Water chiller fault detection method and system based on data image difference

Abstract

The present disclosure relates to the technical field of fault diagnosis, and provides a chiller fault detection method and system based on data image difference, which comprises: performing image conversion on the obtained normal working condition data set and online test data set respectively, performing nonlinear transformation on the image using a symmetric stack auto-encoding network, and mapping the obtained gray images to a high-dimensional feature space respectively; for the online test gray image and the normal working condition gray image that have been nonlinearly transformed to the high-dimensional space, extracting slow feature information using slow feature analysis and subtracting the slow feature information, expanding the obtained slow feature information difference image, calculating a monitoring statistic according to the sub-image obtained after expansion, comparing the monitoring statistic with a set fault detection statistic threshold, and determining whether a fault has occurred. The method can mine the spatial correlation and nonlinear characteristics of system operation data, eliminate the time-varying dynamic characteristics of the data, and improve the fault detection effect.

Description

Technical Field

[0001] This disclosure relates to the technical field of fault diagnosis, specifically to a method and system for detecting faults in chiller units based on differences in data images. Background Technology

[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.

[0003] Chillers are crucial components and major energy-consuming devices in HVAC systems. Their complex structure and variable operating environment inevitably lead to various malfunctions. If a malfunctioning chiller system is not maintained promptly, its control strategies will fail, impacting the HVAC system's efficiency and resulting in significant energy waste. Therefore, monitoring the operating status of chillers is essential. Applying fault detection technology to chillers can quickly identify system faults and minimize their impact. Thus, researching chiller fault diagnosis is of paramount importance for ensuring the normal operation of HVAC systems and reducing system energy consumption.

[0004] Although fault detection technology has been widely used in chiller systems, fault detection in chillers still suffers from complex fault phenomena, difficulty in obtaining fault data, and multiple causes of faults, resulting in relatively low accuracy in directly applying existing fault detection methods to identify faults.

[0005] The inventors found in their research that the existing methods for detecting chiller unit faults mainly have the following problems: (1) When the operating data of the chiller unit is measured and collected by multiple sensors at different locations in space, there is spatial correlation between different process variables; the existing fault detection methods directly build models using the operating data of the chiller unit, and cannot utilize the spatial correlation feature information between various process variables; (2) Due to the frequent changes in operating conditions and set temperature, the chiller unit exhibits significant nonlinear characteristics during operation, while most existing methods are linear methods, which cannot effectively handle the nonlinear characteristics of the chiller unit's operating data and cannot extract more useful feature information from high-dimensional nonlinear operating data; (3) The chiller unit has time-varying dynamic characteristics within a batch during operation, and the existing dynamic monitoring methods cannot effectively capture its time-varying dynamic characteristics within a batch, let alone effectively highlight fault information, resulting in a low fault detection rate. Summary of the Invention

[0006] To address the aforementioned issues, this disclosure proposes a fault detection method and system for chiller units based on data-image differences. It employs a strategy combining data-image conversion, nonlinear mapping based on a symmetric stacked autoencoder network, and slow feature analysis based on image differences to construct a novel fault detection method for chiller units. This method can uncover the spatial correlation and nonlinear characteristics of system operating data, eliminate the time-varying dynamic characteristics of the data, and improve fault detection performance.

[0007] To achieve the above objectives, the present disclosure adopts the following technical solution:

[0008] One or more embodiments provide a method for detecting faults in chiller units based on differences in data images, including the following steps:

[0009] The acquired normal operating condition dataset and online test dataset were converted into images respectively to obtain the online test grayscale image M0 and the normal operating condition grayscale image M3.

[0010] A symmetric stacked autoencoder network is used to perform nonlinear transformations, and the resulting grayscale images are mapped to high-dimensional feature spaces respectively;

[0011] For the online test grayscale image and the normal operating condition grayscale image after nonlinear transformation to high-dimensional space, slow feature analysis is used to extract slow feature information and subtract them to obtain the slow feature information difference map of online test data and normal operating condition data.

[0012] The obtained slow feature information difference map is expanded, and the monitoring statistics are calculated based on the expanded difference map sub-map. The statistics are compared with the set threshold of the fault detection statistics to determine whether a fault has occurred.

[0013] One or more embodiments provide a chiller unit fault detection system based on data image differences, including:

[0014] Image conversion module: configured to convert the acquired normal operating condition dataset and online test dataset into images respectively, resulting in online test grayscale image M0 and normal operating condition grayscale image M3;

[0015] The nonlinear mapping module based on a stacked autoencoder network is configured to perform nonlinear transformations using a symmetric stacked autoencoder network, mapping the resulting grayscale images to high-dimensional feature spaces.

