A visual fault detection method and system for air conditioning units

Abstract

The present invention proposes a visual fault detection method and system for air-conditioning units, the method comprising collecting multi-source data of air-conditioning units and storing them in matrix form, and then preprocessing the stored data; adaptively fusing features to obtain a comprehensive feature vector, performing feature enhancement, dimensionality reduction and regularization processing on the comprehensive feature vector to obtain the final comprehensive feature vector; constructing a causal relationship model, calculating causal effects, performing multi-level fault diagnosis based on the results of the causal effects, and constructing a fault prediction model for fault inference; performing local model training on each edge computing node based on the results of the fault diagnosis and fault prediction models; constructing a fault heat map generation algorithm to visualize the fault, and designing a multi-mode intelligent interactive interface, and adjusting the visualization interface and parameters according to the user's behavior and feedback. The present invention introduces a number of innovative technologies such as multi-source data fusion, edge computing, causal inference and federated learning, which improves the system's fault location and prevention capabilities.

Description

Technical Field

[0001] The invention belongs to the field of fault detection, and in particular relates to a visual fault detection method and system for an air-conditioning unit. Background Art

[0002] With the development of science and technology and the improvement of people's quality of life, air conditioning units have become an indispensable basic equipment in modern buildings, industrial facilities and smart homes. The operating environment of air conditioning units is usually complex and changeable, and the equipment needs to work continuously for a long time, which makes the fault detection and maintenance management of air conditioning units particularly important. Traditional methods for fault detection of air conditioning units mostly rely on single sensor data, such as temperature, pressure, current and vibration sensors. These methods usually determine whether the equipment is faulty by setting a fixed threshold. Once a parameter exceeds the threshold, an alarm signal will be triggered. However, this simple threshold-based monitoring method shows obvious limitations in complex practical applications. First, due to the complex interactions between the various components inside the air conditioning unit and the operating environment, the data of a single sensor often cannot fully reflect the overall operating status of the equipment, resulting in low accuracy of fault detection. Especially when there are interdependencies between multi-sensor data, traditional methods have difficulty identifying these dependencies and are prone to false positives or false negatives.

[0003] Secondly, the response speed of traditional fault detection systems is also difficult to meet actual needs. Most existing systems rely on a centralized data processing architecture, and all sensor data needs to be transmitted to a central server for analysis. Although this method can centrally process a large amount of data, it also brings significant communication delays and computing burdens. When an air conditioning unit fails, it is impossible to respond in real time, which may cause further damage to the equipment or even cause a safety accident. In addition, traditional centralized analysis also faces data privacy and security issues, especially in smart homes and commercial buildings, where users have an increasing demand for data privacy.

[0004] In trying to solve these problems, some existing improved methods, such as machine learning-based fault detection models, have improved the intelligence level of detection to a certain extent, but these methods often rely on a large amount of labeled data and powerful computing resources, and are difficult to be widely promoted in practical applications. In addition, these models are usually trained offline and cannot adapt to the dynamic changes in the equipment operating environment, resulting in the effectiveness of the models in practical applications gradually decreasing over time. Furthermore, the existing system lacks the ability to analyze the root causes of faults and prevent faults. Even if a fault is detected, it is difficult to provide accurate diagnostic information and effective prevention strategies. In summary, the existing air-conditioning unit fault detection technology still has many shortcomings in terms of real-time, accuracy, fault location, data privacy and system intelligence, and an innovative solution that can solve these problems is urgently needed. Summary of the invention

[0005] The present invention overcomes many shortcomings of existing air-conditioning unit fault detection technology by introducing a number of innovative technologies such as multi-source data fusion, edge computing, causal inference, federated learning and visual diagnosis, significantly improves the real-time performance, accuracy, fault location and prevention capabilities of the system, and provides a complete and advanced solution for the efficient operation and intelligent maintenance of air-conditioning units.

[0006] In order to achieve the above object, a first aspect of the present invention provides a visual fault detection method for an air conditioning unit, the method comprising:

[0007] S1, collecting multi-source data of the air-conditioning unit and storing it in a matrix form, and then preprocessing the stored data;

[0008] S2, extracting features from the preprocessed data, and adaptively fusing the extracted multi-source features to obtain a comprehensive feature vector, then constructing a second-order tensor network to enhance the comprehensive feature vector, and performing dimensionality reduction processing on the comprehensive feature vector after feature enhancement, and then constructing a regularized noise suppression model to regularize the comprehensive feature vector after dimensionality reduction processing to obtain the final comprehensive feature vector;

[0009] S3. Construct a causal relationship model based on the final comprehensive feature vector, calculate the causal effect based on the causal relationship model, perform multi-level fault diagnosis based on the results of the causal effect, and construct a fault prediction model, perform fault inference based on causal inference based on the results of the fault diagnosis, and predict future faults;

[0010] S4. Perform local model training on each edge computing node according to the results of the fault diagnosis and fault prediction models, wherein the local model must be trained independently on each edge computing unit. After each edge computing unit completes the local training, it does not share the original data, but only uploads the local model parameters to the central server. The central server aggregates the global model based on the local model, and after completing the aggregation of the global model, the central server sends the updated global model parameters to each edge computing unit. When data anomalies or model failures occur in some edge computing units, an adaptive anomaly detection and adjustment mechanism is designed to detect and correct the abnormal data.

[0011] S5. Construct a fault heat map generation algorithm based on the final comprehensive feature vector and the results of the fault prediction model to visualize the fault, and design a multi-mode intelligent interactive interface. Adaptive adjustment is performed based on the designed adaptive data focusing algorithm. Finally, adjust the parameters of the visualization interface and the multi-mode intelligent interactive interface based on the user's behavior and feedback during use.

[0012] The adaptive anomaly detection and adjustment mechanism includes:

[0013] In each edge computing unit, calculate the anomaly score A of the current model i (t), the anomaly score A of the current model i (t) is used to monitor the stability of model parameters and training quality, and is expressed as follows:

[0014]

[0015] in, is the local model parameter w i (t) and global model parameters The Euclidean distance of Var(L i (w i (t))) is the local model loss function L i (w i The variance of (t)), β i is the adjustment coefficient, which is used to balance the contribution of these two parts to the anomaly score;

[0016] For the model with the highest anomaly score, an adaptive adjustment mechanism is used to correct it. The specific adjustment formula is:

[0017]

[0018] Among them, δ is the adjustment step size, are the adjusted model parameters, is the local model loss function L i (w i (t)) gradient optimization;

[0019] A dynamic anomaly threshold θ(t) is designed so that the system can automatically adjust its tolerance to abnormal models according to changes in the global model, as shown below:

[0020]

[0021] Among them, θ0 is the initial threshold, κ is the adjustment parameter of the edge computing node, is the variance of the sum of anomaly scores of all edge computing units.

[0022] Furthermore, the storage in matrix form is expressed as follows:

[0023] Assuming the number of sensors is N, the time step is T, and the spatial resolution of the visual data is W×H, the data of various sensors are stored in matrix form:

[0024]

[0025] Dvis ∈R H×W×H×T

[0026] Among them, D temp , D press , D current , D vib are the time series data matrices of temperature, pressure, current, and vibration sensors, respectively, x is the number of sensors of the corresponding type, D vis is the visual data, M is the number of cameras, W and H are the image width and height respectively, and T is the time step;

[0027] The preprocessing includes time synchronization and data alignment, data standardization and normalization processing, outlier detection and correction processing, and regularization and dimensionality reduction processing.

