Water chilling unit fault detection method and system based on working condition and fault feature separation

Abstract

The invention belongs to the technical field of heating ventilation air conditioner fault detection, and discloses a water chilling unit fault detection method and system based on working condition and fault feature separation, and the method comprises the steps: collecting the operation data of a water chilling unit; working condition features and fault features of the training data are extracted and decoded into a working condition mode and a fault mode, operation data are reconstructed, a loss function is defined, and the model is trained; fault features of normal operation data and test data in the training data are extracted, low-dimensional flow patterns of the normal operation data and the test data are extracted, the fault features are reconstructed, reconstruction errors are calculated, distribution of the low-dimensional flow patterns and distribution of the reconstruction errors are fitted respectively, and fault detection indexes are calculated; comparing whether the fault detection index value exceeds a threshold value or not, and judging whether the water chilling unit breaks down or not. The method fully considers the influence of the working condition on the fault features, can extract the complete fault features from the operation data of the water chilling unit and eliminate the working condition features, and improves the fault detection performance of the model in an unknown variable working condition scene.

Description

Technical Field

[0001] The present invention relates to the technical field of HVAC fault detection, and in particular to a chiller fault detection method and system based on separation of operating conditions and fault characteristics. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] The chiller is the core component of the HVAC system and is responsible for regulating the indoor state of the air supply room. However, due to its complex structure, changeable operating environment, and human errors that may occur during maintenance, the chiller will inevitably have various faults. Chillers that are not repaired in time will cause the performance of the HVAC system to decline, affecting indoor comfort, work efficiency and quality of life. In addition, faulty chillers will increase energy consumption and bring economic pressure. If the chiller is in a faulty state for a long time, it may also cause equipment damage and production interruption. Therefore, it is very important to monitor the chiller in real time and take effective measures to maintain it in time. Fault detection technology can quickly detect the fault of the chiller, reduce losses, save resources, and improve system reliability. Therefore, studying the fault detection technology of the chiller is of great practical significance.

[0004] The working condition of the chiller is affected by many factors during its operation. First, the temperature setting value set by the user is one of the important factors affecting the operation of the chiller. According to the user's needs, the chiller needs to adjust the indoor air conditioning system according to the set temperature to ensure that the indoor temperature is within the ideal range; secondly, the indoor occupancy will also affect the working environment of the chiller. When the number of people in the room increases, the temperature and humidity of the indoor air will increase, thereby increasing the workload of the chiller; in addition, changes in the external environment will also affect the operation of the chiller. Due to the above reasons, the actual operating conditions and actual working status of the chiller are constantly changing and cannot be predicted in advance. Although fault detection technology has been widely used in chiller systems, the following problems still exist: (1) The implicit features in the normal operating data and fault operating data of the chiller can be decomposed into operating condition features and fault features. Existing fault detection methods rarely consider the impact of the chiller's operating conditions on fault features, resulting in poor fault detection results when the operating conditions are unknown. (2) Due to the frequent changes in the environmental conditions and temperature setting values ​​of the chiller, the chiller often operates in unknown conditions. Most existing methods are based on models trained based on known conditions, and the fault detection effect is poor when facing unknown conditions. (3) The fault feature dimension is high, and existing methods rarely reduce its dimension, which is prone to the "curse of dimensionality". Autoencoders are a common dimensionality reduction method, but when trained only with data reconstruction loss, the resulting flow pattern often does not conform to the normal distribution, resulting in inaccurate modeling. The accuracy of data reconstruction also needs to be improved, which affects the detection effect. Summary of the invention

[0005] In order to solve the above problems, the present invention proposes a chiller fault detection method and system based on the separation of operating conditions and fault characteristics, which fully considers the influence of operating conditions on fault characteristics, can extract complete fault characteristics from the chiller operation data and eliminate operating condition characteristics, thereby improving the fault detection performance of the model in unknown variable operating condition scenarios.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides a chiller fault detection method based on separation of operating conditions and fault characteristics, comprising the following steps:

[0008] Collect the operating data of the chiller and divide it into training data and test data;

[0009] Extracting the working condition characteristics and fault characteristics of the training data, decoding the working condition characteristics and fault characteristics into working condition modes and fault modes respectively, reconstructing the operating data using the working condition modes and fault modes, defining a loss function based on the reconstructed operating data and eliminating loss terms, and training the feature extraction model;

[0010] The trained feature extraction model is used to extract the fault features of the normal operation data and the test data in the training data, and the second fault feature and the first fault feature are obtained respectively. The low-dimensional flow types of the first and second fault features are extracted, and the first fault feature and the second fault feature are reconstructed based on the low-dimensional flow type, and the reconstruction error is calculated. The distribution of the low-dimensional flow type of the second fault feature and the distribution of the reconstruction error of the second fault feature are fitted, and the reconstruction score of each sample point of the first fault feature and its adjacent points is calculated and smoothed, and the fault detection index is calculated based on the above parameters;

[0011] Compare whether the fault detection index value exceeds the threshold value to determine whether the chiller fails.

