random noise vector corresponding to each fault type;

[0059] For each of the fault types, reshape the random noise vector corresponding to the fault type according to the target sample generation model to obtain a target multidimensional tensor corresponding to the random noise vector, and perform feature amplification on the target multidimensional tensor to obtain a generated sample of the first target type corresponding to the original sample set;

[0060] The generated samples are added to the original sample set to obtain a target sample set; the target sample set is used to train the initial evaluation model.

[0061] In a fifth aspect, the present application further provides a computer program product, including a computer program, which implements the following steps when executed by a processor:

[0062] Obtaining an original sample set of the target object and a random noise vector corresponding to each fault type;

[0063] For each of the fault types, reshape the random noise vector corresponding to the fault type according to the target sample generation model to obtain a target multidimensional tensor corresponding to the random noise vector, and perform feature amplification on the target multidimensional tensor to obtain a generated sample of the first target type corresponding to the original sample set;

[0064] The generated samples are added to the original sample set to obtain a target sample set; the target sample set is used to train the initial evaluation model.

[0065] The object evaluation method, device, computer equipment, computer-readable storage medium and computer program product obtain the original sample set of the target object and the random noise vector corresponding to each fault type; for each fault type, the random noise vector corresponding to the fault type is reshaped according to the target sample generation model to obtain the target multidimensional tensor corresponding to the random noise vector, and the target multidimensional tensor is feature-amplified to obtain the generated sample of the first target type corresponding to the original sample set; the generated sample is added to the original sample set to obtain the target sample set; the target sample set is used to train the target evaluation model. With this method, the original sample set is data enhanced by the target generator and the target discriminator to obtain the target generated sample set, which solves the problem of unbalanced distribution of fault data, and then the initial evaluation model is trained by the target generated sample set and the original sample set obtained by data enhancement, which reduces the requirements for sample data collection, improves the training effect for the initial evaluation model, and thus improves the accuracy of air-conditioning system fault detection under unbalanced fault data distribution. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the drawings required for use in the embodiments of the present application or related technical descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0067] Figure 1 A schematic diagram of a flow chart of an object evaluation method in one embodiment;

[0068] Figure 2 A schematic diagram of a process of obtaining an original sample set in one embodiment;

[0069] Figure 3 A schematic diagram of a process for training a target sample to generate a model in one embodiment;

[0070] Figure 4 is a schematic diagram of a generator architecture in one embodiment;

[0071] Figure 5 is a schematic diagram of a discriminator architecture in one embodiment;

[0072] Figure 6 A schematic diagram of an example of generating a sample according to a target sample generation model in one embodiment;

[0073] Figure 7 A schematic diagram of a process for determining a second loss of a discriminator in one embodiment;

[0074] Figure 8is a schematic diagram of a process of screening generated samples according to a screening discriminator in one embodiment;

[0075] Fig. 9 A schematic diagram of a process of evaluating a target object in one embodiment;

[0076] Fig.10 is a schematic diagram of the architecture of a fault classification model in one embodiment;

[0077] Fig.11 A schematic diagram of the change in the SHAP value of the feature value of each sample in one embodiment;

[0078] Fig.12 It is a schematic diagram of the overall influence of each feature of the overall sample on the evaluation result in one embodiment;

[0079] Fig.13 is a structural block diagram of an object evaluation device in one embodiment;

[0080] Fig.14 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0081] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0082] The data center air conditioning system is responsible for precisely adjusting the temperature and humidity parameters in the data center room to ensure the safe and stable operation of IT (Information Technology) equipment under a reasonable temperature and humidity environment. This special role in ensuring the reliable operation of IT equipment also places strict requirements on the reliability and maintenance of the air conditioning system itself. Due to factors such as equipment aging and external environmental erosion, air conditioning equipment will inevitably experience failures in certain components or even the entire system. Fault diagnosis, as an important means of detecting, isolating and identifying faults, has always been one of the main concerns of air conditioning system operation and maintenance. In recent years, with the vigorous development of big data and artificial intelligence technologies, equipment anomaly detection and fault identification using massive operation and maintenance data and artificial intelligence algorithms have become a commonly used method for air conditioning system fault diagnosis.

[0083] According to the different business scenarios and the form of data used, the current intelligent fault diagnosis methods for air-conditioning systems can be divided into two categories: supervised models and unsupervised models. For the fault diagnosis method of supervised models, it is usually applicable to the situation where the system operation and maintenance data has clear labels. Specifically, the collected system status data can clearly label whether the system is faulty or even what type of fault it is in this state. At this time, the fault diagnosis task is regarded as a classification task, and some supervised machine learning and deep learning classification models are widely used, including: SVM, decision tree, random forest, deep neural network, etc. They use labeled business data to learn the data representation of the system when it is normal and faulty, and finally realize the autonomous judgment of whether the system is faulty and the fault category. However, the above methods usually require ideal distribution of labeled data, that is, the ratio of fault data to normal data is balanced, but in reality, faults are occasional events, and the ratio of fault data and non-fault data as sample data is seriously unbalanced. The supervised machine learning and deep learning classification models cannot learn the state characteristics of the air-conditioning system when it is faulty. This imbalance of sample data leads to high errors in the diagnosis of supervised learning models. In addition, the above classification model has low interpretability, and the fault diagnosis process is an opaque "black box" that cannot display the basis for the model's fault classification, posing risks to fault diagnosis decisions.