[0028] Furthermore, the outlier detection and correction process is expressed as follows:

[0029]

[0030] in, is the corrected data, z i,i , is normal data, k is the threshold factor for outlier detection, α is an adjustment coefficient used to correct outliers; μ i is the mean; σ i is the standard deviation;

[0031] The regularization and dimensionality reduction processing is expressed as follows:

[0032] Assume that the sensor data at time t is vector z t =[z 1,t ,z 2,t ,…,z N,t ], processed by the regularization formula, the regularization formula is expressed as follows:

[0033]

[0034] in, is the regularized data, L is the Laplace matrix used to capture the spatial relationship between sensors, and λ is the regularization parameter.

[0035] Furthermore, the specific steps of S2 are as follows:

[0036] S21. For physical sensor data, a one-dimensional convolutional neural network is used to process time series features, and each type of data generates preliminary feature vectors of temperature, pressure, current, and vibration data; for visual data, a two-dimensional convolutional neural network is used to process the spatial features of the image and generate preliminary visual feature vectors;

[0037] S22, normalizing the preliminary feature representations of the temperature, pressure, current, and vibration data and the preliminary visual feature vectors to obtain feature vectors and visual feature vectors of the temperature, pressure, current, and vibration data, and designing an adaptive weighted multi-source feature fusion model to dynamically adjust the fusion weights of different feature vectors to obtain a first comprehensive feature vector;

[0038] S23, construct a second-order tensor network to enhance the first comprehensive feature vector, and then perform dimensionality reduction processing to obtain a second comprehensive feature vector, which is expressed as follows:

[0039]

[0040] in, is the outer product operation, F enhanced (t) is the second comprehensive eigenvector, F fused (t) is the first comprehensive eigenvector;

[0041] S24, constructing a regularized noise suppression model to perform regularization processing on the second comprehensive feature vector, which is expressed as follows:

[0042]

[0043] Where L is the Laplace matrix, λ is the regularization parameter, is the final comprehensive feature vector.

[0044] Furthermore, the steps of constructing the causal relationship model are as follows:

[0045] First, a causal relationship model is constructed based on the final comprehensive feature vector, where each feature is a node in the graph, and the edges between the nodes represent the causal relationship between the features. The causal relationship is represented by the causal relationship matrix C as follows:

[0046]

[0047] Among them, C ij is the strength of the causal relationship between feature i and feature j, Granger(·) is a function based on the Granger causality test, which is used to detect the causal relationship between time series. is the final comprehensive feature vector of feature i, is the final comprehensive feature vector of feature j;

[0048] According to the specific application scenario of the air-conditioning unit, a regular term Ω based on physical constraints is designed in the causal relationship matrix C to enhance the interpretability of the causal relationship, which is expressed as follows:

[0049]

[0050] in, is the enhanced causal relationship matrix element, reflecting the causal influence strength between feature i and feature j, Ω ij is a correction term based on the specific physical structure of the air-conditioning unit or the sensor location constraints, C ij is the causal influence strength between feature i and feature j, and θ is the parameter that adjusts the causal relationship matrix.

[0051] Furthermore, the calculation method of the causal effect is expressed as follows:

[0052]

[0053] Among them, CE i (t) is the causal effect size of feature i on health state y(t), It is the enhanced causal relationship matrix element, reflecting the causal influence strength between feature i and feature j; by calculating CE i (t), determine the contribution of each feature to the system health status at the current moment and identify the root cause of the fault;

[0054] The multi-level fault diagnosis specifically includes:

[0055] Rank the k features with the largest causal effects and use them as key features for preliminary fault diagnosis;

[0056] An adaptive regularization term Γ is introduced to correct the causal effect, which is expressed as follows:

[0057]

[0058] Among them, α i is the weight of each feature, which is dynamically adjusted based on its historical performance, Γ i is a regularization term used to smooth causal effects and reduce noise interference; This is the corrected fault diagnosis result.

[0059] Furthermore, according to the fault diagnosis results Construct a fault prediction model, which is expressed as follows:

[0060] Assume the fault state at the future time t+Δt is The prediction model is based on the historical causal effect vectors CE(t), CE(t-Δt), …, CE(t-nΔt), which are expressed as follows:

[0061]

[0062] Among them, g(·) is a time series prediction function based on a multi-layer recurrent neural network. By backtracking and analyzing past causal effects and combining current causal effects, the model outputs the fault prediction value at the future moment.

[0063] The fault sensitivity adjustment term ΔS(t+Δt) is introduced to dynamically adjust the prediction results to adapt to the behavior characteristics of the air-conditioning unit under special working conditions, which is expressed as follows:

[0064]

[0065] Among them, ΔS(t+Δt) is based on modeling the historical data of specific working conditions, reflecting the operating characteristics of the air-conditioning unit under abnormal conditions. It is the forecast result after dynamic adjustment;

[0066] according to The value of exceeds the set fault threshold τ. When the value exceeds the set fault threshold τ, the system will automatically issue a fault warning signal and transmit the prediction results and current operating data to the central monitoring system for decision analysis.

[0067] Furthermore, when the global model is aggregated, a regularization term is constructed It is used to modify the aggregation weight of each local model, which is expressed as follows:

[0068]

[0069] Among them, γ is the adjustment parameter of the global model, It is an adjustment item designed based on the performance and differences of the local model. is the updated global model, N is the total number of samples, and p i is the weight factor, are local model parameters;

[0070] After completing the aggregation of the global model, the central server will update the global model parameters It is sent to each edge computing unit. After receiving the updated global model, each unit performs fine-tuning based on local data. The update formula of the local model is expressed as follows:

[0071]

[0072] Where η is the learning rate, is the direction of gradient descent based on global model parameters and local data, It is a newer local model;

[0073] An adaptive step size adjustment mechanism is constructed to dynamically adjust the learning rate η to adapt to different training stages and data conditions. The specific expression is as follows:

[0074]

[0075] in, is the updated learning rate, is the current learning rate, ζ is the step size adjustment coefficient, is the variance of the loss function, which is used to reflect the instability in model training.

[0076] Furthermore, the step S5 specifically includes:

[0077] S51, construct a fault heat map generation algorithm based on the final comprehensive feature vector and the results of the fault prediction model to visualize the fault; wherein, the fault heat map generation algorithm is to convert the fault diagnosis results Converted into a heat map, it is represented as follows:

[0078]

[0079] Among them, H(x,y,t) is the thermal map value of the air conditioning unit at the position (x,y), α i (t) is the contribution weight of feature i to the fault at time t, M i (x,y) is the component position matrix corresponding to feature i;

[0080] S52. Design a multi-mode intelligent interactive interface to perform adaptive adjustment according to the designed adaptive data focusing algorithm. The adaptive data focusing algorithm automatically adjusts the information displayed on the interface according to the user's operation behavior and the severity of the current fault. At the same time, design a voice and text interaction module to support users to query the current device status and fault cause through natural language, and build a natural language fault explanation model in the voice and text interaction module to convert complex diagnostic results into understandable language output;

[0081] S53. Design a user satisfaction score based on the user's behavior and feedback during use, and adjust the parameters of the visualization interface and the multi-modal intelligent interaction interface based on the user satisfaction score; wherein the user satisfaction score is expressed as follows:

[0082]

[0083] Among them, γ j (t) is the weight of each user behavior, m is the total number of books used, Behavior j (t) is the specific operation record of the user at time t.

