A building equipment monitoring fault point identification method

Abstract

The application provides a building equipment monitoring fault point identification method, and relates to the technical field of intelligent building operation and maintenance, which realizes real-time collection of equipment operation data, environment data and video stream data through an edge computing node, generates a state feature vector through multimodal fusion processing, extracts a dynamic trend to calculate a composite feature index, identifies equipment abnormal behavior by using a pre-trained abnormality detection model, constructs an equipment correlation graph (including physical connection and data coupling relationship) when the abnormal confidence exceeds a threshold, locates a root fault point in combination with a historical fault library, generates a hierarchical alarm signal and pushes it to an operation and maintenance terminal, and finally updates the abnormality detection model (through retraining of a negative sample library) and the correlation graph (dynamic adjustment of edge weight) based on operation and maintenance feedback data. The application solves the problems of lagging fault identification, high false alarm rate, difficult positioning and low maintenance efficiency in the traditional monitoring mode, and realizes real-time, accuracy and adaptability of building equipment operation and maintenance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent building operation and maintenance, and in particular to a building equipment monitoring fault point identification method. BACKGROUND

[0002] A building refers to a multi-story or high-rise building, covering residential, commercial and office purposes. Its functions include equipment management, security monitoring and automation systems, which can be classified as traditional, modern and intelligent office buildings. Evaluation standards include building area, degree of intelligentization and supporting facilities. Traditional building monitoring equipment mostly relies on manual patrol methods, which have the following problems: 1. Fault identification lag: Traditional monitoring relies on manual inspection or fixed threshold alarms, which cannot capture device abnormalities (such as sensor drift and communication interruption) in real time, resulting in delayed fault discovery. 2. High false alarm rate: Simple threshold judgments based on rules (such as temperature exceeding 30°C for alarm) are easily disturbed by the environment, resulting in a large number of invalid alarms and increasing maintenance costs. 3. Difficulty in positioning: There is strong correlation between devices (such as air conditioner failure caused by water pump abnormalities), and traditional systems lack multi-dimensional data analysis capabilities, making it difficult to quickly locate the root cause. 4. Low maintenance efficiency: Manual troubleshooting requires detecting each device, with an average repair time (MTTR) of several hours, affecting normal operation of the building. In the prior art, patent CN119807990A of Beijing University of Aeronautics and Astronautics discloses a multi-modal data fusion method based on a large model, which includes: converting the multi-modal data to be processed into embedding vectors corresponding to each modality; based on a pre-set CNN model, extracting specific features of the embedding vectors corresponding to each modality to obtain specific features; based on the specific features, determining the weight of the specific features through a pre-set reinforcement learning algorithm; based on the weight of the specific features, adding a gate control to extract features from the specific features to obtain shared features, which are the fusion results of the multi-modal data to be processed. However, the application scenario is target retrieval and tracking, which is not directly related to building monitoring fault diagnosis. Therefore, a building equipment monitoring fault point identification method is proposed to solve the above problems. SUMMARY

[0003] Therefore, the purpose of the present application is to provide a building equipment monitoring fault point identification method to at least solve the above problems.

[0004] The technical solution adopted by the present application is as follows:

[0005] A building monitoring equipment fault point identification method, the method comprising the following steps:

[0006] Step S1: collecting real-time operation data, environmental data and video stream data of building equipment at the edge computing node, the operation data including voltage and current signals, and the environmental data including temperature and humidity;

[0007] Step S2: performing multi-modal fusion processing on the operation data, environmental data and video stream data to generate a device state feature vector;

[0008] Step S3: extracting the dynamic change trend of the feature vector based on a sliding time window and calculating a composite feature index;

[0009] Step S4: analyzing the composite feature index through a pre-trained anomaly detection model to identify device abnormal behavior and generate an abnormal confidence;

[0010] Step S5: when the abnormal confidence exceeds a preset threshold, constructing a device correlation graph, the device correlation graph including device physical connection relationship and data coupling relationship;

[0011] Step S6: traversing the device correlation graph in combination with a historical fault database to locate a root fault point causing the anomaly;

[0012] Step S7: generating a hierarchical alarm signal according to the fault point positioning result and pushing it to an operation and maintenance terminal;

[0013] Step S8: updating the anomaly detection model and the device correlation graph based on operation and maintenance feedback data.