[0012] As an optional implementation, the operation data of the chiller includes normal operation data and fault operation data.

[0013] As an optional implementation, the low-dimensional flow patterns of the first and second fault features are extracted, specifically:

[0014] The first fault feature and the second fault feature are respectively cut into feature sequence pieces of different lengths, and different feature sequence pieces are generated by multiple sub-manifold generators. The multiple sub-manifold generators are integrated with different weights to obtain low-dimensional manifold data. The low-dimensional manifold data is distinguished from the low-dimensional manifold data sampled from the prior distribution by a manifold discriminator. The low-dimensional manifold of the second fault feature is used for training, and the loss function is optimized to obtain the final low-dimensional manifolds of the first and second fault features.

[0015] As an optional implementation, the first fault feature and the second fault feature are reconstructed based on the low-dimensional flow pattern, specifically:

[0016] The reconstruction generator is used to reconstruct the first fault feature and the second fault feature, the reconstruction discriminator is used to distinguish the reconstructed first fault feature from the original fault feature, and the reconstructed second fault feature from the original fault feature, the reconstructed second fault feature is used for training, the loss function is optimized, and the final reconstructed first fault feature and the final reconstructed second fault feature are obtained.

[0017] As an optional implementation, a loss function is defined based on the reconstructed operation data, and the loss function includes a normal operation data reconstruction loss, a fault operation data reconstruction loss, and an elimination loss.

[0018] As an optional implementation, the fault detection index is calculated, specifically:

[0019] Calculating a first sub-fault index based on the distribution of the low-dimensional flow pattern of the second fault feature and the low-dimensional flow pattern of the first fault feature;

[0020] The reconstruction score of each sample point and its neighboring points of the first fault feature is calculated by using the reconstruction discriminator, and the kernel density estimation method is used to smooth the sample points to obtain the second sub-fault index;

[0021] A third sub-fault indicator is calculated based on the distribution of the reconstruction error of the second fault feature and the reconstruction error of the first fault feature.

[0022] In a second aspect, the present invention provides a chiller fault detection system based on separation of operating conditions and fault characteristics, comprising:

[0023] The data collection module is configured to: collect the operation data of the chiller and divide it into training data and test data;

[0024] The joint training module is configured to: extract the working condition characteristics and fault characteristics of the training data, decode the working condition characteristics and the fault characteristics into working condition modes and fault modes respectively, reconstruct the operation data using the working condition modes and the fault modes, define a loss function based on the reconstructed operation data and the elimination of loss terms, and train the feature extraction model;

[0025] The fault detection index calculation module is configured to: use the trained feature extraction model to extract the fault features of the normal operation data and the test data in the training data, obtain the second fault feature and the first fault feature respectively, extract the low-dimensional flow types of the first and second fault features, reconstruct the first fault feature and the second fault feature based on the low-dimensional flow type, calculate the reconstruction error, and fit the distribution of the low-dimensional flow type of the second fault feature and the distribution of the reconstruction error of the second fault feature, calculate the reconstruction score of each sample point of the first fault feature and the adjacent points thereof and perform smoothing on the reconstruction score, and calculate the fault detection index based on the above parameters;

[0026] The fault detection module is configured to: compare whether the fault detection index value exceeds a threshold value, and determine whether a fault occurs in the chiller.

[0027] In a third aspect, the present invention provides an electronic device comprising a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method described in the first aspect is performed.

[0028] In a fourth aspect, the present invention provides a computer-readable storage medium for storing computer instructions, wherein when the computer instructions are executed by a processor, the method described in the first aspect is performed.

[0029] In a fifth aspect, the present invention provides a computer program product, comprising a computer program, which, when executed by a processor, implements the method described in the first aspect.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] The present disclosure proposes a chiller fault detection method and system based on the separation of working conditions and fault characteristics. In view of the problem that the existing methods do not fully consider the influence of working conditions on fault characteristics, the present disclosure constructs a working condition autoencoder composed of a working condition encoder and a working condition decoder. The working condition autoencoder is used to capture working condition characteristics, deeply mine the working condition information in the operation of the chiller, extract working condition characteristics and working condition modes, and then combine the working condition mode with the fault mode to reconstruct the input data. Such a design fully considers the influence of the operating conditions of the chiller on the fault characteristics, and lays a reliable foundation for the fault autoencoder to eliminate the influence of working conditions, extract fault characteristics and subsequent fault detection.