[0084] Compared with the fault diagnosis method of supervised models, the fault diagnosis method of unsupervised models is usually applicable to the case where the data is not labeled or only has normal data. In this business scenario, unsupervised machine learning and deep learning models are usually used for fault identification and classification. Typical results include clustering algorithms, deep autoencoders, etc. For example, by performing unsupervised clustering on business data, the normal and fault categories of the equipment are automatically divided according to the assumption that the same state of the air-conditioning system has similar data characteristics. In addition, some solutions use the autoencoder method to only learn the data representation when the air-conditioning system is in a normal state. In actual use, when the data state of the air-conditioning system is obviously different from the normal data characteristics learned by the autoencoder model, the current state is judged to be abnormal. However, the above two types of methods of unsupervised machine learning, deep learning models and autoencoders also have two defects: on the one hand, unsupervised machine learning, deep learning models and autoencoder models can only realize fault detection, and cannot solve the problem of multi-type fault identification; on the other hand, due to the overfitting problem of autoencoder learning, the autoencoder may be judged as a fault for data states that are slightly different from normal training samples, resulting in false alarms of faults, resulting in poor accuracy in air-conditioning system evaluation.

[0085] In summary, the current data center air conditioning system lacks a practical and logically reliable computer room fault diagnosis method that can overcome data imbalance and unclear fault modes, and achieve air conditioning system anomaly detection and accurate identification of multiple types of faults that can be explained by diagnostic logic.

[0086] In one embodiment, Figure 1 As shown, an object evaluation method is provided. This embodiment uses the method applied to a terminal as an example for illustration. It can be understood that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0087] Step 102: Obtain an original sample set of the target object and a random noise vector corresponding to each fault type.

[0088] In the embodiment of the present application, the target object may be a computer room air conditioning system, and the terminal obtains the operating parameters of the computer room air conditioning system in normal state and fault state, including the cabinet measuring point temperature, water valve opening, air conditioning fan real-time speed, air conditioning fan speed setting, current air conditioning setting temperature, current air conditioning return air temperature, current air conditioning outlet temperature, adjacent air conditioning setting temperature, adjacent air conditioning return air temperature, adjacent air conditioning outlet temperature, forming an original sample set in the form of time series data. Specifically, various sensors are deployed in the computer room air conditioning system to collect parameters such as cabinet measuring point temperature, water valve opening, air conditioning fan real-time speed, etc. The terminal records sensor data based on a data acquisition system (e.g., a data recorder, an Internet of Things platform, etc.) to form time series data, and stores the collected data in a database or file to form an original sample set in the form of time series data.

[0089] The random noise vector is a time-series random noise vector with the same data dimension as the real fault state. The random noise vector can be a random number of Gaussian distribution (normal distribution) or a random number of other distributions. The random noise vector is used as the input of the target sample generation model to generate samples of specific fault types. For example, for three fault types (fault 1, fault 2, fault 3), the terminal can use the random number generation function in Python's numpy library to generate three random noise vectors.

[0090] Step 104, for each fault type, reshape the random noise vector corresponding to the fault type according to the target sample generation model to obtain a target multidimensional tensor corresponding to the random noise vector, and perform feature amplification on the target multidimensional tensor to obtain a generated sample of the first target type corresponding to the original sample set.

[0091] In an embodiment of the present application, the target sample generation model includes a fault data generator. After the terminal inputs the random noise vector into the target sample generation model, the fault data generator reshapes the one-dimensional random noise vector into a target multidimensional tensor to meet the input requirements of the target sample generation model. After obtaining the reshaped target multidimensional tensor, the fault data generator performs feature amplification on the multidimensional tensor to generate a generated sample of the first target type similar to the original sample set. Among them, the first target type can be sample data that is difficult to obtain under normal conditions of the target object. For example, if the target object is a computer room air conditioning system, the number of fault data of the computer room air conditioning system is relatively small, and the fault data is used as a negative sample in the original sample set, then the first target type is a negative sample type.

[0092] Optionally, when the terminal generates the generated sample of the first target type through the target sample generation model, the terminal may simultaneously train the fault data discriminator and the fault data generator of the target sample generation model.

[0093] Step 106: Add the generated samples to the original sample set to obtain the target sample set.

[0094] Among them, the target sample set is used to train the initial evaluation model.

[0095] In an embodiment of the present application, the terminal uses the generated samples generated by the target sample generation model as additional supplementary samples of the first target type, for example, negative samples of the computer room air conditioning system in a faulty state, and adds the generated samples to the original sample set to form an extended target sample set, thereby supplementing the fault data that is difficult to obtain for the computer room air conditioning system, and making the ratio of positive samples to negative samples tend to be balanced.

[0096] In the above object evaluation method, the original sample set is data enhanced by the target generator and the target discriminator to obtain the target generated sample set, which solves the problem of unbalanced distribution of fault data. Then, the initial evaluation model is trained by the target generated sample set and the original sample set obtained by data enhancement, which reduces the requirements for sample data collection and improves the training effect of the initial evaluation model, thereby improving the accuracy of fault detection of the computer room air conditioning system under unbalanced distribution of fault data.

[0097] In an exemplary embodiment, Figure 2 As shown, step 102 includes steps 202 to 204. Among them:

[0098] Step 202: obtain operation and maintenance data of the target object in different states, and perform standardization processing on the operation and maintenance data to obtain a first sample set.