[0016] The feature extraction module based on slow feature analysis is configured to use slow feature analysis to extract slow feature information from online test grayscale images and normal operating condition grayscale images that have been transformed to high-dimensional space through nonlinear transformation. The slow feature information is then subtracted to obtain a difference map of slow feature information between online test data and normal operating condition data.

[0017] Fault detection module: It is configured to expand the obtained slow feature information difference map, calculate the monitoring statistics based on the expanded difference map sub-map, and compare it with the set threshold of the fault detection statistics to determine whether a fault has occurred.

[0018] An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the steps in the above-described chiller unit fault detection method based on data image differences.

[0019] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the steps in the above-described chiller unit fault detection method based on data image differences.

[0020] Compared with the prior art, the beneficial effects of this disclosure are as follows:

[0021] By employing a strategy that combines data-image transformation, nonlinear mapping based on symmetric stacked autoencoder networks, and slow feature analysis based on image differences, the data-image transformation fully explores the neighborhood information and spatial correlation characteristics between process variables. Nonlinear mapping can effectively handle the nonlinearity of chiller unit operating data, and slow feature analysis can eliminate the time-varying dynamic characteristics of the data. The combination of these methods effectively improves the fault detection effect of chiller units.

[0022] The advantages of this disclosure, as well as its additional advantages, will be described in detail in the following specific embodiments. Attached Figure Description

[0023] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute a limitation thereof.

[0024] Figure 1 This is a block diagram of the fault detection system according to Embodiment 2 of this disclosure;

[0025] Figure 2 This is an example of a grayscale image converted from the normal operating condition dataset of Embodiment 1 of this disclosure;

[0026] Figure 3 In Embodiment 1 of this disclosure Figure 1 The diagram illustrates the process of converting a dataset into a two-dimensional grayscale image using the image conversion module.

[0027] Figure 4 In Embodiment 1 of this disclosure Figure 1 A schematic diagram of the method flow for performing nonlinear mapping on grayscale images using the nonlinear mapping module based on stacked autoencoder networks;

[0028] Figure 5 In Embodiment 1 of this disclosure Figure 1 A flowchart illustrating the method for extracting slow feature difference information using the feature extraction module based on slow feature analysis;

[0029] Figure 6 In Embodiment 1 of this disclosure Figure 1 A flowchart illustrating the fault detection process performed by the fault detection module in the system;

[0030] Figure 7 This is an example of using a sliding window to expand the difference map of normal operating conditions in Embodiment 1 of this disclosure;

[0031] Figure 8 This is a flowchart of a chiller unit fault detection method based on data image differences according to Embodiment 1 of this disclosure. Detailed Implementation

[0032] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0033] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of this disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains.

[0034] It should be noted that the terminology used herein is for descriptive purposes only and is not intended to limit the exemplary embodiments according to this disclosure. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. It should be noted that, without conflict, the various embodiments and features within those embodiments can be combined with each other. The embodiments will now be described in detail with reference to the accompanying drawings.

[0035] By analyzing the actual operating status of the chiller unit, it is found that the process variables of the chiller unit are interconnected, and there is a potential spatial interdependence between different process variables. That is, the various process variables of the chiller unit exhibit strong spatial correlation characteristics during actual operation. In addition, during the operation of the chiller unit, changes in indoor load and set parameters, changes in ambient temperature and humidity caused by outdoor weather, and changes in the operating mode of the chiller unit under closed-loop control all lead to significant nonlinear and time-varying dynamic characteristics of the chiller unit's process variables. The chiller unit fault detection method proposed in this disclosure adopts a strategy that combines a data-image conversion method, nonlinear mapping based on a symmetric stacked autoencoder network, and slow feature analysis based on image differences, which improves the accuracy of fault detection. Specific embodiments are described below.

[0036] Example 1

[0037] In one or more of the technical solutions disclosed in the embodiments, such as Figures 1 to 8 As shown, a method for detecting faults in chiller units based on differences in data images includes the following steps:

[0038] Step 1, Image Conversion: The acquired normal operating condition dataset and online test dataset are converted into images respectively to obtain online test grayscale images and normal operating condition grayscale images;

[0039] Step 2, Nonlinear mapping based on symmetric stacked autoencoder network: Nonlinear transformation is performed using a symmetric stacked autoencoder network to map the resulting grayscale images to high-dimensional feature spaces.

[0040] Step 3: Feature extraction based on slow feature analysis: For the online test grayscale image and the normal operating condition grayscale image that have been transformed to a high-dimensional space by nonlinear transformation, slow feature analysis is used to extract slow feature information and subtract them to obtain the slow feature information difference map of online test data and normal operating condition data.

[0041] Step 4, Fault Detection: Expand the obtained slow feature information difference map, calculate the monitoring statistics based on the expanded difference map sub-map, and compare it with the set threshold of the fault detection statistics to determine whether a fault has occurred.