[0032] The present disclosure proposes a chiller fault detection method and system based on the separation of working conditions and fault features. In order to solve the problem that the fault detection performance of the chiller fault detection model trained under known working conditions is poor in unknown variable working condition scenarios, the present disclosure constructs a fault autoencoder and designs an elimination loss term. The fault features are specifically extracted through the fault autoencoder. Thanks to the elimination loss function introduced during the training process, the fault feature extraction module can eliminate the influence of working conditions as much as possible when extracting fault features, effectively improving the adaptability to different working conditions. Such a design enables the fault autoencoder to eliminate the working condition mode as much as possible under the drive of the working condition autoencoder, as the total loss function including the reconstruction loss function and the elimination loss term is minimized, thereby focusing on the extraction of fault features and improving the fault detection performance of the model in unknown variable working condition scenarios.

[0033] The present disclosure proposes a chiller fault detection method and system based on the separation of working conditions and fault features. The fault feature dimensionality of the chiller is high. After the traditional autoencoder is used for dimensionality reduction, the resulting low-dimensional flow pattern usually does not obey the normal distribution and the data reconstruction is not accurate enough, which affects the modeling accuracy and fault detection effect. The present disclosure constructs an adversarial autoencoder fault detection module. Through the confrontation between the flow pattern discriminator and the flow pattern generator, the low-dimensional flow pattern is approximated to the normal distribution, which improves the accuracy of modeling. Through the confrontation between the reconstruction discriminator and the reconstruction generator, the reconstruction generator generates accurate reconstructed data, which improves the fault detection effect.

[0034] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0036] Figure 1 A flow chart of a chiller fault detection method based on separation of operating conditions and fault characteristics provided in Example 1 of the present invention;

[0037] Figure 2 A network structure diagram of the improved dense block and channel attention component provided in Example 1 of the present invention;

[0038] Figure 3 The network structure diagram of the sub-flow pattern generator and variational attention mechanism components provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0039] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0040] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0041] It should be noted that the terms used herein are only for describing specific embodiments, and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that the terms "include" and "have" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0042] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0043] Example 1

[0044] like Figure 1 As shown, this embodiment provides a chiller fault detection method based on separation of operating conditions and fault characteristics, comprising the following steps:

[0045] S1 collects the operating data of the chiller and divides it into training data and test data;

[0046] S2 extracts the working condition characteristics and fault characteristics of the training data, decodes the working condition characteristics and fault characteristics into working condition modes and fault modes respectively, reconstructs the operating data using the working condition modes and fault modes, defines a loss function based on the reconstructed operating data and the elimination of loss terms, and trains the feature extraction model;

[0047] S3 uses the trained feature extraction model to extract the fault features of the normal operation data and the test data in the training data, obtains the second fault feature and the first fault feature respectively, extracts the low-dimensional flow types of the first and second fault features, reconstructs the first fault feature and the second fault feature based on the low-dimensional flow type, calculates the reconstruction error, and fits the distribution of the low-dimensional flow type of the second fault feature and the distribution of the reconstruction error of the second fault feature, calculates the reconstruction score of each sample point of the first fault feature and its adjacent points and performs smoothing on it, and calculates the fault detection index based on the above parameters;

[0048] S4 compares whether the fault detection index value exceeds a threshold value, and determines whether a fault occurs in the chiller.

[0049] The collected operation data of the chiller includes normal operation data and fault operation data. The operation condition features of the operation data are extracted by the operation condition autoencoder, and the fault features of the operation data are extracted by the fault autoencoder.

[0050] In the process of detecting the operation of the chiller, due to the complex and changeable working conditions, different environmental conditions and operating parameters will affect the extraction of fault features. Therefore, the present disclosure adopts a working condition feature extraction module to specifically extract working condition features and working condition modes, fully considering the impact of the chiller's operating conditions on the fault features, laying a reliable foundation for the fault autoencoder to eliminate the impact of working conditions and extract fault features, thereby improving the accuracy of subsequent fault detection.

[0051] The present disclosure names the encoder and decoder of the working condition feature autoencoder as working condition encoder and working condition decoder respectively. For the operating data x of the chiller under specific working conditions, the working condition feature of the input data is first extracted by the working condition encoder. Then the working condition characteristic v w Mapped to the output layer through the working condition decoder, the working condition mode is obtained Working mode will be related to the failure mode In the joint training phase, they jointly participate in the reconstruction of the input data, define the loss function based on the reconstructed operating data and the elimination of loss terms, and train the model to extract pure fault features that eliminate the influence of working conditions.