[0099] In an embodiment of the present application, the terminal obtains operation and maintenance data of the computer room air-conditioning system in different states. The different states of the target object may be normal state, fault state, critical state, etc. For example, the normal state is that the computer room air-conditioning system is cooling normally, the fault state is that the air-conditioning system of the computer room fails to cool, and the critical state is that the temperature of the computer room air-conditioning system gradually rises and is close to overheating.

[0100] The operation and maintenance data may be the time series data of the operation and maintenance of the air conditioner in the computer room. Since the dimensions and value ranges of different parameters may vary greatly, for example, the temperature may be in degrees Celsius, the speed may be in revolutions per minute (RPM), and the water valve opening may be between 0 and 100%. In order to ensure that the weights of various operation parameters are balanced during the initial evaluation model training process, these operation parameters need to be standardized, and the standardization method may be Z-score standardization. That is, the terminal performs feature standardization on the time series data of the operation and maintenance of the air conditioner in the computer room to obtain a first sample set, and each sample in the first sample set may be a multidimensional feature vector, which represents the state of the target object at a certain moment.

[0101] Step 204 , divide the first sample set according to the sliding time pane form to obtain first sample data in multiple time series, and use the state of the target object in each time series as a sample label to obtain an original sample set.

[0102] In an embodiment of the present application, the terminal divides the first sample set into multiple time series data, and assigns a label to each time series data, and finally forms an original sample set. The original sample set will be used for the subsequent training of the initial evaluation model, and the target evaluation model will be obtained after training. The target evaluation model is a neural network model for fault detection of the computer room air conditioning system. Sliding time pane is a commonly used data partitioning method, which is used to divide continuous time series data into multiple subsequences.

[0103] The width of the sliding pane indicates the number of time steps of each subsequence. For example, a pane width of 100 indicates that each subsequence contains data of 100 time steps. The step length of the sliding pane indicates the number of time steps of each slide. For example, a step length of 10 indicates 10 time steps of each slide. The terminal divides the first sample set by the sliding pane and can extract multiple time series data from the first sample set. Specifically, the terminal applies the sliding time pane to the first sample set to extract the first sample data under multiple time series.

[0104] The terminal applies a sliding time pane to the first sample set after the normalization according to the preset time pane width and compensation to extract multiple time series data. For example, if the first sample set is a two-dimensional array with dimensions (T, F), where T is the number of time steps and F is the number of features (i.e., the various operating parameters mentioned above), then using a pane width of 100 and a step size of 10, multiple subsequences of shape (100, F) can be generated.

[0105] For each extracted time series data, the terminal determines its corresponding label, that is, the state of the computer room air conditioning system in the time series. Among them, the sample label can be "normal" or "faulty", depending on the actual state of the system in the time period. Finally, the terminal pairs each time series data with its corresponding label to form the original sample set.

[0106] In this embodiment, the accuracy and comparability of the data are ensured by acquiring and standardizing the operation and maintenance data of the computer room air conditioner in different states, and the sliding time pane technology is used to divide the standardized data into multiple time series samples, and a corresponding state label is assigned to each sample. Not only the time series characteristics of the computer room air conditioner system are effectively extracted, but also a high-quality data set is provided for subsequent model training, thereby improving the accuracy and efficiency of fault detection.

[0107] In an exemplary embodiment, Figure 3 As shown, the target sample generation model includes a generator and a discriminator. Before step 102, the method further includes steps 302 to 310. Among them:

[0108] Step 302: Generate an initial generated sample based on the generator, and analyze the initial generated sample based on the discriminator to obtain the first loss of the generator in the current round.

[0109] In an embodiment of the present application, the terminal generates fault data through the fault data enhancement module of the Atten-GAN mesh wheel architecture as a negative sample for training the initial evaluation model. Before applying the Atten-GAN mesh wheel architecture, the Atten-GAN mesh wheel architecture needs to be trained. Specifically, the Atten-GAN mesh wheel architecture includes a generator and a discriminator. The terminal alternately trains the generator and the discriminator. First, the terminal inputs a noise vector into the initial generator and randomly outputs a multi-dimensional time series data as the generated data. Then, the discriminator discriminates the difference between the data generated by the initial generator and the real data, and obtains the first loss of the generator of the current round. Figure 4 As shown, Figure 4The figure is the generator architecture diagram, in which the generator includes a linear dimension increase module, a content perception module of the self-attention mechanism, a convolution time series perception module, and a feature attention perception module. The function of the linear dimension increase module is to transform the random Gaussian noise through multiple layers of linear transformation to generate random data of the same size as the original time series data; the content perception module of the self-attention mechanism calculates the important relationship between the internal features and the data at each moment through the self-attention mechanism and the feedforward neural network; the function of the convolution time series information module is to use a one-dimensional convolution with a convolution kernel size of 1 to capture the time dimension relationship of the data in the time series direction; the feature attention perception module includes three components: global average pooling, one-dimensional convolution, and sigmoid function. The feature attention perception module uses global average pooling and sigmoid to weight the data according to the feature dimension, and multiplies the calculation result of the feature attention with the result of the convolution time series information module to capture the relationship between data features. Finally, a one-dimensional convolution module is used to obtain the generated data of the same size as the real fault data. The generator is continuously optimized to gradually make the generated data similar to the actual sample until the discriminator cannot distinguish the generated data from the real data. In addition, through the generative model and learning strategy of the Atten-GAN mesh wheel architecture, the generated data can also have a certain noise change, increasing the diversity of fault sample (negative sample) features.