[0042] Steps 1 to 4 above constitute the online detection process for chiller unit faults. The threshold for fault detection statistics is calculated by analyzing the normal operating condition dataset during the offline modeling phase. The threshold for determining the normal operating state of the chiller unit is the threshold for preventing faults from occurring.

[0043] In this embodiment, a strategy combining a data-graph conversion method, a nonlinear mapping based on a symmetric stacked autoencoder network, and a slow feature analysis based on image differences is adopted. The data-graph conversion method fully explores the neighborhood information and spatial correlation characteristics between process variables, the nonlinear mapping method can effectively handle the nonlinearity of chiller unit operating data, and the slow feature analysis method can eliminate the time-varying dynamic characteristics of the data. The combination of the above methods effectively improves the fault detection effect of chiller units.

[0044] A further technical solution involves determining the threshold for fault detection statistics during the offline modeling phase, including the following steps:

[0045] Step S1, Image Conversion: The two batches of normal working condition datasets are converted into images respectively to obtain the corresponding normal working condition grayscale image M1 and normal working condition grayscale image M2.

[0046] In this embodiment, different batches of data refer to the operating data of the chiller unit at different time periods, such as the operating data of different days. In this embodiment, two batches of normal operating condition datasets were acquired during the offline modeling stage and are denoted as X1 and X2, respectively.

[0047] Step S2, Nonlinear mapping based on symmetric stacked autoencoder network: Nonlinear transformation is performed using a symmetric stacked autoencoder network to map the resulting grayscale images to high-dimensional feature spaces.

[0048] Step S3: Calculation of feature difference based on slow feature analysis: For two normal operating condition grayscale images that have been transformed into high-dimensional space through nonlinear transformation, slow feature analysis is used to extract slow feature information and subtract them to obtain the slow feature information difference map of the two normal operating condition data.

[0049] Step S4: Expand the obtained slow feature information difference map, and calculate the control limit of the monitoring statistics based on the expanded difference map sub-map, which is used as the threshold of the fault detection statistics.

[0050] In this embodiment, the data processing process of the offline modeling stage and the online testing stage is basically the same. The offline modeling stage provides the online testing stage with monitoring statistical thresholds to determine whether a fault has occurred. The online testing stage identifies the online detection data and determines whether a fault has occurred.

[0051] In steps 1 and S1, the image conversion method performs image conversion on the acquired normal operating condition dataset and online test dataset separately. Optionally, the following process is performed for each dataset:

[0052] Step 11: Preprocess the data in the dataset by converting it into an unsigned integer type. The value of the converted integer data corresponds to the grayscale value of the image.

[0053] Step 12: Convert the preprocessed process data collected by different sensors at the same sampling time into a row in the grayscale image;

[0054] Step 13: Convert the preprocessed process data collected by the same sensor at different sampling times into a column of the grayscale image, thereby converting the normal operating condition dataset and the online test dataset into the corresponding normal operating condition grayscale image and online test grayscale image.

[0055] This embodiment transforms the dataset into an image through data-to-image conversion, which can fully explore the spatial correlation information between various process variables. In other words, it explores the neighborhood information and spatial correlation of each process variable, providing the necessary conditions for improving the accuracy of fault detection.

[0056] Optionally, the process operation data collected by sensors in the normal operating condition dataset and the online test dataset may include operating data such as the supply air temperature, mixing air temperature, return air temperature, and supply fan speed of the chiller unit.

[0057] In step 1, the normal operating condition dataset can be obtained from any normally operating dataset X3 in the historical operating data, the online test dataset is denoted as Y, and the normal operating condition datasets in the offline modeling stage are X1 and X2.

[0058] Before performing the data-to-image conversion, the data needs to be preprocessed. The preprocessing includes standardizing and normalizing the data in the dataset, and multiplying the normalization result by 255.

[0059] The following uses a normal operating condition dataset X1 as an example to illustrate the specific steps involved in obtaining a grayscale image M1 under normal operating conditions through image transformation:

[0060] The normal operating condition dataset X1 can be represented as:

[0061]

[0062] Where t∈[1,2,…,T] represents the time range.

[0063] The value of the p-th process variable in a continuous T time series is represented as:

[0064]

[0065] (1) Data preprocessing. For each process variable x in the normal operating condition dataset X1... p The data corresponding to (t) is Z-score standardized as follows:

[0066]

[0067] Where, mean(x) p (t) represents the process variable x. p The mean of (t), std(x) p (t) represents the process variable x. p The standard deviation of (t).

[0068] The standardized data is defined as:

[0069]

[0070] in:

[0071]

[0072] Standardized dataset Each process variable The corresponding data is then normalized using Min-Max and the output is multiplied by 255, as shown in the following formula.