[0052] The operating condition encoder uses an improved residual network with a channel spatial attention mechanism as the basic structure, and the operating condition decoder adopts the mirror structure of the operating condition encoder. For the input normal operation data and fault operation data of the chiller, they are first converted into a 32×32 matrix and then input into the improved residual network. Specifically, the input high-dimensional data is first convolved, and the feature extraction ability of the data is enhanced through the channel attention component and spatial attention component modules. Then, it is further processed through the improved single-layer convolution residual block and ELU activation function to reduce the model complexity and calculation amount. Subsequently, the data is further feature extracted and optimized through the convolution layer and two attention components. Finally, the normal data operating condition features and fault data operating condition features are extracted, and reconstructed into normal data operating mode and fault data operating mode respectively through the operating condition decoder.

[0053] The working condition encoder uses an improved residual network as its basic structure. Compared with the residual network structure before the improvement, it only retains one 3×3 convolution, one batch normalization layer and one activation layer. This simplification not only reduces the complexity of the model, but also significantly reduces the amount of calculation and training time. In addition, the original ReLU activation function of the activation layer is replaced by the ELU activation function with better performance. Through this optimization, the residual block is not only improved in computational efficiency, but also shows stronger feature extraction capabilities when processing complex working condition data, which provides a guarantee for the efficient operation of the working condition encoder and decoder.

[0054] The working condition feature extraction module of the present disclosure incorporates an attention mechanism component, specifically including a channel attention component and a spatial attention component:

[0055] ① Channel attention mechanism component: The channel attention mechanism extracts global features of each channel through global average pooling and global maximum pooling to enhance the extraction of working condition features. First, the input feature map is subjected to global average pooling and global maximum pooling operations, and the results are input into a shared multi-layer perceptron (MLP). After that, the two feature maps are added element by element, and the attention weights are obtained by normalization. Finally, these weights are multiplied with the original feature map to enhance important features and suppress irrelevant features, thereby improving the accuracy of working condition feature extraction.

[0056] ② Spatial attention mechanism component: The spatial attention mechanism extracts the importance of each spatial position through convolution operations to ensure that the spatial distribution information of the feature map is fully utilized. Specifically, first, the feature map output by the channel attention mechanism component is used as input, and the maximum pooling and average pooling operations are performed. After that, the concatenation is input into the convolution layer, and the result is normalized to obtain the spatial attention weight. Finally, these weights are multiplied with the input feature map, so as to fully utilize the spatial distribution information of the feature map.

[0057] The relevant parameters of the operating condition feature extraction module are updated only by using normal operation data through formula (1) and are jointly trained with the fault mode extraction module in the offline modeling stage.

[0058]

[0059] Among them, n Nor represents the number of normal samples, x Nor,i represents the i-th normal sample, Represents the reconstruction result of the i-th normal sample.

[0060] Fault features are often affected by constantly changing working conditions, making fault detection complicated and difficult. Traditional fault detection algorithms are usually difficult to effectively eliminate these interference factors when faced with such a complex environment, resulting in poor diagnostic accuracy. Therefore, the present disclosure constructs a fault feature extraction module, which can eliminate the influence of working conditions on fault features and accurately extract pure fault features that are unrelated to working conditions from input data under the drive of a working condition autoencoder.

[0061] The fault autoencoder consists of a fault encoder and a fault decoder. Different from the operating mode extraction module, this module focuses on the extraction of fault features. The fault encoder extracts the fault features v from the data. f , which is then mapped to a fault mode by a fault decoder The extracted fault features v f will be used as the input to the fault detection module, and the fault mode In the joint training phase, Together they participate in the reconstruction of the input data, thereby training this module to purify fault features and eliminate operating interference.

[0062] The fault encoder uses an improved dense network structure and introduces an improved channel attention component to the dense block, so that the fault encoder can pay more attention to the features that are more important for fault detection. The fault decoder adopts the mirror structure of the fault encoder. For the input preprocessed normal operation data and fault operation data of the chiller, this module first uses the convolution layer to perform preliminary feature extraction, and then passes through the combination structure of three improved dense block structures and two transition layers in sequence. Finally, the fault features of normal data and fault data are obtained through the pooling layer, and reconstructed into the fault mode of normal data and the fault mode of fault data respectively through the fault decoder.