[0110] The goal of the generator is to make the discriminator unable to distinguish between generated samples and real samples, so the loss function of the generator is based on the discriminator's judgment of the generated samples. The loss function of the generator can be a cross entropy loss, which represents the probability that the generated sample is misclassified as a real sample by the discriminator. The calculation formula is as follows:

[0111]

[0112] in, is a random noise sample, is the output of the discriminator for the initial generated sample, and m is the batch size.

[0113] Step 304: Update the weight of the generator of the current iteration round based on the first loss to obtain the updated generator of the current iteration round.

[0114] In the embodiment of the present application, the terminal uses the first loss of the generator for back propagation, calculates the gradient of the generator parameters, and the gradient indicates the direction and magnitude of the generator parameters that need to be adjusted, and updates the weight of the generator of the current round to obtain the updated generator of the current iteration round. The update rule for the generator weight update is:

[0115]

[0116] in, is the batch size, is the learning rate, represents the gradient of the generator, is a real data sample, is a random noise sample.

[0117] Step 306: discriminate each original sample and each initially generated sample in the original sample set based on the discriminator, and determine the second loss of the discriminator of the current round according to the discrimination result, the first sample label of the original sample and the second sample label of the initially generated sample.

[0118] In the embodiment of the present application, the discriminator takes each original sample and each initially generated sample in the original sample set as input, analyzes the input samples through its network structure, and outputs a score value for each input sample, indicating the probability that the input sample is a true sample. Figure 5 As shown, Figure 5 : is a diagram of the discriminator architecture, wherein the discriminator includes a linear residual connection module, a content perception module of a self-attention mechanism, a convolutional time series perception module, and a feature attention perception module. Compared with the generator structure, the first module of the discriminator network is a linear residual connection module. The purpose of the linear residual connection module is to use linear residual links to perform some shallow data feature extraction on the data. The structures of the remaining three modules are the same as those of the generator network and will not be repeated in this embodiment. In addition, the final output of the discriminator network is a score value given by a Sigmoid unit, which represents the authenticity discrimination score of the sample. The input of the discriminator is real fault data and generated fault data. The purpose of optimizing the discriminator is to detect as accurately as possible whether the input sample is a real sample composed of real fault data or a generated sample generated by the generator.

[0119] The goal of the discriminator is to maximize its ability to correctly distinguish between real samples and generated samples. The loss function of the discriminator is based on its discrimination results and the first sample label and the second sample label. The second loss calculation formula is:

[0120]

[0121] in, It is a real sample. is the output of the discriminator for each original sample in the original sample set, is the output of the discriminator for each initially generated sample.

[0122] Step 308: Update the weight of the discriminator of the current iteration round based on the second loss to obtain the updated discriminator of the current iteration round.

[0123] In the embodiment of the present application, the terminal uses the second loss of the discriminator for back propagation to calculate the gradient of the discriminator parameters, and the discriminator update rule is:

[0124]

[0125] in, is the batch size, is the learning rate, represents the gradient of the discriminator, is a real data sample, is a random noise sample.

[0126] Step 310, execute the step of generating an initial generated sample according to the generator, and analyzing the initial generated sample based on the discriminator to obtain the first loss of the generator of the current round, until the initial sample generation model meets the preset training conditions and obtains the target sample generation model.

[0127] In the embodiment of the present application, under the guidance of the discriminator, the data generated by the generator gradually becomes more realistic than the real data. At the same time, the discriminator also gradually improves its discrimination performance until the discriminator cannot distinguish between the generated data and the real data. Specifically, let the generator be , which receives random noise as input and produces data samples . Let the discriminator be , accept the sample And output a scalar, indicating the probability that the sample comes from the true data distribution. During the training process, and The optimization is performed with the following objective function:

[0128]

[0129] in, is the actual data distribution, is the random noise distribution of the generator input. The goal of the generator is to minimize this function, while the goal of the discriminator is to maximize it.

[0130] Through the iterative process of the generator and the discriminator in step 304 and step 308, the generator will generate more and more realistic data, while the discriminator will better distinguish between real and generated data.

[0131] In a specific embodiment, Figure 6As shown in the figure, the model structure of the target sample generation model is the Atten-GAN network, and the model structure of the initial evaluation model is the TabNet network. The terminal receives the input data, standardizes the input data, and obtains a large number of normal samples and a small number of fault samples. Then, the terminal inputs the fault sample into the Atten-GAN network, generates the input noise through the Atten-GAN network, obtains the generated data G(z), and discriminates the real data x and the generated data G(z) according to the discriminator to obtain the first loss of the generator, and updates the generator and the discriminator according to the first loss until the discriminator cannot distinguish the generated data from the real data, and obtains a large number of fault samples generated by the Atten-GAN network. The TabNet network is trained with the generated fault samples and the original normal samples to obtain the trained TabNet network, and the importance of the fault features is explained according to the SHAP method. The SHAP value of each feature in the feature vector to the model output is calculated to indicate the contribution of each feature to the prediction result, so as to achieve the explainability of important features.