[0073]

[0074] in, and They are The minimum and maximum values ​​of the vectors are determined, and the normal operating condition dataset after Min-Max normalization is denoted as:

[0075] (2) Convert the preprocessed normal operating condition dataset Z1 into the corresponding grayscale image. Specifically, convert the data from different sensors at the same time in the normal operating condition dataset Z1 into a row in the grayscale image, and convert the process operation data of the same sensor at different sampling times into a column in the grayscale image. Figure 2 This is an example of a grayscale image M1 converted from the normal operating condition dataset Z1.

[0076] The data-to-image conversion is performed by converting the input data into an unsigned integer type, followed by a data format conversion instruction. The resulting integer data value is used as the grayscale value; the larger the value, the darker the grayscale.

[0077] (3) Repeat the above steps to convert the normal operating condition dataset and the online test dataset into grayscale images M1, M2, M3, and M0, as follows: Figure 3 As shown.

[0078] Considering the nonlinearity of chiller unit data, the powerful nonlinear representation capability of a symmetric stacked autoencoder network is utilized to nonlinearly map the obtained normal operating condition grayscale images (M1, M2, and M3) and the online test grayscale image M0 to a high-dimensional feature space, effectively handling the nonlinear characteristics of normal operating condition and online test data. By nonlinearly projecting the grayscale images into the high-dimensional feature space, the nonlinear characteristics of the process data are processed, thereby improving the feature representation capability of the fault detection model.

[0079] In steps 2 and S2, based on the nonlinear mapping of the symmetric stacked autoencoder network, the following process is performed on each obtained grayscale image:

[0080] Step 21: The input grayscale image is nonlinearly mapped through an encoder to obtain the nonlinear features of the grayscale image, thereby fully displaying the nonlinear features of the grayscale image in a high-dimensional space.

[0081] Step 22: Based on the obtained nonlinear features, the original input grayscale image is reconstructed through the decoder, thereby realizing the mapping of the normal working condition grayscale image and the online test grayscale image to a high-dimensional feature space;

[0082] Specifically, in this embodiment, the input layer of the stacked autoencoder network initially acquires sample points from three different batches of normal operating condition grayscale images and one online test grayscale image. First, the encoder nonlinearly maps these grayscale images to an intermediate layer to obtain the nonlinear features of the grayscale images. Then, the intermediate layer data is mapped to the output layer through a decoder to obtain the reconstructed grayscale images, thereby nonlinearly mapping the normal operating condition and online test grayscale images to a high-dimensional feature space.

[0083] The following uses one of the normal operating condition grayscale images, M1, as an example to illustrate the nonlinear mapping process based on a symmetric stacked autoencoder network, which nonlinearly maps M1 to a high-dimensional feature space, as follows:

[0084] (2.1) The input grayscale image is non-linearly mapped through an encoder, and the non-linear mapping is applied to the intermediate layer;

[0085] The encoder has n-1 hidden layers. During the nonlinear mapping process, the output of each encoder layer is:

[0086]

[0087] Where f(.) is a nonlinear activation function, h n This refers to the non-linear features of the grayscale image obtained from the intermediate layer. Here is the encoder weight matrix. Let h be the bias vector of the encoder. Therefore, the stacked autoencoder network maps the input normal operating condition grayscale image M1 through the encoder to the intermediate layer represented as h. n .

[0088] (2.2) The decoder will process the intermediate layer data h n Mapped to the output layer.

[0089] The decoder has n-1 hidden layers, and the weight matrix is... The bias vector is The output of each layer of the decoder is The input grayscale image for reconstruction;

[0090] The decoder decoding process is as follows:

[0091]

[0092] The grayscale image obtained by nonlinearly mapping the normal operating condition grayscale image M1 to the high-dimensional feature space through the above steps is as follows: This effectively addresses the nonlinear characteristics of the data.

[0093] (2.3) Repeat the above steps to obtain three different batches of normal operating condition grayscale images in the high-dimensional feature space. And an online test grayscale image

[0094] In this embodiment, nonlinear mapping is used to map the grayscale images of normal operating conditions and online testing to a high-dimensional feature space, which can effectively handle the nonlinearity problem of chiller unit operating data. The grayscale images of normal operating conditions and online testing after nonlinear transformation by a stacked autoencoder network are used as inputs for slow feature analysis (SFA), which improves the feature expression capability of the fault detection model.

[0095] To address the time-varying dynamic characteristics within batches of chiller units during operation and highlight fault feature information, a slow feature analysis method is used to calculate the slow feature difference between the normal operating condition grayscale image and the online test grayscale image after nonlinear transformation to a high-dimensional space. This eliminates the influence of dynamic data changes, highlights fault feature information, and effectively improves fault detection results.