[0063] This module updates parameters through the loss function (2) and the elimination of loss terms (3), ensuring that pure fault features that are not affected by working conditions are extracted, providing accurate and reliable data support for subsequent fault detection.

[0064]

[0065]

[0066] Among them, n Fau represents the number of fault samples, x Fau,i represents the i-th fault sample, represents the reconstruction result of the i-th fault sample, represents the output of the fault encoder for the i-th normal sample, represents the failure mode of the i-th normal sample.

[0067] The principle of the improved channel attention component in this module is as follows: First, the input is normalized through the batch normalization layer. Then, average pooling and maximum pooling are applied to compress the feature matrix in the spatial dimension and fuse the spatial information. The feature matrix is ​​then summed and merged elements through a shared network to further compress the spatial dimension and generate a channel attention matrix. In order to retain spatial information, a shared convolutional layer is used instead of a shared fully connected layer, thereby reducing model parameters and improving computational efficiency. Finally, referring to the squeeze incentive mechanism, the channel attention matrix is ​​multiplied channel by channel with the original input features to generate features with channel attention. The network structure of the improved dense block and channel attention component is as follows: Figure 2 As shown:

[0068] This module needs to be trained jointly with the working condition feature extraction module. Specifically:

[0069] ① In the joint training stage, the working condition encoder is trained from the normal data x Nor With fault data x Fau Separate the working condition characteristics of normal data and operating characteristics of fault data And reconstruct them into normal data mode respectively Fault data mode At the same time, the fault encoder separates the fault characteristics of normal data Fault characteristics with fault data And reconstruct them into normal data failure modes respectively Failure modes with failure data After that, normal data x Nor pass Refactored to Fault data x Fau pass Refactored to The joint training is performed through the total loss function (4), which includes the normal operation data reconstruction loss, the fault operation data reconstruction loss and the elimination loss:

[0070] L = ζ(LR_F +L R_N )+ξL E (4)

[0071] In the formula, ζ and ξ are two hyperparameters in the overall optimization object. By minimizing L R_F With L R_N , both fault data and normal data can be well reconstructed; by minimizing the loss term L E ,The fault encoder can extract as few working condition features as possible during the training process. In the offline modeling stage, the adaptive moment estimation algorithm is used to optimize the total optimization object and update the parameters. When the training reaches the set maximum number of iterations, the training is terminated and the joint training stage is completed.

[0072] ②After training, the fault autoencoder can extract pure fault features from the input chiller operation data and use them as the input of the adversarial autoencoder fault detection module.

[0073] The fault feature dimensionality of the chiller is relatively high. The features after dimensionality reduction using traditional autoencoders usually do not obey the normal distribution and the reconstructed data is not accurate enough, which affects the modeling accuracy and detection effect. To solve this problem, the present invention continues to construct an adversarial autoencoder fault detection module. Through the confrontation between the flow pattern discriminator and the flow pattern generator, the flow pattern is made to approach the normal distribution, which improves the modeling accuracy; through the confrontation between the reconstruction discriminator and the reconstruction generator, the reconstruction generator generates accurate reconstruction data, which improves the detection effect of the model.

[0074] This module consists of two pairs of generative adversarial structures. The first pair of generative adversarial network structures consists of an integrated flow pattern generator and a flow pattern discriminator. Through the supervision of the flow pattern discriminator during the adversarial training of the two, it is ensured that the low-dimensional flow pattern extracted by the flow pattern generator is close to the normal distribution. The second pair of generative adversarial network structures consists of a reconstruction generator and a reconstruction discriminator. Through the supervision of the reconstruction discriminator during the adversarial training of the two, it is ensured that the model can accurately reconstruct the fault characteristics of normal working conditions.

[0075] The first pair of generative adversarial networks is used to extract the low-dimensional manifold of fault features, specifically:

[0076] The fault features are cut into feature sequence pieces of different lengths, and different feature sequence pieces are generated using multiple sub-manifold generators. The multiple sub-manifold generators are integrated with different weights to obtain low-dimensional manifold data. The low-dimensional manifold data is distinguished from the low-dimensional manifold data sampled from the prior distribution (normal distribution) using a manifold discriminator, and the loss function is optimized to obtain the low-dimensional manifold of the fault features.

[0077] The integrated flow pattern generator consists of multiple sub-flow pattern generators, each of which adopts an improved Transformer encoder structure to generate fake feature sequence slices. Specifically, the improved Transformer encoder structure consists of multiple stacked encoding layers, and each layer of the encoder contains a variational attention mechanism component, layer normalization, and residual connection components. The variational attention component generates the query vector q and the key vector k through feature extraction, linear transformation, and nonlinear activation function of the input data, and uses the Dirichlet distribution parameters to sample and regularize the attention weights. By minimizing the KL divergence, the approximate posterior distribution is close to the true posterior distribution, achieving sparse regularization and improving the performance and interpretability of the model. The network structure of the sub-flow pattern generator and the variational attention component is shown in the figure. Figure 3 shown.