[0014] Further, the video stream data processing in step S1 includes:

[0015] S11, dividing a dynamic monitoring area in a device working area to extract continuous gray scale images at a fixed frame rate;

[0016] S12, performing optical flow field analysis on adjacent frame images to obtain displacement vectors of each pixel point;

[0017] S13, calculating a device vibration intensity index according to the displacement vectors, specifically including: taking the average value of the module length of the displacement vectors as a basic vibration amount;

[0018] superimposing a compensation factor of the area change rate of the monitoring area to generate a final vibration intensity;

[0019] S14, spatiotemporally aligning the vibration intensity with the operation data corresponding to the time stamp to form a four-element synchronous data group: vibration intensity value, voltage instantaneous value, current effective value and environmental temperature value.

[0020] Further, the multi-modal fusion processing of step S2 specifically includes:

[0021] S21. The timestamps of the running data and the environmental data are aligned by the dynamic time warping algorithm. The alignment degree A is calculated to meet the preset conditions. The maximum allowable time delay threshold between the sensor data timestamp and the environmental data timestamp is 500ms. Time deviation compensation is performed by the S-shaped compensation function.

[0022] S22. The vibration intensity extracted from the video stream data is Gaussian normalized, where the normalization parameter is determined based on the historical 30-day mean and standard deviation of vibration intensity.

[0023] S23. The voltage and current data, temperature data, and normalized vibration intensity are fused in the latent space to generate the equipment state feature vector F. Vector splicing is used during fusion, and the weight coefficients of each data source are dynamically adjusted based on the voltage fluctuation rate and the proportion of vibration intensity.

[0024] Furthermore, the calculation of the composite feature index in step S3 includes:

[0025] S31. Perform wavelet packet decomposition on the voltage and current signals, extract the energy distribution within the specified frequency band of 0.5-1.5Hz, and calculate the electrical anomaly index E;

[0026] S32. Calculate the environmental stability coefficient S based on the temperature and humidity change rate, where the weighting factor for the humidity change rate is 0.33.

[0027] S33. Integrate the environmental stability coefficient and the electrical anomaly index to generate the equipment health H, where the environmental weight is 0.4 and the electrical weight is 0.6. When the equipment health H is lower than 0.2, anomaly detection is triggered.

[0028] Furthermore, the anomaly detection model in step S4 adopts a structure combining a temporal convolutional network and a self-attention mechanism, and its processing flow includes:

[0029] S41. Input a composite feature index matrix of N consecutive time windows, where the number of time windows N=20 and the feature dimension D=5;

[0030] S42. Extract local temporal features through a temporal convolutional network, using a dilated convolutional structure and the ReLU activation function;

[0031] S43. Input the output of the temporal convolutional network into the attention layer and calculate the attention weights for each time window;

[0032] S44. Output the anomaly confidence level C through the fully connected layer. When C is greater than 0.85, trigger the fault location process.

[0033] Furthermore, the device association map construction in step S5 includes:

[0034] S51, parse the building equipment topology configuration file to generate a matrix representing the physical connection relationship between devices;

[0035] S52, calculate the coupling weight of the historical operation data between devices by the mutual information algorithm;

[0036] S53, fuse the physical connection relationship and the data coupling weight to construct a weighted adjacency matrix, and the edge weight in the graph represents the correlation strength between devices.

[0037] Further, the step S6 of fault point positioning comprises:

[0038] S61, taking the device with an abnormal confidence exceeding 0.85 as a starting node, performing a random walk algorithm on the device correlation graph, and determining the state transition probability based on the edge weight;

[0039] S62, calculating the fault propagation probability of each node after T=100 steps of walking;

[0040] S63, screening the nodes with a fault propagation probability exceeding 0.7 to form a candidate fault point set;

[0041] S64, calculating the matching degree of the candidate node and the historical fault feature library, and selecting the node with a matching degree exceeding 0.9 and the highest fault propagation probability as the root fault point.

[0042] Further, the step S7 of hierarchical alarm signal comprises:

[0043] First-level alarm: triggered when the device health degree H is lower than 0.15 or the fault propagation probability is greater than 0.95, and the key device fault instruction is pushed;

[0044] Second-level alarm: triggered when the fault propagation probability is between 0.6 and 0.95 and the number of associated devices is not less than 3, and the chain failure risk is prompted;

[0045] Third-level alarm: triggered when the decline rate of the device health degree in the last 5 monitoring cycles exceeds 10%, and the performance degradation warning is triggered.

[0046] Further, the step S8 of updating operation comprises:

[0047] S81, when receiving the operation and maintenance feedback confirming false alarm, adding the feature vector at this moment to the negative sample library, and retraining the anomaly detection model based on the cross-entropy loss function and the regularization term;

[0048] S82, correcting the edge weight in the device correlation graph according to the maintenance record, and introducing a time decay factor and a maintenance enhancement coefficient to reflect the change of the device correlation relationship.