[0096] Specifically, in the offline modeling and analysis phase, i.e., steps S1 to S4: First, the normal operating condition datasets from two different batches (different workdays) are converted into two different normal operating condition grayscale images through a graph transformation. These are then nonlinearly mapped to a high-dimensional feature space via a symmetric stacked autoencoder network. Next, slow feature analysis is used in the feature space to extract the slow feature information from these two normal operating condition grayscale images, and the difference is subtracted to obtain the slow feature information difference between the two normal batches of data (i.e., the normal operating condition slow feature difference map). This eliminates the dynamic changes in the data and establishes a fault detection model, which is a corresponding... Figure 1The various modules include a graph-to-data conversion module, a nonlinear mapping module based on a stacked autoencoder network, a feature extraction module based on slow feature analysis, and a fault detection module.

[0097] Specifically, in the online detection stage, i.e. steps 1 to 4: First, the online test dataset and a normal operating condition dataset are processed by image-data conversion and symmetric stacked autoencoder network mapping, respectively. Then, slow feature analysis is used in the feature space to extract the slow feature information of the online test grayscale image and the normal operating condition grayscale image, and the difference is subtracted to obtain the difference of slow feature information between the online test data and the normal operating condition data (i.e., the online test slow feature difference map), which can highlight the fault feature information of the online test data.

[0098] In this embodiment, a slow feature analysis method is used in a nonlinear high-dimensional space. First, slow feature information is extracted from two different grayscale images of normal operating conditions and the difference is calculated to eliminate the dynamic changes in the data and establish a fault detection model. Then, slow feature information is extracted from the grayscale images of the online test dataset and another normal operating condition dataset and the difference is calculated. This can highlight the fault feature information of the online test data.

[0099] The slow feature analysis methods for the offline modeling stage and the online testing stage are explained below.

[0100] In step S3, during the offline modeling stage, based on the feature difference calculation of slow feature analysis, a slow feature information difference map of two normal operating condition data is obtained, including the following steps:

[0101] Step S31: Based on the two normal operating condition grayscale images that have undergone nonlinear transformation to high-dimensional space... and Calculate the projection matrix W1 together;

[0102] 3-1) In the high-dimensional feature space, firstly, two grayscale images of normal operating conditions from different batches are compared. Centralized processing, in which Let n be a matrix containing only 1s, and n represent the number of sample points.

[0103]

[0104]

[0105] 3-2) Calculate the covariance matrix and cross-covariance matrix of the two matrices obtained after centering, and solve for the projection vector;

[0106] Calculate matrix and Covariance matrix and cross-covariance matrix:

[0107]

[0108]

[0109]

[0110] Where I is the identity matrix and r is the regularization constant.

[0111] make The objective function of the Slow Feature Analysis (SFA) can then be expressed as:

[0112]

[0113]

[0114] This optimization problem is transformed into a problem of solving generalized eigenvalues:

[0115] A1ω j =B1ω j λ j (14)

[0116] Where, λ j It is the generalized eigenvalue corresponding to the j-th generalized eigenvector.

[0117] When the generalized eigenvector ω is solved j By using ω j Normalization yields the final projection vector.

[0118] 3-3) Using the obtained projection vectors, further construct the projection matrix. in This represents the first p smallest eigenvalues ​​λ. j The corresponding feature vector.

[0119] Step S32: Multiply the transpose of the obtained projection matrix W1 by the two normal operating condition grayscale images that have undergone nonlinear transformation to high-dimensional space. and After obtaining the corresponding slow feature information and subtracting it, a slow feature difference map under normal operating conditions is obtained;

[0120] That is, during the offline modeling stage, slow feature information of grayscale images from two different batches of normal operating condition datasets is extracted.

[0121] During the offline modeling phase, the slow feature difference information D between grayscale images under normal operating conditions after slow feature processing is calculated. n =T1-T2, which is the slow feature difference map under normal operating conditions, thus processing the time-varying dynamic characteristics of the data;

[0122] Step 3, the online testing phase, involves calculating the feature difference based on slow feature analysis to obtain a slow feature information difference map between the online test data and the normal operating condition data. This includes the following steps:

[0123] Step 31: Based on the online test grayscale image that has undergone nonlinear transformation to a high-dimensional space Grayscale image under normal operating conditions The projection matrix W2 is obtained through joint calculation;

[0124] 3.1) Test grayscale image online Grayscale image under normal operating conditions Centralized processing:

[0125]

[0126]

[0127] 3.2) Calculate the covariance matrix and cross-covariance matrix of the two matrices obtained after centering, and solve for the projection vector;

[0128] The matrix was obtained through calculation. and covariance matrix and cross-covariance matrix

[0129] make The objective function of SFA can then be expressed as:

[0130]

[0131]

[0132] This optimization problem is transformed into a problem of solving generalized eigenvalues:

[0133] A2ω i =B2ω i λ i (20)

[0134] When the generalized eigenvector ω is solved i Then, through ω i Normalization yields the final projection vector.