[0132] In the embodiment of the present application, the generator and discriminator of the Atten-GAN network not only capture the time series relationship before and after the time series data when generating data and performing identification tasks, but also capture the relationship between features for multi-dimensional time series feature data, ensuring the similarity between the generated data and the original fault data in terms of time series changes. At the same time, the fault data obtained by the generation mechanism also contains a certain diversity of noise, enriching the types of fault data. The generator and discriminator of the Atten-GAN network solve the problems of fault detection and fault type identification of air-conditioning systems under unbalanced distribution of fault data. Compared with the existing fault diagnosis method based on self-encoding, this method can not only realize fault detection, but also classify and determine specific fault types. Compared with the existing deep learning classification method, this method overcomes the problem of low fault classification accuracy caused by data imbalance.

[0133] The Atten-GAN network not only realizes the learning and perception of the temporal relationship of the data through the convolutional temporal perception module, but also designs a feature perception module based on feature attention. It realizes the capture of the relationship between features in multi-dimensional temporal feature data, ensuring the high similarity between the generated data and the original fault data in terms of temporal changes.

[0134] In an exemplary embodiment, Figure 7 As shown, step 306 includes steps 702 to 706. Among them:

[0135] Step 702: discriminate each original sample and each initially generated sample in the original sample set based on the discriminator to obtain a first discrimination result corresponding to the original sample and a second discrimination result corresponding to the initially generated sample.

[0136] In an embodiment of the present application, the discriminator receives two groups of input samples, including one group of original samples consisting of real samples in the original sample set, and the other group of initial generated samples generated by the generator. The discriminator analyzes each input sample through its network structure (including a linear dimension increase module, a content perception module of the self-attention mechanism, a convolutional temporal perception module, and a feature attention perception module) to determine whether the input sample is a real sample or a generated sample. For each original sample, the discriminator outputs a first discrimination result, indicating the probability that the original sample is judged as a real sample; for each initial generated sample, the discriminator outputs a second discrimination result, indicating the probability that the initial generated sample is judged as a real sample.

[0137] Step 704: A first initial loss is obtained based on the first discrimination result and the first sample label of the original sample, and a second initial loss is obtained based on the second discrimination result and the second sample label of the initially generated sample.

[0138] In an embodiment of the present application, for the original sample, the terminal uses a loss function such as cross entropy to calculate a first initial loss, and the first initial loss reflects the ability of the discriminator to correctly distinguish the original sample as a real sample; for the initial generated sample, the terminal also uses a loss function to calculate a second initial loss, and the second initial loss reflects the ability of the discriminator to correctly distinguish the generated sample as a non-real sample.

[0139] Step 706: Obtain a second loss of the discriminator of the current round according to the first initial loss and the second initial loss.

[0140] In the embodiment of the present application, the terminal combines the first initial loss and the second initial loss to obtain the second loss of the discriminator in the current round. The second loss is the overall performance indicator of the discriminator in the current round of training, reflecting the overall ability of the discriminator in distinguishing between real samples and generated samples. It is used to update the discriminator weights in subsequent steps to optimize the performance of the discriminator.

[0141] In this embodiment, the discriminator can calculate the loss based on its discrimination results and sample labels, thereby guiding its own weight update and performance improvement. In the training process of the target sample generation model, this self-optimization of the discriminator is mutually antagonistic and mutually reinforcing with the optimization of the generator, and jointly improves the accuracy of sample generation of the target sample generation model.

[0142] In an exemplary embodiment, step 306 includes step 3061. In which:

[0143] Step 3061, if the first discrimination result of the generator for the original sample and the second discrimination result for the initial generated sample tend to the preset interval, it is determined that the initial sample generation model meets the preset training conditions, and the target sample generation model is obtained.

[0144] In the embodiment of the present application, the goal of the generator is to make the generated samples as close to the real samples as possible, so that the discriminator is difficult to distinguish. Therefore, the discrimination result of the generator should be close to the discrimination result of the real sample. The preset interval means that the difference between the discrimination result of the generator and the discrimination result of the real sample should be within an acceptable range. For example, if the output of the discriminator for the real sample is close to 1 and the output for the generated sample is close to 0, then the generator needs to make the discrimination result of the generated sample close to 1.

[0145] As the training progresses, the quality of samples generated by the generator improves, and the output of the discriminator for the generated samples will gradually approach the output for the real samples. When the difference between the discrimination results of the generated samples and the real samples enters the preset interval, it indicates that the generator is already able to generate high-quality samples. At this time, it means that the generator is already able to generate high-quality generated samples. Furthermore, once the difference in the discrimination results meets the preset interval, it can be considered that the generator has been fully trained and the generated samples are sufficient to confuse the discriminator. At this point, the terminal obtains a generator and discriminator that meet the training conditions. The current generator and discriminator models can be used as target sample generation models to generate new fault samples to enhance the negative samples in the original sample set.

[0146] In this embodiment, by training the generator and the discriminator, the generator can generate fault samples that are highly similar to real samples. When the difference between the generator's discrimination results for the generated samples and the real samples enters the preset range, it indicates that the generator has the ability to generate high-quality samples and can meet the training conditions, thereby obtaining a target sample generation model. The target sample generation model can be used to generate new fault samples, effectively enhancing the negative samples in the original sample set, and improving the sample diversity and the fault detection capability of the model.