[0078] The present disclosure uses one-dimensional deconvolution as the framework of the reconstruction generator, and expands the dimension by using an interpolation method. and They are the input features and convolution kernel of the lth deconvolution layer respectively. First, l-1 The method of inserting s-1 zeros between every two adjacent elements in the first dimension will l-1 Expand to After that in v' l-1 By using the kernel as κ l , stride 1 and "valid" fill mode for v' l-1 By performing the convolution operation, we can get v l .

[0079] This module is divided into the offline modeling stage, the online test preparation stage, and the online test stage. In each round of training in the offline modeling stage, the first set of adversarial learning strategies is used for the low-dimensional flow pattern. First, the features input into this module by the fault feature extraction module are cut into feature sequence slices of different lengths, and then different feature sequence slices are generated using multiple sub-flow pattern generators. Next, multiple flow pattern generators are integrated with different weights to obtain the final low-dimensional flow pattern data. The weight is calculated as follows:

[0080]

[0081] In the formula, is the binary cross entropy loss corresponding to the h-th sub-manifold generator in the β-1th round of training, is the hth sub-manifold generator, q is the number of sub-manifold generators, is the weight corresponding to the hth sub-manifold generator in the βth round.

[0082] Finally, the flow discriminator distinguishes the data generated by the integrated flow generator from the data from the prior distribution (normal distribution). This set of adversarial learning strategies is trained by the following formula:

[0083]

[0084] Where D manifold is the flow type discriminator, represents the true fault feature corresponding to the hth sub-flow generator in the βth iteration, A is the total number of iterations, m (β) is the manifold sampled from the prior distribution (normal distribution) in the βth iteration.

[0085] The second pair of generative adversarial networks is used to reconstruct the fault features based on the low-dimensional flow pattern, specifically:

[0086] The reconstruction generator is used to reconstruct the fault features with the low-dimensional manifold as input, and the reconstruction discriminator is used to distinguish the reconstructed fault features from the real fault features.

[0087] In each round of training, the second set of adversarial learning strategies is used for the reconstruction features of the chiller, and the output of the reconstruction discriminator is used to construct the sub-fault index, which is obtained by the output of the trained reconstruction discriminator in the online detection phase. During training, the reconstruction discriminator is first trained to distinguish whether the signal it receives is from the reconstruction generator or the original fault feature signal, and then the reconstruction generator is trained to deceive the reconstruction discriminator as much as possible. This adversarial strategy is trained using the following loss function:

[0088]

[0089] Among them, D recon To reconstruct the discriminator, G recon For the refactoring generator. The low-dimensional manifold representing the fault characteristics of normal data in the training data, D recon To reconstruct the discriminator, G recon For the refactoring generator.

[0090] At the end of each round of training, the integrated flow generator and reconstruction generator are trained with reconstruction loss, and the loss function is as follows:

[0091]

[0092] By doing this, the information loss during training is minimized.

[0093] In the online test preparation stage, the fault features of the chiller under normal conditions extracted by the fault feature extraction module are first input, and the low-dimensional flow patterns are calculated through the model established in the offline modeling stage. and reconstruction error The distribution of low-dimensional flow patterns and the distribution of reconstruction errors are fitted respectively.

[0094] In the online detection phase, the chiller is judged to be faulty by calculating whether the index value of the unknown state data exceeds the limit. The fault detection index is calculated as follows:

[0095] Based on the distribution of the low-dimensional flow type of the second fault feature and the low-dimensional flow type of the first fault feature, the first sub-fault index P1 is calculated; the reconstruction score of each sample point of the first fault feature and its nearby adjacent points is calculated using the reconstruction discriminator, and the kernel density estimation method is used to smooth it to obtain the second sub-fault index P2; based on the distribution of the reconstruction error of the second fault feature and the reconstruction error of the first fault feature, the third sub-fault index P3 is calculated.

[0096] Taking the fault feature v0 of the unknown state data of the chiller proposed by the fault feature extraction module as an example, firstly, the low-dimensional flow pattern z0 and reconstruction error e0 of v0 are further calculated through the previously established model. For the reconstruction discriminator, the reconstruction scores of it and its neighboring points are collected, and then the kernel density estimation method is used to smooth it to obtain the sub-fault index P2:

[0097]

[0098] Where R(v0) is the reconstruction score of the output of the reconstruction discriminator at v0, t is the number of selected neighboring points, c is the bandwidth parameter, K(·) is the Gaussian kernel function, To reconstruct the discriminator at the neighboring points of v0 The reconstruction score output at .