[0135] 3.3) Using the obtained projection vectors, further construct the projection matrix. in This indicates the relationship between the first q smallest eigenvalues ​​λ. i The corresponding feature vector.

[0136] Step 32: Multiply the transpose of the obtained projection matrix W2 by the online test grayscale image that has undergone nonlinear transformation to a high-dimensional space. Grayscale image under normal operating conditions After obtaining the corresponding slow feature information, the difference is calculated to obtain the online test slow feature difference map.

[0137] That is, in the online detection phase, the slow feature information extracted from the online test grayscale image is: The slow feature information of the grayscale image under normal operating conditions is:

[0138] During the online testing phase, the slow feature difference information D between the grayscale images under normal operating conditions and those under online testing after slow feature processing is calculated. o =T0-T3, which is the online test slow feature difference map, thus highlighting the fault information in the online test data and effectively improving the performance of the fault detection model.

[0139] Based on the above steps, the grayscale images under normal operating conditions were analyzed respectively. and Calculate the projection matrix W1 and obtain the online test grayscale image. And another normal operating condition grayscale image After calculating the projection matrix W2, slow feature information of grayscale images from two different batches of normal operating condition datasets is extracted during the offline modeling stage. Slow feature information is extracted from the online test grayscale image during the online detection phase. And slow feature information of another normal operating condition grayscale image

[0140] During the offline modeling phase, the slow feature difference information D between grayscale images under normal operating conditions after slow feature processing is calculated. n =T1-T2, i.e., the slow feature difference map under normal operating conditions, thus handling the time-varying dynamic characteristics of the data; during the online testing phase, the slow feature difference information of the normal operating conditions and online testing grayscale images after slow feature processing is calculated, i.e., the online testing slow feature difference map D. o =T0-T3, thus highlighting the fault information in the online test data and effectively improving the performance of the fault detection model.

[0141] In step 4 and step S4, in order to further improve the accuracy of chiller unit fault detection, the sliding window method is used to expand the feature difference map after slow feature analysis to enrich the data in the slow feature difference map of the chiller unit.

[0142] The method for expanding the slow feature difference map using the sliding window approach includes the following steps:

[0143] Step 41: Set the hysteresis parameter to L and the sliding distance of the sliding window to q;

[0144] Step 42: The first subplot of the difference map is generated from the data in rows M to (M+L-1) of the feature difference map;

[0145] Step 43: Starting from the second difference map subgraph, slide each data row q rows backward as a difference map subgraph until the last row of the feature difference map, and obtain multiple difference map subgraphs in sequence;

[0146] Specifically, the second subplot of the difference map is generated by extracting data from row (M+q) of the feature difference map and ending at row (M+L+q-1), and so on, to obtain the subsequent subplots of the difference map.

[0147] For example, a slow feature difference plot has P process variables and T samples. Repeating the above operation can yield... Each subgraph has a size of L×P. Figure 7 Example diagram for expanding the slow feature difference map under normal operating conditions using the sliding window method.

[0148] Repeat the above steps to obtain the slow feature difference map D under normal operating conditions. n And online test slow feature difference map D o A series of subgraphs D after being expanded using the sliding window method n (i), i = 1, 2, ..., Q and D o (j), j=1,2,…,Q.

[0149] In step 4, during the online detection phase, optionally, the modulus ||D of the difference features of each online test slow feature difference map submap obtained using the sliding window method is used. o (j)||,j∈[1,2,…,Q] is used as a monitoring statistic, and the operating status of the chiller unit is monitored based on this monitoring statistic.

[0150] In step S4, during the offline modeling stage, it is necessary to determine the control limits of the monitoring statistics to determine whether the chiller unit is operating under normal conditions.

[0151] Considering the uncertainty distribution of process variables in chiller units, this embodiment uses the kernel density estimation method to determine the control limits of monitoring statistics. The expression for the univariate kernel estimator is:

[0152]

[0153] Where h is the smoothing parameter, K is the Gaussian kernel function, and Q is the number of subgraphs after expansion.

[0154] In some embodiments, the control limits are determined using a kernel density estimation method, including the following steps:

[0155] Step S41: Use the magnitude of the difference feature of each normal operating condition slow feature difference map sub-map obtained by the sliding window method as the monitoring statistic ||D n (i)||, i∈[1,2,…,Q];

[0156] Step S42: Calculate the normal operating condition monitoring statistic ||D| for the offline modeling phase using the univariate kernel density estimator. n (i) Density function of ||;

[0157] Step S42: Search for data points in the density function whose coverage area ratio is the set value to determine the control limit δ with the set confidence level in the normal operating condition data. Limit (||D n ||).