[0084] In a second aspect of the present invention, a visual fault detection system for an air conditioning unit is provided, the system comprising:

[0085] A data acquisition module is used to collect multi-source data of the air-conditioning unit and store it in a matrix form, and then pre-process the stored data;

[0086] The data fusion and feature extraction module is used to extract features from the preprocessed data, and adaptively fuse the extracted multi-source features to obtain a comprehensive feature vector, then build a second-order tensor network to enhance the comprehensive feature vector, and perform dimensionality reduction on the comprehensive feature vector after feature enhancement, and then build a regularized noise suppression model to regularize the comprehensive feature vector after dimensionality reduction to obtain the final comprehensive feature vector;

[0087] The fault diagnosis and prediction module is used to build a causal relationship model based on the final comprehensive feature vector, calculate the causal effect based on the causal relationship model, perform multi-level fault diagnosis based on the results of the causal effect, and build a fault prediction model, perform fault inference based on causal inference based on the results of the fault diagnosis, and predict future faults;

[0088] The model optimization and upgrade module is used to perform local model training on each edge computing node according to the results of the fault diagnosis and fault prediction models. The local model must be trained independently on each edge computing unit. After each edge computing unit completes the local training, it does not share the original data, but only uploads the local model parameters to the central server. The central server aggregates the global model based on the local model. After the central server completes the aggregation of the global model, it sends the updated global model parameters to each edge computing unit. When some edge computing units have data anomalies or model failures, an adaptive anomaly detection and adjustment mechanism is designed to detect and correct the abnormal data.

[0089] The fault visualization and intelligent interaction module is used to construct a fault heat map generation algorithm based on the final comprehensive feature vector and the results of the fault prediction model to visualize the fault, design a multi-mode intelligent interaction interface, perform adaptive adjustments based on the designed adaptive data focusing algorithm, and finally adjust the parameters of the visualization interface and the multi-mode intelligent interaction interface based on the user's behavior and feedback during use;

[0090] The adaptive anomaly detection and adjustment mechanism includes:

[0091] In each edge computing unit, calculate the anomaly score A of the current model i (t), the anomaly score A of the current model i (t) is used to monitor the stability of model parameters and training quality, and is expressed as follows:

[0092]

[0093] in, is the local model parameter w i (t) and global model parameters The Euclidean distance of Var(L i (w i (t))) is the local model loss function L i (w i The variance of (t)), β i is the adjustment coefficient, which is used to balance the contribution of these two parts to the anomaly score;

[0094] For the model with the highest anomaly score, an adaptive adjustment mechanism is used to correct it. The specific adjustment formula is:

[0095]

[0096] Among them, δ is the adjustment step size, are the adjusted model parameters, is the local model loss function L i (w i (t)) gradient optimization;

[0097] A dynamic anomaly threshold θ(t) is designed so that the system can automatically adjust its tolerance to abnormal models according to changes in the global model, as shown below:

[0098]

[0099] Among them, θ0 is the initial threshold, κ is the adjustment parameter of the edge computing node, is the variance of the sum of anomaly scores of all edge computing units.

[0100] The beneficial technical effects of the present invention are at least as follows:

[0101] (1) In order to solve the problem that traditional methods rely on single sensor data and cannot fully reflect the operating status of equipment, the present invention uses multi-source data fusion technology to deeply integrate multiple physical quantity data such as temperature, pressure, current, vibration, etc. with visual data (such as camera video and infrared images). This fusion of multi-source data enables the system to fully monitor the operating status of the air-conditioning unit and significantly improves the accuracy of fault detection.

[0102] (2) In order to solve the problems of communication delay and slow response speed in traditional centralized fault detection systems, the present invention introduces edge computing technology. The edge computing unit is deployed at the key nodes of the air-conditioning unit, which can process and analyze multi-source data in real time locally, reduce the delay of data transmission, realize real-time detection and response of faults, and thus improve the overall response speed of the system.

[0103] (3) In order to solve the problem that existing systems lack the ability to accurately diagnose and prevent complex faults, the present invention uses a causal inference model to conduct in-depth analysis of the integrated multi-source data, identify the causal relationship between the data, and determine the root cause of the fault. At the same time, the causal inference model can also predict possible fault risks in advance, helping users take preventive measures before the fault occurs.

[0104] (4) In order to solve the problem that traditional models become ineffective over time, the present invention also uses federated learning technology. Through federated learning, the edge computing units of multiple air-conditioning units can share the parameters of the detection model while protecting data privacy, realize cross-device collaborative learning and model optimization, and ensure that the system can continuously adapt to changes in the operation of the equipment.

[0105] (5) In order to improve the operability and user experience of the system, the present invention designs an intuitive visual diagnosis system. Through the visual interface, users can view the operating status, fault information and diagnosis results of the air-conditioning unit in real time, and obtain maintenance suggestions and operation instructions in an interactive way. This visualization and intelligent interactive function makes fault diagnosis more intuitive and easy to understand, helping users to take quick action. BRIEF DESCRIPTION OF THE DRAWINGS

[0106] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0107] Figure 1 The present invention is a flow chart of a visual fault detection method for an air-conditioning unit.

[0108] Figure 2 The present invention is a framework diagram of a visual fault detection system for an air conditioning unit. DETAILED DESCRIPTION

[0109] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0110] like Figure 1 As shown, an embodiment of the present invention provides a visual fault detection method for an air conditioning unit, the method comprising:

[0111] S1. Collect multi-source data of the air-conditioning unit and store it in a matrix form, and then pre-process the stored data.

[0112] Data processing and collection are the foundation of the entire system. By deploying a variety of sensors and cameras on key components of the air conditioning unit, the system can collect multi-source data including temperature, pressure, current, vibration, visual data, etc. After these data are standardized, noise and inconsistency are removed, ensuring that different types of data can be compared and analyzed on the same timeline. This step lays the foundation for subsequent analysis, ensuring the reliability and consistency of the data, thus providing the prerequisite for accurate fault detection.

[0113] Furthermore, in said S1, the specific steps are:

[0114] S101. The collected physical quantity data and visual data are stored in a matrix form, which is expressed as follows:

[0115] Various sensors are deployed on key components of air-conditioning units, such as compressors, condensers, fans, etc., to collect physical quantity data (such as temperature, pressure, current, vibration) and visual data (such as images and videos captured by visible light and infrared cameras).

[0116] Specifically, assuming that the number of sensors is N, the time step is T, and the spatial resolution of the visual data is W×H, the data of various sensors are stored in matrix form, as shown below:

[0117]

[0118] D vis ∈R M×W×H×T

[0119] Among them, D temp , D press , D current , D vib are the time series data matrices of temperature, pressure, current, and vibration sensors, respectively, x is the number of sensors of the corresponding type, D vis is the visual data, M is the number of cameras, W and H are the image width and height respectively, and T is the time step;

[0120] The preprocessing includes time synchronization and data alignment, data standardization and normalization processing, outlier detection and correction processing, and regularization and dimensionality reduction processing;

[0121] S102. Various sensors of the air conditioning unit may have different sampling frequencies, and the data needs to be synchronized. Define a unified time base τ, and use linear interpolation or piecewise high-order polynomial interpolation to align all sensor data to the same time point, as shown below:

[0122]

[0123] in, is the data matrix after time synchronization, τ j and τ i+1 is the original sampling time point, t is the unified reference time, D x (i,t) is the data value of sensor i at time t.

[0124] The data matrix after time synchronization is expressed as follows:

[0125]

[0126] Through time alignment, the data of all sensors are unified on the same timeline, ensuring that subsequent fusion and analysis operations can be performed on a consistent time basis;

[0127] S103, perform standardization to eliminate the differences in data scales of different sensors. First, calculate the mean μ of each sensor i at time t i and standard deviation σ i , for each data point d i,t Normalized to zero mean and unit variance, expressed as follows:

[0128]

[0129] in,

[0130] For visual data, considering that the range of image pixel values ​​is usually fixed (such as 0-255), the visual data is normalized and mapped to the [0,1] interval, as shown below:

[0131]

[0132] Among them, Z vis is the normalized visual data matrix. The standardized and normalized data is:

[0133] Z temp ,Z press ,Z current ,Z vib ,Z vis

[0134] These standardized data eliminate dimensional differences, making subsequent fusion processing more consistent;

[0135] S104. Detect and process abnormal values ​​that may appear in the actual collection process. In order to effectively identify and correct abnormal values, an innovative abnormal detection formula is used as follows:

[0136]

[0137] Among them, k is the threshold factor for outlier detection, and α is an adjustment coefficient, which is usually set to a constant less than 1, used to correct outliers instead of simply removing them. This correction method can reduce the impact of outliers on subsequent analysis while retaining the overall trend of the data;

[0138] S105. In order to further improve the quality of data, an innovative regularization process is introduced to smooth the data matrix. Assume that the sensor data at time t is vector z t =[z 1,t ,z 2,t ,…,z N,t ], the regularization formula used for processing is expressed as follows:

[0139]

[0140] Where L is the Laplacian matrix of the sensor network, which is used to capture the spatial relationship between sensors, and λ is the regularization parameter. This formula can smooth the noise in the data and retain the structural characteristics of the data.