[0147] In an exemplary embodiment, Figure 8 As shown, step 106 includes steps 802 to 804. Among them:

[0148] Step 802 : training a screening discriminator based on original samples of the second target type in the original sample set.

[0149] In an embodiment of the present application, the second sample type may be a positive sample type, and the terminal trains a screening discriminator through the second sample type, and the screening discriminator only learns data features of the positive sample type.

[0150] Step 804 , filtering and screening the generated samples based on the screening discriminator to obtain target generated samples, and adding the target generated samples to the original sample set to obtain the target sample set.

[0151] In the embodiment of the present application, the terminal inputs the generated sample of each negative sample type into the screening discriminator, and filters the generated sample based on the screening discriminator. If the generated sample is misjudged as normal, it indicates that the generated sample is too similar to the non-fault data, and the generated sample is removed to obtain the target generated sample. The target generated sample has obvious fault characteristics and can effectively enhance the fault data in the original sample set.

[0152] In this embodiment, the generated samples are filtered and screened by a screening discriminator to obtain target generated samples, and a target sample set is constructed based on the target generated samples. This can effectively improve the quality of the generated samples and enhance the fault detection capability of the model, while ensuring that the generated samples can better represent the characteristics of the fault data.

[0153] In an exemplary embodiment, Fig. 9 As shown, after step 106, the method further includes steps 902 to 904. Among them:

[0154] Step 902: Train the initial evaluation model based on the target sample set to obtain a trained target evaluation model.

[0155] In the embodiment of the present application, the initial evaluation model is a fault classification model based on the TabNet encoder, such as Fig.10 As shown in the figure, the fault classification model includes a feature transformer and an attention transformer. The feature transformer consists of five fully connected layers, and then connects the LayerNorm and ReLU activation functions to calculate the output; the attention transformer consists of one fully connected layer, one LayerNorm layer, and Sparsemax to select features. Through the combination of three layers of feature transformers and attention transformers, important features are extracted and learned, and finally the model classification is completed through the Softmax layer.

[0156] The terminal trains the initial evaluation model based on the target sample set. After training with the target sample set, the parameters of the initial evaluation model are continuously optimized, and finally a trained target evaluation model is obtained.

[0157] Step 904: Evaluate the target object according to the target evaluation model to obtain an evaluation result.

[0158] In an embodiment of the present application, the trained target evaluation model is used to perform fault evaluation on the target object. The target object is the time series data of the operation and maintenance of the computer room air conditioner to be evaluated. The terminal inputs the time series data of the operation and maintenance of the computer room air conditioner to be evaluated into the target evaluation model, and processes it through the feature converter and attention converter of the target evaluation model. After extracting the key features, it is classified through the Softmax layer to finally obtain the evaluation result. The evaluation result reflects whether there is a fault in the target object and the type of fault, which provides an accurate judgment basis for fault detection and classification. Specifically, the target evaluation model first adjusts the structure of the multi-dimensional time series data of the target object and flattens the data in sequence. The purpose of the above adjustment is to make the target evaluation model consider not only the factors between the features in the calculation process, but also the changing factors of each feature at different times. On the other hand, the normalization method of the feature converter and the attention module adopts the LayerNorm module, which has the advantage of adapting to the inconsistent length of time series data.

[0159] After the multi-dimensional time series data is input into the target evaluation model, the feature converter first passes the initial input features to obtain the initial feature representation. The output of this feature converter will become the input of the attention module, and the hyperparameter step sets the number of model iterations. Fig.10 As shown in the figure, in each step, the target evaluation model receives data features as input and uses the output of the previous step to weight the data features. The output of each step is accumulated for the final decision.

[0160] In an optional embodiment, the SHAP method is used to interpret the importance of fault features. The SHAP value of each feature in the feature vector to the output of the target evaluation model is calculated to indicate the contribution of each feature to the prediction result, thereby achieving interpretability of important features. Fig.11 and Fig.12 As shown, Fig.11 The values ​​in show the changes in the size of the eigenvalues ​​for each sample. Fig.12 For the overall sample, the overall impact of each feature on the evaluation results is shown.

[0161] In this embodiment, the initial evaluation model is trained based on the target sample set to obtain a trained target evaluation model. The target evaluation model adopts the fault classification model of the TabNet encoder, including a feature transformer and an attention transformer, which can effectively extract and learn key features in multidimensional time series data. The target evaluation model not only considers the factors between each feature, but also considers the changing factors of each feature at different times, and adapts to the problem of inconsistent length of time series data through the LayerNorm module. Finally, the target evaluation model can accurately evaluate whether the target object has a fault and its fault type, provide a reliable judgment basis for fault detection and classification, and significantly improve the accuracy and reliability of fault detection.

[0162] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0163] Based on the same inventive concept, the embodiment of the present application also provides an object evaluation device for implementing the object evaluation method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more object evaluation device embodiments provided below can refer to the limitations of the object evaluation method above, and will not be repeated here.

[0164] In an exemplary embodiment, Fig.13 As shown, an object evaluation device 1300 is provided, comprising: an acquisition module 1301, a sample generation module 1302 and a sample construction module 1303, wherein:

[0165] An acquisition module 1301 is used to acquire an original sample set of a target object and a random noise vector corresponding to each fault type;

[0166] The sample generation module 1302 is used to reshape the random noise vector corresponding to each fault type according to the target sample generation model to obtain a target multidimensional tensor corresponding to the random noise vector, and perform feature amplification on the target multidimensional tensor to obtain a generated sample of the first target type corresponding to the original sample set;

[0167] The sample construction module 1303 adds the generated samples to the original sample set to obtain a target sample set; the target sample set is used to train the initial evaluation model.