[0049] Compared with the prior art, the beneficial effects of the present application are:

[0050] 1. Through edge computing and multi-modal data fusion technology, real-time capture of equipment abnormalities and generation of feature vectors, combined with dynamic time warping and wavelet packet decomposition algorithm, significantly reduce fault recognition lag and false alarm rate.

[0051] 2. The root fault point can be quickly traced through the constructed device association graph, solving the positioning difficulty problem in traditional methods, and further shortening the maintenance response time.

[0052] 3. Based on the closed-loop updating mechanism (model retraining and graph edge weight dynamic adjustment) of operation and maintenance feedback, the system can continuously adapt to dynamic factors such as equipment aging and environmental changes, and the maintenance efficiency is significantly improved.

[0053] 4. The hierarchical alarm signal is pushed to the corresponding operation and maintenance level according to the fault severity, avoiding resource waste and reducing operation and maintenance cost. BRIEF DESCRIPTION OF DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only preferred embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0055] Figure 1 It is a whole flowchart of the building monitoring equipment fault point identification method proposed in the embodiment of the present application. DETAILED DESCRIPTION

[0056] The principles and characteristics of the present application are described below in conjunction with the drawings, and the listed embodiments are only used to explain the present application, and are not used to limit the scope of the present application.

[0057] Referring to Figure 1 , the present application provides a building monitoring equipment fault point identification method, which comprises the following steps:

[0058] Step S1: Real-time collection of operation data, environment data and video stream data of building equipment at edge computing node, the operation data including voltage, current signal, and the environment data including temperature and humidity;

[0059] Step S2: Multi-modal fusion processing of the operation data, environment data and video stream data to generate device state feature vector;

[0060] Step S3: Based on the sliding time window, the dynamic change trend of the feature vector is extracted, and the composite feature index is calculated;

[0061] Step S4: Analyzing the composite feature indicators by the pre-trained anomaly detection model to identify device abnormal behaviors and generate anomaly confidence;

[0062] Step S5: When the anomaly confidence exceeds a preset threshold, constructing a device correlation graph containing device physical connection relationships and data coupling relationships;

[0063] Step S6: Locating the root fault point causing the anomaly by traversing the device correlation graph in combination with a historical fault database;

[0064] Step S7: Generating a hierarchical alarm signal according to the fault point positioning result and pushing it to an operation and maintenance terminal;

[0065] Step S8: Updating the anomaly detection model and the device correlation graph based on operation and maintenance feedback data.

[0066] Exemplarily, the application realizes accurate fault positioning through multi-source data fusion and dynamic correlation analysis. In step S1, data collection can be performed. Real-time operation data, environmental data, and video stream data can be acquired at the edge computing node close to the equipment end. The operation data includes voltage / current signal data, which can directly reflect the electrical state of the equipment. The environmental data includes temperature and humidity data, which can reveal changes in the working conditions of the equipment. The video stream data can provide visual verification of the physical state of the equipment. Edge computing can reduce the transmission delay of the cloud. Multi-source data can provide a stereoscopic basis for fault analysis. In step S2, multi-modal fusion can be performed. Different modal data (numerical operation parameters, time-series environmental data, and image-type video stream) can be integrated into a unified dimension device state feature vector through feature extraction and alignment technology. Thus, data islands can be eliminated, and cross-domain feature correlation can be established. For example, the synchronization of device vibration and current anomalies can be confirmed through video. In step S3, dynamic feature extraction can be performed. A sliding time window is used to capture the trend of the feature vector over time, and a composite feature index of the statistical quantity is calculated. Thus, static features can be converted into dynamic behavior descriptions. For example, the current mutation mode in the starting stage of the equipment can be identified. In step S4, anomaly detection can be performed. A pre-trained machine learning model can be used to identify the pattern of the composite feature index, and an anomaly confidence in the interval [0, 1] can be output. Through the pre-training mechanism, the model can ensure prior knowledge of common fault patterns, and the confidence can quantify the severity of the anomaly. In step S5, a correlation graph can be constructed. When the confidence exceeds a preset threshold, a device correlation graph containing two layers of relationships can be automatically generated: physical connection (such as pipeline series connection, circuit parallel connection), and data coupling (such as vibration data of sensor A affecting temperature reading of sensor B). The graph visualizes the direct and indirect correlations between devices, and can provide a structured framework for root cause analysis. In step S6, root cause positioning can be performed. Through a graph traversal algorithm, similar cases in the historical fault library can be matched, and historical fault nodes that frequently appear in the correlation graph can be preferentially checked. Historical experience can be used to quickly converge the scope of the investigation. For example, when an air conditioning unit is abnormal, the pump machine that shares the cooling water pipe with it can be preferentially checked. In step S7, hierarchical alarm can be generated. According to the fault impact range (such as single machine fault / regional cascading fault), three-level alarm signals can be generated: first-level alarm (red): immediate shutdown and maintenance, second-level alarm (orange): processing within 2 hours, and third-level alarm (yellow): investigation within 24 hours. Thus, the allocation of operation and maintenance resources can be optimized, and over-reaction can be avoided. In step S8, model iteration can be performed. The fault results confirmed by the operation and maintenance personnel (such as "cooling water pump bearing damage") can be fed back to the historical fault library, and the parameters of the anomaly detection model and the weights of the correlation graph can be updated synchronously.