[0141] In addition, the visual data is reduced in dimensionality through dimensionality reduction techniques such as principal component analysis (PCA) or non-negative matrix factorization (NMF) to extract the main features for subsequent analysis to reduce the computational burden.

[0142] Finally, the data sets that have been time-synchronized, standardized, corrected for outliers, and regularized are stored. All data are uniformly stored as multi-dimensional time series data to ensure seamless integration with subsequent real-time analysis and fusion operations.

[0143] S2. Perform feature extraction on the preprocessed data, and adaptively fuse the extracted multi-source features to obtain a comprehensive feature vector. Then, construct a second-order tensor network to enhance the comprehensive feature vector, and perform dimensionality reduction on the comprehensive feature vector after feature enhancement. Then, construct a regularized noise suppression model to regularize the comprehensive feature vector after dimensionality reduction to obtain the final comprehensive feature vector.

[0144] Relying on the data collected and preprocessed in step 1, the system can process and analyze the data in real time by deploying edge computing units locally on the device. Since edge computing reduces the delay in data transmission, the system can detect and make preliminary diagnoses at the moment of fault occurrence, thus avoiding false alarms or missed alarms caused by communication delays. Further, the data from different sensors are deeply fused. Through the feature extraction algorithm, the key features closely related to the fault are extracted from the fused data. The key to this step is to improve the real-time performance of the system so that the detection results can be quickly fed back. By utilizing the real-time analysis results obtained, the feature weights are dynamically adjusted to ensure that the most representative fault features can be extracted under different working conditions.

[0145] Specifically, in S2, the specific steps are:

[0146] S201. In step 1, the data has been preprocessed by standardization, time synchronization, anomaly detection, etc. to obtain a standardized data matrix Z. temp , Z press , Z current , Z vib and Z vis These data serve as input to the edge computing unit.

[0147] Specifically, the present invention performs preliminary feature extraction on physical sensor data (temperature, pressure, current, vibration) and visual data in the edge computing unit. For physical sensor data, a one-dimensional convolutional neural network (1D-CNN) is used to process time series features, and each type of data generates a corresponding feature vector, which are expressed as follows:

[0148] F temp =1D-CNN(Z temp ),F press =1D-CNN(Z press ),

[0149] F current =1D-CNN(Z current ),F vib =1D-CNN(Z vib )

[0150] Among them, F temp 、F press 、F current and F vib They represent the preliminary features of temperature, pressure, current, and vibration data, respectively.

[0151] For the visual data Z vis , a two-dimensional convolutional neural network (2D-CNN) is used to process the spatial features of the image and generate a preliminary visual feature vector, which is expressed as follows:

[0152] F vis =2D-CNN(Z vis )

[0153] In this way, a multi-dimensional preliminary feature vector set F is generated in the edge computing unit. temp 、F press 、F current 、F vib and F vis , which represent the operating status of various aspects of the air-conditioning unit at the current moment.

[0154] S202, the edge computing unit has extracted preliminary feature vectors F for temperature, pressure, current, and vibration data through a one-dimensional convolutional neural network (1D-CNN). temp 、F press 、F current and F vib , and extracted the feature vector F of the visual data through a two-dimensional convolutional neural network (2D-CNN) vis These feature vectors represent various sensor information of the air-conditioning unit at the current moment.

[0155] Since these feature vectors come from different sensors and have different dimensions and characteristics, direct fusion may cause mutual influence between features or information loss. Therefore, each feature vector is first normalized so that their scales are within the same range to ensure that there is no weight imbalance due to the magnitude difference between features during fusion.

[0156] Specifically, the preliminary feature representations of temperature, pressure, current, and vibration data and the preliminary visual feature vectors are normalized to obtain feature vectors and visual feature vectors of temperature, pressure, current, and vibration data, which are expressed as follows:

[0157]

[0158] Among them, F x ′ is the normalized eigenvector, min(F x ) and max(F x ) are the minimum and maximum values ​​of the eigenvectors, respectively.

[0159] Then, an adaptive weighted multi-source feature fusion model is designed to dynamically adjust the fusion weights of different feature vectors so that the data characteristics of each sensor can be fully utilized in the fusion process.

[0160] Specifically, the initial weights of different sensor features are set to w = [w temp ,w press ,w current ,wvib ,w vis ], these initial weight values ​​are set through historical data statistics and domain expert experience. In order to make the fusion process adapt to the current operating status of the device, the weight adjustment function ψ(F x ′,t), the weights are adaptively adjusted based on the feature vector at the current moment, as shown below:

[0161]

[0162] Among them, ψ(F x ′,t) is based on the eigenvector F x ′ is a dynamic adjustment function that takes into account the fault sensitivity and historical performance at the current time point t. The adjusted weight w′ x The weights are optimized according to the actual situation during the fusion process.

[0163] The first integrated feature vector F after fusion fused The calculation formula is as follows:

[0164]

[0165] Among them, F x ′(t) is the normalized eigenvector, F fused (t) is the final multi-source fusion feature, which represents the overall operating status of the air-conditioning unit at the current moment.

[0166] S203. In addition, for complex faults (such as correlated faults of multi-source signals), it is difficult to capture its deep correlation information by simple feature fusion. Therefore, a correlation-enhanced feature extraction model (CE-FE) is introduced to capture the nonlinear relationship between multi-source features.

[0167] Specifically, the fused comprehensive feature vector F is used fused (t), construct a second-order tensor network to enhance the first comprehensive feature vector, and then perform dimensionality reduction processing to obtain the second comprehensive feature vector, which is expressed as follows:

[0168]

[0169] in, is an outer product operation, through which a higher-dimensional feature representation can be obtained. enhanced (t) is the second comprehensive eigenvector, F fused (t) is the first comprehensive eigenvector;

[0170] After feature enhancement, enhanced(t) Dimensionality reduction is performed to reduce redundant information and computational complexity, and the final extracted feature vector F final (t) will serve as the input for subsequent fault diagnosis and prediction models.

[0171] S204. In the process of multi-source data fusion and enhancement, noise and redundant features may be introduced, which may affect the accuracy of subsequent fault diagnosis. In order to overcome this problem, a regularized noise suppression model (RNS) is used to perform the feature vector F final (t) Perform regularization.

[0172] A regularized noise suppression model is constructed to regularize the second comprehensive feature vector, which is expressed as follows:

[0173]

[0174] Among them, L is the Laplace matrix between features, λ is the regularization parameter, is the final comprehensive feature vector. Through this regularization process, noise and redundant features can be effectively suppressed and the final feature quality can be improved.

[0175] Finally, The features are passed to the fault diagnosis model to ensure that the system can accurately detect and predict faults based on high-quality features.

[0176] S3. Construct a causal relationship model based on the final comprehensive feature vector, calculate the causal effect based on the causal relationship model, perform multi-level fault diagnosis based on the results of the causal effect, and construct a fault prediction model. Perform fault inference based on causal inference based on the results of the fault diagnosis to predict future faults.