[0099] Continue to calculate sub-fault indicators And the total fault index. If the calculated total fault index P is less than the chiller threshold, it is determined that the chiller system has a fault; otherwise, the chiller operates normally.

[0100] Example 2

[0101] This embodiment provides a chiller fault detection system based on separation of operating conditions and fault characteristics, including:

[0102] The data collection module is configured to: collect the operation data of the chiller and divide it into training data and test data;

[0103] The joint training module is configured to: extract the working condition characteristics and fault characteristics of the training data, decode the working condition characteristics and the fault characteristics into working condition modes and fault modes respectively, reconstruct the operation data using the working condition modes and the fault modes, define a loss function based on the reconstructed operation data and the elimination of loss terms, and train the feature extraction model;

[0104] The fault detection index calculation module is configured to: use the trained feature extraction model to extract the fault features of the normal operation data and the test data in the training data, obtain the second fault feature and the first fault feature respectively, extract the low-dimensional flow types of the first and second fault features, reconstruct the first fault feature and the second fault feature based on the low-dimensional flow type, calculate the reconstruction error, and fit the distribution of the low-dimensional flow type of the second fault feature and the distribution of the reconstruction error of the second fault feature, calculate the reconstruction score of each sample point of the first fault feature and the adjacent points thereof and perform smoothing on the reconstruction score, and calculate the fault detection index based on the above parameters;

[0105] The fault detection module is configured to: compare whether the fault detection index value exceeds a threshold value, and determine whether a fault occurs in the chiller.

[0106] Among them, the joint training module includes the training of the working condition feature extraction module and the fault feature extraction module, and the fault detection index calculation module and the fault detection module both belong to the adversarial autoencoding fault detection module.

[0107] It should be noted that the above modules correspond to the steps described in Example 1, and the examples and application scenarios implemented by the above modules and the corresponding steps are the same, but are not limited to the contents disclosed in the above Example 1. It should be noted that the above modules, as part of the system, can be executed in a computer system such as a set of computer executable instructions.

[0108] In further embodiments, there is also provided:

[0109] An electronic device includes a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method described in Embodiment 1 is performed. For the sake of brevity, it will not be described in detail here.

[0110] It should be understood that in this embodiment, the processor may be a central processing unit CPU, and the processor may also be other general-purpose processors, digital signal processors DSP, application-specific integrated circuits ASIC, off-the-shelf programmable gate arrays FPGA or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0111] The memory may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

[0112] A computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, the method described in Example 1 is completed.

[0113] The method in Example 1 can be directly embodied as a hardware processor, or a combination of hardware and software modules in the processor. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware. To avoid repetition, it is not described in detail here.

[0114] A computer program product includes a computer program, and when the computer program is executed by a processor, the method described in embodiment 1 is implemented.

[0115] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer executable instructions, such as instructions included in a program module, which are executed in a device on a real or virtual processor of the target to perform the process / method as described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc. that perform specific tasks or implement specific abstract data types. In various embodiments, the functions of program modules can be combined or divided between program modules as needed. Machine executable instructions for program modules can be executed in local or distributed devices. In distributed devices, program modules can be located in local and remote storage media.

[0116] The computer program code for implementing the method of the present invention can be written in one or more programming languages. These computer program codes can be provided to the processor of a general-purpose computer, a special-purpose computer or other programmable data processing device, so that the program code, when executed by the computer or other programmable data processing device, causes the function / operation specified in the flow chart and / or block diagram to be implemented. The program code can be executed completely on a computer, partially on a computer, as an independent software package, partially on a computer and partially on a remote computer or completely on a remote computer or server.

[0117] In the context of the present invention, computer program codes or related data may be carried by any appropriate carrier to enable a device, apparatus or processor to perform the various processes and operations described above. Examples of carriers include signals, computer readable media, and the like. Examples of signals may include electrical, optical, radio, acoustic or other forms of propagation signals, such as carrier waves, infrared signals, and the like.