[0158] Optionally, the coverage area ratio can be set to 97%-99%, corresponding to a confidence level of 97%-99%. In this embodiment, the preferred setting is 99%.

[0159] In this embodiment, the kernel density estimation method is used to determine the threshold of the fault detection statistic based on the slow feature difference map of the two extended normal operating condition datasets; then, the statistic calculated by the slow feature difference map of the online test dataset and the normal operating condition dataset is compared with the determined threshold to detect whether the chiller unit has malfunctioned.

[0160] During the online testing phase, in step 4, the monitoring statistics of the online test data ||D o (j)||and control limit δ Limit (||D n Compare ||) . If the monitoring statistic ||D o (j)||greater than the control limit δ Limit (||D n If the indicator shows a value of ||, then the chiller system is malfunctioning. Otherwise, the chiller is operating normally. The fault detection procedure for the chiller system is as follows: Figure 8 As shown.

[0161] Example 2

[0162] Based on Example 1, this embodiment provides a chiller unit fault detection system based on data image differences, such as... Figure 1 As shown, it includes:

[0163] Image conversion module: configured to convert the acquired normal operating condition dataset and online test dataset into images respectively, resulting in online test grayscale image M0 and normal operating condition grayscale image M3;

[0164] The nonlinear mapping module based on a stacked autoencoder network is configured to perform nonlinear transformations using a symmetric stacked autoencoder network, mapping the resulting grayscale images to high-dimensional feature spaces.

[0165] The feature extraction module based on slow feature analysis is configured to use slow feature analysis to extract slow feature information from online test grayscale images and normal operating condition grayscale images that have been transformed to high-dimensional space through nonlinear transformation. The slow feature information is then subtracted to obtain a difference map of slow feature information between online test data and normal operating condition data.

[0166] Fault detection module: It is configured to expand the obtained slow feature information difference map, calculate the monitoring statistics based on the expanded difference map sub-map, and compare it with the set threshold of the fault detection statistics to determine whether a fault has occurred.

[0167] It should be noted that each module in this embodiment corresponds one-to-one with each step in embodiment 1, and their specific implementation process is the same, so it will not be repeated here.

[0168] Example 3

[0169] Based on Embodiment 1, this embodiment provides an electronic device, including a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When the computer instructions are executed by the processor, they complete the steps in the above-described chiller unit fault detection method based on data image differences.

[0170] Example 4

[0171] Based on Embodiment 1, this embodiment provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps in the above-described chiller unit fault detection method based on data image differences.

[0172] The above description is merely a preferred embodiment of this disclosure and is not intended to limit this disclosure. Various modifications and variations can be made to this disclosure by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

[0173] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.

Claims

1. A method for detecting faults in chiller units based on differences in data images, characterized in that, Includes the following steps: The acquired normal operating condition dataset and online test dataset were respectively converted into image datasets to obtain online test grayscale images. M Grayscale images of 0 and normal operating conditions M 3; A symmetric stacked autoencoder network is used for nonlinear transformation, mapping the resulting grayscale images to high-dimensional feature spaces, specifically including: The input grayscale image is non-linearly mapped through an encoder to obtain the non-linear features of the grayscale image, thereby fully displaying the non-linear features of the grayscale image in a high-dimensional space. Based on the obtained nonlinear features, the original input grayscale image is reconstructed through a decoder, mapping the normal operating condition grayscale image and the online test grayscale image to a high-dimensional feature space. For the online test grayscale image and the normal operating condition grayscale image after nonlinear transformation to the high-dimensional space, slow feature analysis is used to extract slow feature information and subtract them to obtain a slow feature information difference map between the online test data and the normal operating condition data. Specifically, this includes: The projection matrix is ​​calculated by combining the online test grayscale image (which has been nonlinearly transformed to a high-dimensional space) and the normal operating condition grayscale image. ; The obtained projection matrix The transpose of the image is multiplied by the online test grayscale image and the normal operating condition grayscale image after nonlinear transformation to high-dimensional space, respectively, and the corresponding slow feature information is obtained. The difference is then calculated to obtain the online test slow feature difference map. The obtained slow feature information difference map is expanded, and the monitoring statistics are calculated based on the expanded difference map sub-map. The statistics are compared with the set fault detection statistics threshold to determine whether a fault has occurred. The process of determining the threshold for fault detection statistics using offline modeling and analysis includes the following steps: Two batches of normal operating condition datasets were acquired and image converted separately to obtain corresponding grayscale images of the normal operating conditions. M 1. Grayscale image of normal operating conditions M 2; A symmetric stacked autoencoder network is used to perform nonlinear transformations, and the resulting grayscale images are mapped to high-dimensional feature spaces respectively; For two grayscale images of normal operating conditions that have been transformed into high-dimensional space by nonlinear transformation, slow feature analysis is used to extract slow feature information and subtract them to obtain a slow feature information difference map of the two normal operating condition data. The obtained slow feature information difference map is expanded, and the control limit of the monitoring statistics is calculated based on the expanded difference map sub-map, which is used as the threshold of the fault detection statistics.