[0177] Based on the key features extracted in step S2, the causal inference model is used to conduct in-depth analysis of the data to identify the causal relationship between different data features, thereby accurately locating the root cause of the failure. In addition, by combining historical data and current status, the causal inference model can predict possible future failures and provide preventive suggestions.

[0178] Specifically, the specific steps of S3 include:

[0179] S301, in step S2, a high-dimensional comprehensive feature vector is generated through multi-source data fusion and feature extraction These feature vectors reflect the complex operating status of the air conditioning unit at a specific moment. In order to further understand the mutual influence between these features and accurately locate the possible root causes of failures, it is necessary to build a causal relationship model. The steps of model construction are as follows:

[0180] First, based on Construct a causal relationship model, in which each feature is a node in the graph, and the edges between nodes represent the causal relationship between features. The causal relationship is represented by the causal relationship matrix C as follows:

[0181]

[0182] Among them, C ij is the strength of the causal relationship between feature i and feature j, and Granger(·) is a function based on the Granger causality test, which is used to detect the causal relationship between time series. By calculating the causal relationship matrix C, a causal relationship model that reflects the causal relationship between each feature in multi-source data can be constructed.

[0183] According to the specific application scenario of the air-conditioning unit, a regular term Ω based on physical constraints is designed in the causal relationship matrix C to enhance the interpretability of the causal relationship, which is expressed as follows:

[0184]

[0185] Among them, Ω ij is a correction term based on the specific physical structure of the air-conditioning unit or the sensor location constraints, and θ is a parameter that adjusts the causal relationship matrix. This adjustment can better reflect the actual causal relationship in the operation of the equipment and improve the reliability of the model;

[0186] S302. Based on the above causal relationship model, calculate the impact of each feature on the overall operating state of the air-conditioning unit, i.e., the causal effect. The causal effect is used to identify which features are potential sources of failure. Define the target variable y(t) to represent the health state of the air-conditioning unit at time t. The larger the value, the healthier the system and the lower the possibility of failure.

[0187] For each feature i, the calculation formula of the causal effect is expressed as follows:

[0188]

[0189] Among them, CE i (t) is the causal effect size of feature i on health state y(t), It is the enhanced causal relationship matrix element, reflecting the causal influence strength between feature i and feature j. By calculating CE i (t), the contribution of each feature to the system health status at the current moment can be determined, thereby identifying the possible root causes of failures.

[0190] On this basis, multi-level fault diagnosis is performed. First, the k features with the largest causal effects are ranked and used as key features for preliminary fault diagnosis. Furthermore, in order to improve the accuracy of diagnosis, an adaptive regularization term Γ is introduced to correct the causal effect to avoid overfitting or misjudgment of causal relationships. The specific regularization formula is expressed as follows:

[0191]

[0192] Among them, α i is the weight of each feature, which is dynamically adjusted based on its historical performance, Γ i is a regularization term used to smooth causal effects and reduce noise interference. Finally, It is the corrected fault diagnosis result, reflecting the fault risk of the air-conditioning unit at the current moment;

[0193] S303: Based on the fault diagnosis, further causal inference is performed to predict possible future faults and improve the preventive maintenance capability of the system. The fault status at future moments is predicted by performing time series analysis on the causal effects at multiple moments in the past.

[0194] Specifically, according to the fault diagnosis results Construct a fault prediction model, which is expressed as follows:

[0195] Assume the fault state at the future time t+Δt is The prediction model makes predictions based on the historical causal effect vectors CE(t), CE(t-Δt), …, CE(t-nΔt):

[0196]

[0197] Among them, g(·) is a time series prediction function based on a multi-layer recurrent neural network (RNN). By looking back and analyzing past causal effects and combining them with current causal effects, the model can output the fault prediction value at the future moment.

[0198] In order to improve the accuracy of prediction, an innovative fault sensitivity adjustment term ΔS(t+Δt) is introduced to dynamically adjust the prediction results to adapt to the behavior characteristics of the air-conditioning unit under special working conditions, which is expressed as follows:

[0199]

[0200] Among them, ΔS(t+Δt) is based on modeling the historical data of specific working conditions, reflecting the operating characteristics of the air-conditioning unit under abnormal working conditions. This adjustment item can better adapt to the complex working conditions of the air-conditioning unit and ensure the robustness of the prediction.

[0201] Finally, when When the set fault threshold τ is exceeded, the system will automatically issue a fault warning signal and transmit the prediction results and current operating data to the central monitoring system for further decision-making analysis. This prediction mechanism ensures that the system can take effective measures before a fault occurs to avoid unnecessary losses caused by the fault.

[0202] S4. Perform local model training on each edge computing node according to the results of the fault diagnosis and fault prediction models, wherein the local model must be trained independently on each edge computing unit. After each edge computing unit completes the local training, it does not share the original data, but only uploads the local model parameters to the central server. The central server aggregates the global model based on the local model, and after completing the aggregation of the global model, the central server sends the updated global model parameters to each edge computing unit. When some edge computing units have data anomalies or model failures, an adaptive anomaly detection and adjustment mechanism is designed to detect and correct the abnormal data.

[0203] The data analysis results of the previous steps are used to achieve dynamic optimization of the detection model. Each edge computing unit independently trains the model locally and collaboratively optimizes the detection model through federated learning technology. This method not only protects data privacy, but also enables the system to continuously upgrade the detection model according to environmental changes and maintain long-term high-efficiency performance.

[0204] Specifically, the specific steps of S4 include:

[0205] S401: In the edge computing unit of the air conditioning unit, the fault diagnosis and prediction model based on causal inference in step S3 is used for local training. The data received and the features extracted by each edge computing unit The environment has unique characteristics and operating conditions. Therefore, the local model must be trained independently on each edge computing unit to adapt it to the specific operating environment.

[0206] The local model is trained independently on each edge computing unit to adapt it to the specific operating environment. For each edge computing unit i, the loss function of the local model (w i ), where w i are the parameters of the model. The goal of local model training is to minimize the error of local data in fault prediction and diagnosis by optimizing the loss function. The loss function can be designed as:

[0207]

[0208] Among them, y i (t) is the actual health status, is the predicted health state, Ω(w i) is a regularization term used to control the complexity of the model and prevent overfitting, λ i is the regularization coefficient, the influence weight of the adjustment term.

[0209] The local training process uses stochastic gradient descent (SGD) or its variants (such as Adam optimizer) to update parameters to gradually minimize the loss function. After multiple iterations, the edge computing unit will obtain a preliminary optimized local model. It can better adapt to the operating data of the current air-conditioning unit.

[0210] S402: Because the operating conditions of each air-conditioning unit are different, each edge computing unit independently trains the model Although it has good adaptability to local data, it may lack generalization ability globally. In order to make the model more universal across the entire air-conditioning unit population, a federated learning framework is used to aggregate and optimize the global model.

[0211] The federated learning framework is used to aggregate and optimize the global model. The specific steps are as follows:

[0212] In the federated learning framework, after completing local training, each edge computing unit does not share the original data, but instead stores the local model parameters. Upload to the central server. The central server aggregates the global model based on these parameters. The global model parameters w global The aggregation is represented as follows:

[0213]

[0214] Where N is the total number of samples, p i is the weight factor, is a local model parameter, usually proportional to the amount of data or computing power of each unit, w global are the global model parameters after aggregation.