[0168] In one embodiment, the acquisition module 1301 is specifically used to acquire operation and maintenance data of the target object in different states, and perform standardization processing on the operation and maintenance data to obtain a first sample set;

[0169] The first sample set is divided according to the sliding time pane form to obtain first sample data in multiple time series, and the state of the target object in each time series is used as a sample label to obtain the original sample set.

[0170] In one embodiment, the target sample generation model includes a generator and a discriminator, and the device 1300 further includes:

[0171] A generation module, used to generate an initial generation sample according to the generator, and analyze the initial generation sample based on the discriminator to obtain the first loss of the generator in the current round;

[0172] A first updating module, used to update the weight of the generator of the current iteration round based on the first loss to obtain the updated generator of the current iteration round;

[0173] A discriminator module is used to discriminate each original sample and each initially generated sample in the original sample set based on the discriminator, and determine the second loss of the discriminator of the current round according to the discriminator result, the first sample label of the original sample and the second sample label of the initially generated sample;

[0174] A second updating module is used to update the weight of the discriminator of the current iteration round based on the second loss to obtain the updated discriminator of the current iteration round;

[0175] The iterative module is used to generate initial generation samples according to the generator, analyze the initial generation samples based on the discriminator, and obtain the first loss of the generator in the current round, until the initial sample generation model meets the preset training conditions and obtains the target sample generation model.

[0176] In one embodiment, the discrimination module is specifically used to discriminate each original sample and each initially generated sample in the original sample set based on the discriminator, and obtain a first discrimination result corresponding to the original sample and a second discrimination result corresponding to the initially generated sample;

[0177] A first initial loss is obtained based on the first discrimination result and the first sample label of the original sample, and a second initial loss is obtained based on the second discrimination result and the second sample label of the initially generated sample;

[0178] The second loss of the discriminator of the current round is obtained according to the first initial loss and the second initial loss.

[0179] In one embodiment, the apparatus 1300 further includes:

[0180] The determination module is used to determine that the initial sample generation model meets the preset training conditions and obtain the target sample generation model if the first discrimination result of the generator for the original sample and the second discrimination result for the initial generated sample tend to the preset interval.

[0181] In one embodiment, the sample construction module 1303 is specifically used to train a screening discriminator based on the original samples of the second target type in the original sample set;

[0182] The generated samples are filtered based on the screening discriminator to obtain the target generated samples, and the target generated samples are added to the original sample set to obtain the target sample set.

[0183] In one embodiment, the apparatus 1300 further includes:

[0184] A training module is used to train the initial evaluation model based on the target sample set to obtain a trained target evaluation model;

[0185] The evaluation module is used to evaluate the target object according to the target evaluation model to obtain the evaluation result.

[0186] Each module in the above object evaluation device can be implemented in whole or in part by software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to each module.

[0187] In an exemplary embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as shown in FIG. Fig.14As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store sample data. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, an object evaluation method is implemented.

[0188] Those skilled in the art will understand that Fig.14 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0189] In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0190] Obtaining an original sample set of the target object and a random noise vector corresponding to each fault type;

[0191] For each of the fault types, reshape the random noise vector corresponding to the fault type according to the target sample generation model to obtain a target multidimensional tensor corresponding to the random noise vector, and perform feature amplification on the target multidimensional tensor to obtain a generated sample of the first target type corresponding to the original sample set;

[0192] The generated samples are added to the original sample set to obtain a target sample set; the target sample set is used to train the initial evaluation model.

[0193] In one embodiment, when the processor executes the computer program, the following steps are also implemented:

[0194] Acquire operation and maintenance data of the target object in different states, and perform standardization processing on the operation and maintenance data to obtain a first sample set;

[0195] The first sample set is divided in a sliding time pane form to obtain first sample data in a plurality of time series, and the state of the target object in each of the time series is used as a sample label to obtain an original sample set.

[0196] In one embodiment, when the processor executes the computer program, the following steps are also implemented:

[0197] Generate an initial generated sample according to the generator, and analyze the initial generated sample based on the discriminator to obtain a first loss of the generator in the current round;

[0198] Performing a weight update on the generator of the current iteration round based on the first loss to obtain the updated generator of the current iteration round;

[0199] Based on the discriminator, each original sample in the original sample set and each of the initially generated samples are discriminated, and a second loss of the discriminator of the current round is determined according to the discrimination result, the first sample label of the original sample and the second sample label of the initially generated sample;

[0200] Performing a weight update on the discriminator of the current iteration round based on the second loss to obtain the updated discriminator of the current iteration round;

[0201] Execute the step of generating an initial generated sample according to the generator, and analyzing the initial generated sample based on the discriminator to obtain the first loss of the generator in the current round, until the initial sample generation model meets the preset training conditions, and obtains the target sample generation model.

[0202] In one embodiment, when the processor executes the computer program, the following steps are also implemented:

[0203] Based on the discriminator, each original sample in the original sample set and each initially generated sample are discriminated to obtain a first discrimination result corresponding to the original sample and a second discrimination result corresponding to the initially generated sample;

[0204] Obtain a first initial loss based on the first discrimination result and the first sample label of the original sample, and obtain a second initial loss based on the second discrimination result and the second sample label of the initially generated sample;

[0205] A second loss of the discriminator of the current round is obtained according to the first initial loss and the second initial loss.