2. The chiller unit fault detection method based on data image differences as described in claim 1, characterized in that: The acquired normal operating condition dataset and online test dataset were image transformed separately, and the following process was performed for each dataset: The data in the dataset is preprocessed by converting it into an unsigned integer type. The value of the converted integer data corresponds to the grayscale value of the image. The preprocessed process data collected by different sensors at the same sampling time is converted into a row in a grayscale image; The preprocessed process data collected by the same sensor at different sampling times is converted into a column of a grayscale image, thereby converting the normal operating condition dataset and the online test dataset into corresponding normal operating condition grayscale images and online test grayscale images.

3. The chiller unit fault detection method based on data image differences as described in claim 2, characterized in that: Preprocessing the data in the dataset includes standardizing, normalizing, and multiplying the normalization result by 255.

4. The chiller unit fault detection method based on data image differences as described in claim 1, characterized in that: The sliding window method is used to augment the slow feature difference map, including the following steps: Set the hysteresis parameter as The sliding distance of the sliding window is ; The first subplot of the difference map is composed of the first subplot of the feature difference map. Arrive Data generation for rows; Starting from the second difference map subgraph, each data row that slides q rows backward is treated as a difference map subgraph, until the last row of the feature difference map is obtained, thus obtaining multiple difference map subgraphs in sequence; Alternatively, the modulus of the difference features of each online test slow feature difference map subplot obtained using the sliding window method can be used as a monitoring statistic.

5. The chiller unit fault detection method based on data image differences as described in claim 1, characterized in that: During the offline modeling phase, based on the feature difference calculation of slow feature analysis, a slow feature information difference map of two normal operating condition data is obtained, including the following steps: The projection matrix is ​​calculated by combining two grayscale images of normal operating conditions that have been nonlinearly transformed to a high-dimensional space. The obtained projection matrix The transpose of the image is multiplied by two normal operating condition grayscale images that have undergone nonlinear transformation to high-dimensional space to obtain the corresponding slow feature information. The difference between these two images is then used to obtain the normal operating condition slow feature difference image. Alternatively, during the offline modeling phase, the kernel density estimation method can be used to determine the control limits of the monitoring statistics, including the following process: The modulus of the difference feature of each normal operating condition slow feature difference map sub-map obtained by the sliding window method is used as the monitoring statistic. Based on the univariate kernel density estimator, the density function of the normal operating condition monitoring statistics in the offline modeling stage is calculated; The search density function is used to determine the control limits with a confidence level of the set value for data points whose coverage area ratio is the set value in the normal operating condition data.

6. A chiller unit fault detection system based on data image differences, characterized in that, This system is used to execute the chiller unit fault detection method based on data image differences as described in claim 1, and the system includes: Image conversion module: Configured to convert the acquired normal operating condition dataset and online test dataset into image representations, respectively, to obtain online test grayscale images. M Grayscale images of 0 and normal operating conditions M 3; The nonlinear mapping module based on a stacked autoencoder network is configured to perform nonlinear transformations using a symmetric stacked autoencoder network, mapping the resulting grayscale images to high-dimensional feature spaces. The feature extraction module based on slow feature analysis is configured to use slow feature analysis to extract slow feature information from online test grayscale images and normal operating condition grayscale images that have been transformed to high-dimensional space through nonlinear transformation. The slow feature information is then subtracted to obtain a difference map of slow feature information between online test data and normal operating condition data. Fault detection module: It is configured to expand the obtained slow feature information difference map, calculate the monitoring statistics based on the expanded difference map sub-map, and compare them with the set threshold of the fault detection statistics to determine whether a fault has occurred. The process of determining the threshold for fault detection statistics using offline modeling and analysis includes the following steps: Two batches of normal operating condition datasets were acquired and image converted separately to obtain corresponding grayscale images of the normal operating conditions. M 1. Grayscale image of normal operating conditions M 2; A symmetric stacked autoencoder network is used to perform nonlinear transformations, and the resulting grayscale images are mapped to high-dimensional feature spaces respectively; For two grayscale images of normal operating conditions that have been transformed into high-dimensional space by nonlinear transformation, slow feature analysis is used to extract slow feature information and subtract them to obtain a slow feature information difference map of the two normal operating condition data. The obtained slow feature information difference map is expanded, and the control limit of the monitoring statistics is calculated based on the expanded difference map sub-map, which is used as the threshold of the fault detection statistics.

7. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the steps in the chiller unit fault detection method based on data image differences as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, complete the steps in the chiller unit fault detection method based on data image differences as described in any one of claims 1-5.

Images