[0215] When the global model is aggregated, a regularization term is constructed It is used to modify the aggregation weight of each local model, which is expressed as follows:

[0216]

[0217] Among them, γ is the adjustment parameter, It is an adjustment item designed based on the performance and differences of the local model. is the updated global model;

[0218] After completing the aggregation of the global model, the central server will update the global model parameters It is sent to each edge computing unit. After receiving the updated global model, each unit makes fine adjustments based on local data to maintain high adaptability to the local specific environment. The update formula of the local model is expressed as follows:

[0219]

[0220] Where η is the learning rate, is the direction of gradient descent based on global model parameters and local data, is the updated local model. Through this step, the model is continuously updated and optimized in each cycle, and it has both global knowledge and can adapt to local operating characteristics.

[0221] An adaptive step size adjustment mechanism is constructed to dynamically adjust the learning rate η to adapt to different training stages and data conditions. The specific expression is as follows:

[0222]

[0223] in, is the updated learning rate, is the current learning rate, ζ is the step size adjustment coefficient, is the variance of the loss function, which is used to reflect the instability in model training. By adaptively adjusting the learning rate, the model can better control the oscillation and convergence problems during training;

[0224] S403. During the federated learning process, data anomalies or model failures may occur in some edge computing units, which will affect the aggregation quality of the global model. Therefore, an adaptive anomaly detection and adjustment mechanism (Adaptive Anomaly Detection and Adjustment, AADA) is designed to detect and correct abnormal models to ensure the robustness of federated learning. The specific steps are as follows:

[0225] First, in each edge computing unit, the anomaly score A of the current model is calculated i (t), which is used to monitor the stability of model parameters and training quality. The score calculation formula is as follows:

[0226]

[0227] in, is the Euclidean distance between the local model parameters and the global model parameters, Var(L i (w i (t))) is the variance of the local model loss function, β i is the adjustment coefficient used to balance the contribution of these two parts to the anomaly score.

[0228] For models with high anomaly scores, an adaptive adjustment mechanism is used to correct them. The specific adjustment formula is:

[0229]

[0230] Among them, δ is the adjustment step size, is the adjusted model parameter. This mechanism can effectively reduce the difference between the local model and the global model and avoid the negative impact of individual abnormal models on the global model aggregation.

[0231] In order to further enhance the adaptive ability of the federated learning system, a dynamic anomaly threshold θ(t) is designed so that the system can automatically adjust its tolerance to abnormal models according to changes in the global model, as shown below:

[0232]

[0233] Among them, θ0 is the initial threshold, κ is the adjustment parameter, is the variance of the sum of anomaly scores of all edge computing units. Through this dynamic adjustment, the system can flexibly respond to model changes in different environments during federated learning and ensure the stability and reliability of the global model.

[0234] Through the above-mentioned federated learning-driven model optimization and continuous upgrade mechanism, the entire patented solution can continuously improve the accuracy of fault diagnosis and prediction during the operation of the air-conditioning unit, adapt to different operating environments, and maintain the efficiency and robustness of the system in long-term use. This process is closely connected with the previous steps, and ultimately forms a fault detection and prevention system with strong adaptability and wide applicability.

[0235] S5. According to the results of the final comprehensive feature vector and the fault prediction model, a fault heat map generation algorithm is constructed to visualize the fault, and a multi-mode intelligent interactive interface is designed. Adaptive adjustments are made according to the designed adaptive data focusing algorithm. Finally, according to the user's behavior and feedback during use, the parameters of the visualization interface and the multi-mode intelligent interactive interface are adjusted.

[0236] Based on all the previous steps, the fault diagnosis and prediction results are presented through an intuitive user interface. The system displays fault information in the form of dynamic charts, heat maps, 3D models, etc., and provides an intelligent interactive interface to help users quickly understand and handle faults.

[0237] Specifically, the specific steps of S5 include:

[0238] S501. Through causal inference and federated learning mechanism, the system obtains the multi-dimensional feature vector of the air-conditioning unit. and diagnostic prediction results In order for operators to intuitively understand these complex data and results, effective visualization must be performed.

[0239] According to the final comprehensive feature vector and the results of the fault prediction model, a fault heat map generation algorithm is constructed to visualize the fault; wherein, the fault heat map generation algorithm is to convert the fault diagnosis results Converted into a heat map, it is represented as follows:

[0240]

[0241] Among them, H(x,y,t) is the thermal map value of the air conditioning unit at the position (x,y), α i (t) is the contribution weight of feature i to the fault at time t, M i (x, y) is the component position matrix corresponding to feature i. This heat map can be updated in real time, showing the areas with higher risk of failure in the air conditioning unit, making it easier for operators to quickly locate problems.

[0242] S502. In order to improve operators' understanding and response speed of fault information, a multimodal intelligent interaction interface (MIII) is designed. The interface includes a visual fault heat map, a time series chart, and voice and text interaction modules.

[0243] The time series chart shows the historical and predicted data of each key feature, and the dynamic curve chart shows the changing trend of different sensor data over time. The multi-mode intelligent interactive interface is designed to be adaptively adjusted according to the designed adaptive data focusing algorithm. The adaptive data focusing algorithm automatically adjusts the information displayed on the interface according to the user's operation behavior and the severity of the current fault, as shown below:

[0244]

[0245] Among them, V focus (t) is the feature information of the focused display, β i (t) is the adaptive focusing weight, which is dynamically adjusted according to the user's current operation and system status to ensure that the operator always pays attention to the most critical fault information.

[0246] At the same time, the voice and text interaction module is designed to support users to query the current device status and fault causes through natural language. A natural language fault explanation model is built in the voice and text interaction module to convert complex diagnostic results into understandable language output. The model uses causal inference results to generate a conversational fault description, which is expressed as follows:

[0247]

[0248] Among them, Explanation(t) is the fault explanation generated by natural language, which helps users quickly understand the current status of the system and possible operation steps.

[0249] S503. In order to enhance the intelligence and user experience of the system, a dynamic user feedback mechanism (DUFM) is designed. This mechanism continuously adjusts the parameters of the visualization interface and the interaction model based on the user's behavior and feedback during use.

[0250] Specifically, the system calculates a user satisfaction score US(t) by collecting user behavior data such as clicks, browsing time, and voice commands:

[0251]

[0252] Among them, γ j (t) is the weight of each user behavior, m is the total number of books used, Behavior j (t) is the specific operation record of the user at time t. According to the changing trend of US(t), the system will automatically optimize the interface layout, adjust the data display method, or modify the response strategy of voice interaction to ensure that the user's operation experience in troubleshooting is continuously improved.

[0253] Ultimately, the entire fault visualization and intelligent interaction system, as the last step in the patented solution, not only presents complex fault information to users in an intuitive manner, but also achieves system self-optimization and continuous improvement of user experience through intelligent interaction and dynamic feedback.

[0254] like Figure 2 As shown, an embodiment of the present invention further provides a visual fault detection system for an air conditioning unit, the system comprising:

[0255] The data acquisition module 10 is used to collect multi-source data of the air-conditioning unit and store it in a matrix form, and then pre-process the stored data;

[0256] The fault detection module 11 is used to extract features from the preprocessed data, and adaptively fuse the extracted multi-source features to obtain a comprehensive feature vector, then construct a second-order tensor network to enhance the features of the comprehensive feature vector, and perform dimensionality reduction processing on the comprehensive feature vector after feature enhancement, and then construct a regularized noise suppression model to regularize the comprehensive feature vector after dimensionality reduction processing to obtain a final comprehensive feature vector;

[0257] The data fusion and fault prediction module 12 is used to construct a causal relationship model according to the final comprehensive feature vector, calculate the causal effect according to the causal relationship model, perform multi-level fault diagnosis according to the result of the causal effect, and construct a fault prediction model, perform fault inference based on causal inference according to the result of the fault diagnosis, and predict future faults;

[0258] The strategy execution module 13 is used to perform local model training on each edge computing node according to the results of the fault diagnosis and fault prediction model, wherein the local model must be trained independently on each edge computing unit. After each edge computing unit completes the local training, it does not share the original data, but only uploads the local model parameters to the central server. The central server aggregates the global model based on the local model, and after completing the aggregation of the global model, the central server sends the updated global model parameters to each edge computing unit. When data anomalies or model failures occur in some edge computing units, an adaptive anomaly detection and adjustment mechanism is designed to detect and correct the abnormal data.