[0206] In one embodiment, when the processor executes the computer program, the following steps are also implemented:

[0207] If the first discrimination result of the generator for the original sample and the second discrimination result for the initial generated sample tend to a preset interval, it is determined that the initial sample generation model meets the preset training conditions, and the target sample generation model is obtained.

[0208] In one embodiment, when the processor executes the computer program, the following steps are also implemented:

[0209] Training a screening discriminator based on original samples of a second target type in the original sample set;

[0210] The generated samples are filtered and screened based on the screening discriminator to obtain target generated samples, and the target generated samples are added to the original sample set to obtain a target sample set.

[0211] In one embodiment, when the processor executes the computer program, the following steps are also implemented:

[0212] Training the initial evaluation model based on the target sample set to obtain a trained target evaluation model;

[0213] The target object is evaluated according to the target evaluation model to obtain an evaluation result.

[0214] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0215] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0216] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0217] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, an artificial intelligence (AI) processor, etc., but are not limited to this.

[0218] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0219] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. An object evaluation method, characterized in that: The method comprises: Obtaining an original sample set of the target object and a random noise vector corresponding to each fault type; For each of the fault types, reshape the random noise vector corresponding to the fault type according to the target sample generation model to obtain a target multidimensional tensor corresponding to the random noise vector, and perform feature amplification on the target multidimensional tensor to obtain a generated sample of the first target type corresponding to the original sample set; The generated samples are added to the original sample set to obtain a target sample set; the target sample set is used to train the initial evaluation model.

2. The method according to claim 1, characterized in that The obtaining of the original sample set of the target object includes: Acquire operation and maintenance data of the target object in different states, and perform standardization processing on the operation and maintenance data to obtain a first sample set; The first sample set is divided in a sliding time pane form to obtain first sample data in a plurality of time series, and the state of the target object in each of the time series is used as a sample label to obtain an original sample set.

3. The method according to claim 1, characterized in that The target sample generation model includes a generator and a discriminator. Before obtaining the original sample set of the target object and the random noise vector corresponding to each fault type, the method further includes: Generate an initial generated sample according to the generator, and analyze the initial generated sample based on the discriminator to obtain a first loss of the generator in the current round; Performing a weight update on the generator of the current iteration round based on the first loss to obtain the updated generator of the current iteration round; Based on the discriminator, each original sample in the original sample set and each of the initially generated samples are discriminated, and a second loss of the discriminator of the current round is determined according to the discrimination result, the first sample label of the original sample and the second sample label of the initially generated sample; Performing a weight update on the discriminator of the current iteration round based on the second loss to obtain the updated discriminator of the current iteration round; Execute the step of generating an initial generated sample according to the generator, and analyzing the initial generated sample based on the discriminator to obtain the first loss of the generator in the current round, until the initial sample generation model meets the preset training conditions, and obtains the target sample generation model.

4. The method according to claim 3, characterized in that The discriminating each original sample in the original sample set and each initially generated sample based on the discriminator, and determining the second loss of the discriminator in the current round according to the discrimination result, the first sample label of the original sample and the second sample label of the initially generated sample, comprises: Based on the discriminator, each original sample in the original sample set and each initially generated sample are discriminated to obtain a first discrimination result corresponding to the original sample and a second discrimination result corresponding to the initially generated sample; Obtain a first initial loss based on the first discrimination result and the first sample label of the original sample, and obtain a second initial loss based on the second discrimination result and the second sample label of the initially generated sample; A second loss of the discriminator of the current round is obtained according to the first initial loss and the second initial loss.

5. The method according to claim 4, characterized in that After discriminating each original sample in the original sample set and each initially generated sample based on the discriminator, and determining the second loss of the discriminator in the current round according to the discrimination result, the first sample label of the original sample and the second sample label of the initially generated sample, the method further includes: If the first discrimination result of the generator for the original sample and the second discrimination result for the initial generated sample tend to a preset interval, it is determined that the initial sample generation model meets the preset training conditions, and the target sample generation model is obtained.

6. The method according to claim 1, characterized in that The step of adding the generated samples to the original sample set to obtain a target sample set includes: Training a screening discriminator based on original samples of a second target type in the original sample set; The generated samples are filtered and screened based on the screening discriminator to obtain target generated samples, and the target generated samples are added to the original sample set to obtain a target sample set.

7. The method according to claim 1, characterized in that After adding the generated samples to the original sample set to obtain the target sample set, the method further includes: Training the initial evaluation model based on the target sample set to obtain a trained target evaluation model; The target object is evaluated according to the target evaluation model to obtain an evaluation result.

8. An object evaluation device, characterized in that The device comprises: An acquisition module, used to acquire an original sample set of a target object and a random noise vector corresponding to each fault type; A sample generation module is used to reshape the random noise vector corresponding to each fault type according to a target sample generation model to obtain a target multidimensional tensor corresponding to the random noise vector, and perform feature amplification on the target multidimensional tensor to obtain a generated sample of the first target type corresponding to the original sample set; The sample construction module adds the generated samples to the original sample set to obtain a target sample set; the target sample set is used to train the initial evaluation model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images