[0118] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0119] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A chiller fault detection method based on separation of operating conditions and fault characteristics, characterized in that: The following steps are involved: Collect the operating data of the chiller and divide it into training data and test data; Extracting the working condition characteristics and fault characteristics of the training data, decoding the working condition characteristics and fault characteristics into working condition modes and fault modes respectively, reconstructing the operating data using the working condition modes and fault modes, defining a loss function based on the reconstructed operating data and eliminating loss terms, and training the feature extraction model; The trained feature extraction model is used to extract the fault features of the normal operation data and the test data in the training data, and the second fault feature and the first fault feature are obtained respectively. The low-dimensional flow types of the first and second fault features are extracted, and the first fault feature and the second fault feature are reconstructed based on the low-dimensional flow type, and the reconstruction error is calculated. The distribution of the low-dimensional flow type of the second fault feature and the distribution of the reconstruction error of the second fault feature are fitted, and the reconstruction score of each sample point of the first fault feature and its adjacent points is calculated and smoothed, and the fault detection index is calculated based on the above parameters; Compare whether the fault detection index value exceeds the threshold value to determine whether the chiller fails.

2. The chiller fault detection method based on separation of operating conditions and fault characteristics according to claim 1, characterized in that: The operating data of the chiller includes normal operating data and fault operating data.

3. The chiller fault detection method based on separation of operating conditions and fault characteristics according to claim 1, characterized in that: Extract the low-dimensional manifold of the first and second fault features, specifically: The first fault feature and the second fault feature are respectively cut into feature sequence pieces of different lengths, and different feature sequence pieces are generated by multiple sub-manifold generators. The multiple sub-manifold generators are integrated with different weights to obtain low-dimensional manifold data. The low-dimensional manifold data is distinguished from the low-dimensional manifold data sampled from the prior distribution by a manifold discriminator. The low-dimensional manifold of the second fault feature is used for training, and the loss function is optimized to obtain the final low-dimensional manifolds of the first and second fault features.

4. The chiller fault detection method based on separation of operating conditions and fault characteristics according to claim 1, characterized in that: The first fault feature and the second fault feature are reconstructed based on the low-dimensional flow pattern, specifically: The reconstruction generator is used to reconstruct the first fault feature and the second fault feature, the reconstruction discriminator is used to distinguish the reconstructed first fault feature from the original fault feature, and the reconstructed second fault feature from the original fault feature, the reconstructed second fault feature is used for training, the loss function is optimized, and the final reconstructed first fault feature and the final reconstructed second fault feature are obtained.

5. The chiller fault detection method based on separation of operating conditions and fault characteristics according to claim 1, characterized in that: A loss function is defined based on the reconstructed operation data and the elimination loss term, and the loss function includes normal operation data reconstruction loss, fault operation data reconstruction loss and elimination loss.

6. The chiller fault detection method based on separation of operating conditions and fault characteristics according to claim 4, characterized in that: Calculate the fault detection index, specifically: Calculating a first sub-fault index based on the distribution of the low-dimensional flow pattern of the second fault feature and the low-dimensional flow pattern of the first fault feature; The reconstruction score of each sample point and its neighboring points of the first fault feature is calculated by using the reconstruction discriminator, and the kernel density estimation method is used to smooth the sample points to obtain the second sub-fault index; A third sub-fault indicator is calculated based on the distribution of the reconstruction error of the second fault feature and the reconstruction error of the first fault feature.

7. A chiller fault detection system based on separation of operating conditions and fault characteristics, characterized in that: include: The data collection module is configured to: collect the operation data of the chiller and divide it into training data and test data; The joint training module is configured to: extract the working condition characteristics and fault characteristics of the training data, decode the working condition characteristics and the fault characteristics into working condition modes and fault modes respectively, reconstruct the operation data using the working condition modes and the fault modes, define a loss function based on the reconstructed operation data and the elimination of loss terms, and train the feature extraction model; The fault detection index calculation module is configured to: use the trained feature extraction model to extract the fault features of the normal operation data and the test data in the training data, obtain the second fault feature and the first fault feature respectively, extract the low-dimensional flow types of the first and second fault features, reconstruct the first fault feature and the second fault feature based on the low-dimensional flow type, calculate the reconstruction error, and fit the distribution of the low-dimensional flow type of the second fault feature and the distribution of the reconstruction error of the second fault feature, calculate the reconstruction score of each sample point of the first fault feature and the adjacent points thereof and perform smoothing on the reconstruction score, and calculate the fault detection index based on the above parameters; The fault detection module is configured to: compare whether the fault detection index value exceeds a threshold value, and determine whether a fault occurs in the chiller.

8. An electronic device, characterized in that: The method comprises a memory and a processor and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method according to any one of claims 1 to 6 is completed.

9. A computer-readable storage medium, characterized in that: Used to store computer instructions, which, when executed by a processor, complete the method described in any one of claims 1 to 6.

10. A computer program product, characterized in that The invention comprises a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 6.

Images