[0259] The local control module 14 is used to construct a fault heat map generation algorithm according to the final comprehensive feature vector and the result of the fault prediction model to visualize the fault, design a multi-mode intelligent interactive interface, perform adaptive adjustment according to the designed adaptive data focusing algorithm, and finally adjust the parameters of the visualization interface and the multi-mode intelligent interactive interface according to the user's behavior and feedback during use;

[0260] The adaptive anomaly detection and adjustment mechanism includes:

[0261] In each edge computing unit, calculate the anomaly score A of the current model i (t), the anomaly score A of the current model i (t) is used to monitor the stability of model parameters and training quality, and is expressed as follows:

[0262]

[0263] in, is the local model parameter w i (t) and global model parameters The Euclidean distance of Var(L i (w i (t))) is the local model loss function L i (w i The variance of (t)), β i is the adjustment coefficient, which is used to balance the contribution of these two parts to the anomaly score;

[0264] For the model with the highest anomaly score, an adaptive adjustment mechanism is used to correct it. The specific adjustment formula is:

[0265]

[0266] Among them, δ is the adjustment step size, are the adjusted model parameters, is the local model loss function L i (w i (t)) gradient optimization;

[0267] A dynamic anomaly threshold θ(t) is designed so that the system can automatically adjust its tolerance to abnormal models according to changes in the global model, as shown below:

[0268]

[0269] Among them, θ0 is the initial threshold, κ is the adjustment parameter, is the variance of the sum of anomaly scores of all edge computing units.

[0270] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0271] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a division of logical functions. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the device or unit can be electrical, mechanical or other forms.

[0272] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0273] Although embodiments of the present invention have been shown and described, those skilled in the art will appreciate that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A visual fault detection method for an air conditioning unit, characterized in that: The method comprises: S1, collecting multi-source data of the air-conditioning unit and storing it in a matrix form, and then preprocessing the stored data; S2, extracting features from the preprocessed data, and adaptively fusing the extracted multi-source features to obtain a comprehensive feature vector, then constructing a second-order tensor network to enhance the comprehensive feature vector, and performing dimensionality reduction processing on the comprehensive feature vector after feature enhancement, and then constructing a regularized noise suppression model to regularize the comprehensive feature vector after dimensionality reduction processing to obtain the final comprehensive feature vector; S3. Construct a causal relationship model based on the final comprehensive feature vector, calculate the causal effect based on the causal relationship model, perform multi-level fault diagnosis based on the results of the causal effect, and construct a fault prediction model, perform fault inference based on causal inference based on the results of the fault diagnosis, and predict future faults; S4. Perform local model training on each edge computing node according to the results of the fault diagnosis and fault prediction models, wherein the local model must be trained independently on each edge computing unit. After each edge computing unit completes the local training, it does not share the original data, but only uploads the local model parameters to the central server. The central server aggregates the global model based on the local model, and after completing the aggregation of the global model, the central server sends the updated global model parameters to each edge computing unit. When data anomalies or model failures occur in some edge computing units, an adaptive anomaly detection and adjustment mechanism is designed to detect and correct the abnormal data. S5. Construct a fault heat map generation algorithm based on the final comprehensive feature vector and the results of the fault prediction model to visualize the fault, and design a multi-mode intelligent interactive interface. Adaptive adjustment is performed based on the designed adaptive data focusing algorithm. Finally, adjust the parameters of the visualization interface and the multi-mode intelligent interactive interface based on the user's behavior and feedback during use. The adaptive anomaly detection and adjustment mechanism includes: In each edge computing unit, calculate the anomaly score A of the current model i (t), the anomaly score A of the current model i (t) is used to monitor the stability of model parameters and training quality, and is expressed as follows: in, is the local model parameter w i (t) and global model parameters The Euclidean distance of Var(L i (w i (t))) is the local model loss function L i (w i The variance of (t)), β i is the adjustment coefficient, which is used to balance the contribution of these two parts to the anomaly score; For the model with the highest anomaly score, an adaptive adjustment mechanism is used to correct it. The specific adjustment formula is: Among them, δ is the adjustment step size, are the adjusted model parameters, is the local model loss function L i (w i (t)) gradient optimization; A dynamic anomaly threshold θ(t) is designed so that the system can automatically adjust its tolerance to abnormal models according to changes in the global model, as shown below: Among them, θ0 is the initial threshold, κ is the adjustment parameter of the edge computing node, is the variance of the sum of anomaly scores of all edge computing units, and N is the total number of nodes.

2. A visual fault detection method for an air conditioning unit according to claim 1, characterized in that: The above is stored in matrix form, which is expressed as follows: Assuming the time step is T, the spatial resolution of the visual data is W×H, and the various sensor data are stored in matrix form: D vis ∈R M×W×H×T Among them, D temp , D press , D current , D vib are the time series data matrices of temperature, pressure, current, and vibration sensors, respectively, vis is the visual data, M is the number of cameras, W and H are the image width and height respectively, T is the time step, and R represents a real number; The preprocessing includes time synchronization and data alignment, data standardization and normalization processing, outlier detection and correction processing, and regularization and dimensionality reduction processing.

3. A visual fault detection method for an air conditioning unit according to claim 2, characterized in that: The outlier detection and correction process is represented as follows: in, is the corrected data, z i,t , is normal data, k is the threshold factor for outlier detection, α is an adjustment coefficient used to correct outliers; μ i is the mean; σ i is the standard deviation; The regularization and dimensionality reduction processing is expressed as follows: Assume that the sensor data at time t is vector z t =[z 1,t ,z 2,t ,…,z N,t ], processed by the regularization formula, the regularization formula is expressed as follows: in, is the regularized data, L is the Laplace matrix used to capture the spatial relationship between sensors, and λ is the regularization parameter.

4. A visual fault detection method for an air conditioning unit according to claim 1, characterized in that: The specific steps of S2 are as follows: S21. For physical sensor data, a one-dimensional convolutional neural network is used to process time series features, and each type of data generates preliminary feature vectors of temperature, pressure, current, and vibration data; for visual data, a two-dimensional convolutional neural network is used to process the spatial features of the image and generate preliminary visual feature vectors; S22, normalizing the preliminary feature representations of the temperature, pressure, current, and vibration data and the preliminary visual feature vectors to obtain feature vectors and visual feature vectors of the temperature, pressure, current, and vibration data, and designing an adaptive weighted multi-source feature fusion model to dynamically adjust the fusion weights of different feature vectors to obtain a first comprehensive feature vector; S23, construct a second-order tensor network to enhance the first comprehensive feature vector, and then perform dimensionality reduction processing to obtain a second comprehensive feature vector, which is expressed as follows: in, is the outer product operation, F enhanced (t) is the second comprehensive eigenvector, F fused (t) is the first comprehensive eigenvector; S24, constructing a regularized noise suppression model to perform regularization processing on the second comprehensive feature vector, which is expressed as follows: Where L is the Laplace matrix, λ is the regularization parameter, is the final comprehensive feature vector.

5. A visual fault detection method for an air conditioning unit according to claim 1, characterized in that: The steps of constructing the causal relationship model are as follows: First, a causal relationship model is constructed based on the final comprehensive feature vector, where each feature is a node in the graph, and the edges between the nodes represent the causal relationship between the features. The causal relationship is represented by the causal relationship matrix C as follows: Among them, C ij is the strength of the causal relationship between feature i and feature j, Granger(·) is a function based on the Granger causality test, which is used to detect the causal relationship between time series. is the final comprehensive feature vector of feature i, is the final comprehensive feature vector of feature j; According to the specific application scenario of the air-conditioning unit, a regular term Ω based on physical constraints is designed in the causal relationship matrix C to enhance the interpretability of the causal relationship, which is expressed as follows: in, The enhanced causal relationship matrix element reflects the causal influence strength between feature i and feature j; Ω ij is a correction term based on the specific physical structure of the air-conditioning unit or the sensor location constraints, C ij is the causal influence strength between feature i and feature j, and θ is the parameter that adjusts the causal relationship matrix.

6. A visual fault detection method for an air conditioning unit according to claim 5, characterized in that: The calculation method of the causal effect is expressed as follows: Among them, CE i (t) is the causal effect size of feature i on health state y(t), It is the enhanced causal relationship matrix element, reflecting the causal influence strength between feature i and feature j; by calculating CE i (t), determine the contribution of each feature to the system health status at the current moment and identify the root cause of the fault; The multi-level fault diagnosis specifically includes: Sort the K features with the largest causal effects and use them as key features for preliminary fault diagnosis; An adaptive regularization term Γ is introduced to correct the causal effect, which is expressed as follows: Among them, α i is the weight of each feature, which is dynamically adjusted based on its historical performance, Γ i is a regularization term used to smooth causal effects and reduce noise interference; This is the corrected fault diagnosis result.

7. A visual fault detection method for an air conditioning unit according to claim 6, characterized in that: According to the fault diagnosis results Construct a fault prediction model, which is expressed as follows: Assume the fault state at the future time t+Δt is The prediction model is based on the historical causal effect vectors CE(t), CE(t-Δt), …, CE(t-nΔt), which are expressed as follows: Among them, g(·) is a time series prediction function based on a multi-layer recurrent neural network. By backtracking and analyzing past causal effects and combining current causal effects, the model outputs the fault prediction value at the future moment. n is the nth change moment; The fault sensitivity adjustment term ΔS(t+Δt) is introduced to dynamically adjust the prediction results to adapt to the behavior characteristics of the air-conditioning unit under special working conditions, which is expressed as follows: Among them, ΔS(t+Δt) is based on modeling the historical data of specific working conditions, reflecting the operating characteristics of the air-conditioning unit under abnormal conditions. It is the forecast result after dynamic adjustment; according to The value of exceeds the set fault threshold τ. When the value exceeds the set fault threshold τ, the system will automatically issue a fault warning signal and transmit the prediction results and current operating data to the central monitoring system for decision analysis.

8. A visual fault detection method for an air conditioning unit according to claim 1, characterized in that: When the global model is aggregated, a regularization term is constructed It is used to modify the aggregation weight of each local model, which is expressed as follows: Among them, γ is the adjustment parameter of the global model, It is an adjustment item designed based on the performance and differences of the local model. is the updated global model, M′ is the total number of samples, p i is the weight factor, are local model parameters; After completing the aggregation of the global model, the central server will update the global model parameters It is sent to each edge computing unit. After receiving the updated global model, each unit performs fine-tuning based on local data. The update formula of the local model is expressed as follows: Where η is the learning rate, is the direction of gradient descent based on global model parameters and local data, It is a newer local model; An adaptive step size adjustment mechanism is constructed to dynamically adjust the learning rate η to adapt to different training stages and data conditions. The specific expression is as follows: in, is the updated learning rate, is the current learning rate, ζ is the step size adjustment coefficient, is the variance of the loss function, which is used to reflect the instability in model training.

9. A visual fault detection method for an air conditioning unit according to claim 1, characterized in that: The step S5 specifically includes: S51, construct a fault heat map generation algorithm based on the final comprehensive feature vector and the results of the fault prediction model to visualize the fault; wherein, the fault heat map generation algorithm is to convert the fault diagnosis results Converted into a heat map, it is represented as follows: Among them, H(x,y,t) is the thermal map value of the air conditioning unit at the position (x,y), α i (t) is the contribution weight of feature i to the fault at time t, M i (x,y) is the component position matrix corresponding to feature i; S52. Design a multi-mode intelligent interactive interface to perform adaptive adjustment according to the designed adaptive data focusing algorithm. The adaptive data focusing algorithm automatically adjusts the information displayed on the interface according to the user's operation behavior and the severity of the current fault. At the same time, design a voice and text interaction module to support users to query the current device status and fault cause through natural language, and build a natural language fault explanation model in the voice and text interaction module to convert complex diagnostic results into understandable language output; S53. Design a user satisfaction score based on the user's behavior and feedback during use, and adjust the parameters of the visualization interface and the multi-modal intelligent interaction interface based on the user satisfaction score; wherein the user satisfaction score is expressed as follows: Among them, γ j (t) is the weight of each user behavior, m is the total number of users, Behavior j (t) is the specific operation record of the user at time t.

10. Based on a visual fault detection system for air conditioning units, characterized in that: The system comprises: A data acquisition module is used to collect multi-source data of the air-conditioning unit and store it in a matrix form, and then pre-process the stored data; The data fusion and feature extraction module is used to extract features from the preprocessed data, and adaptively fuse the extracted multi-source features to obtain a comprehensive feature vector, then build a second-order tensor network to enhance the comprehensive feature vector, and perform dimensionality reduction on the comprehensive feature vector after feature enhancement, and then build a regularized noise suppression model to regularize the comprehensive feature vector after dimensionality reduction to obtain the final comprehensive feature vector; The fault diagnosis and prediction module is used to build a causal relationship model based on the final comprehensive feature vector, calculate the causal effect based on the causal relationship model, perform multi-level fault diagnosis based on the results of the causal effect, and build a fault prediction model, perform fault inference based on causal inference based on the results of the fault diagnosis, and predict future faults; The model optimization and upgrade module is used to perform local model training on each edge computing node according to the results of the fault diagnosis and fault prediction models. The local model must be trained independently on each edge computing unit. After each edge computing unit completes the local training, it does not share the original data, but only uploads the local model parameters to the central server. The central server aggregates the global model based on the local model. After the central server completes the aggregation of the global model, it sends the updated global model parameters to each edge computing unit. When some edge computing units have data anomalies or model failures, an adaptive anomaly detection and adjustment mechanism is designed to detect and correct the abnormal data. The fault visualization and intelligent interaction module is used to construct a fault heat map generation algorithm based on the final comprehensive feature vector and the results of the fault prediction model to visualize the fault, design a multi-mode intelligent interaction interface, perform adaptive adjustments based on the designed adaptive data focusing algorithm, and finally adjust the parameters of the visualization interface and the multi-mode intelligent interaction interface based on the user's behavior and feedback during use; The adaptive anomaly detection and adjustment mechanism includes: In each edge computing unit, calculate the anomaly score A of the current model i (t), the anomaly score A of the current model i (t) is used to monitor the stability of model parameters and training quality, and is expressed as follows: in, is the local model parameter w i (t) and global model parameters The Euclidean distance of Var(L i (w i (t))) is the local model loss function L i (w i The variance of (t)), β i is the adjustment coefficient, which is used to balance the contribution of these two parts to the anomaly score; For the model with the highest anomaly score, an adaptive adjustment mechanism is used to correct it. The specific adjustment formula is: Among them, δ is the adjustment step size, are the adjusted model parameters, is the local model loss function L i (w i (t)) gradient optimization; A dynamic anomaly threshold θ(t) is designed so that the system can automatically adjust its tolerance to abnormal models according to changes in the global model, as shown below: Among them, θ0 is the initial threshold, κ is the adjustment parameter of the edge computing node, is the variance of the sum of anomaly scores of all edge computing units, and N is the total number of nodes.

Images