[0067] Thus, a "detection-positioning-feedback-optimization" closed loop can be formed, so that the system gradually adapts to dynamic changes such as equipment aging. The method in the application can break through the limitations of traditional single-point monitoring, realize real-time response through edge computing, reveal implicit correlations through graph reasoning, and ultimately build a fault diagnosis system with self-learning ability, significantly improving the accuracy and efficiency of building equipment operation and maintenance.

[0068] The video stream data processing in step S1 includes:

[0069] S11, a dynamic monitoring area is demarcated in the equipment working area, and continuous gray scale images are extracted at a fixed frame rate;

[0070] S12, optical flow field analysis is performed on adjacent frame images to obtain displacement vectors of each pixel point;

[0071] S13, the displacement vectors are used to calculate the equipment vibration intensity index, specifically including: taking the average value of the displacement vector length as the basic vibration amount;

[0072] A compensation factor of the area change rate of the monitoring area is superimposed to generate the final vibration intensity;

[0073] S14, the vibration intensity is spatiotemporally aligned with the operation data corresponding to the time stamp to form a four-element synchronous data group: vibration intensity value, voltage instantaneous value, current effective value, and environment temperature value.

[0074] For example, taking building air conditioning unit vibration monitoring as an example, when the edge node collects equipment video stream, a 200x200 pixel monitoring frame is demarcated in the air conditioner outdoor unit working area, covering the compressor vibration sensitive area, 25 frames of gray scale images are extracted per second, the Lucas-Kanade algorithm is used to calculate the pixel displacement between adjacent frames to generate a displacement vector field containing horizontal / vertical components, and the typical displacement amount is in the range of 0-5 pixels. Then, the vibration intensity is calculated, the basic vibration amount is taken as the average value of the length of all displacement vectors in the monitoring area (such as 3.2 pixels), the compensation factor is superimposed when the compressor vibration causes the area change rate of the monitoring frame to exceed 15%, the change rate proportion (such as 18%) is superimposed, the final vibration intensity is 3.2x(1+0.18)=3.78 pixel value, and finally the vibration intensity 3.78 at time t is combined with the voltage 220V (instantaneous value), current 8.5A (effective value), and temperature 35℃ (environmental value) collected at the same time to form a four-element data group, which is used for subsequent multi-modal fusion analysis. Through the spatiotemporal correlation of video vibration characteristics and electrical parameters, the problem of easy missed reporting in traditional single parameter monitoring can be effectively solved, and the fault recognition accuracy can be further improved in the early detection of air conditioner compressor bearing wear.

[0075] The multi-modal fusion processing of step S2 specifically includes:

[0076] S21, align the timestamps of the running data and the environmental data by a dynamic time warping algorithm, and the alignment degree A satisfies:

[0077]

[0078] wherein, is the sensor data timestamp, is the environmental data timestamp, is an S-shaped compensation function: , is a maximum allowed delay threshold, and the value is 500 ms;

[0079] S22, Gaussian normalization is performed on the vibration intensity η extracted from the video stream data:

[0080]

[0081] wherein, μ is the historical 30-day vibration intensity mean, and σ is the historical data standard deviation;

[0082] S23, the physical data and the video features are fused in the hidden space to generate a device state feature vector F:

[0083] F = α × [voltage, current] || β × temperature || γ × normalized vibration intensity

[0084] wherein, || represents a vector splicing operation, α, β, and γ are dynamic weight coefficients, and satisfy: α = 0.7 - 0.3 × (voltage fluctuation rate), and γ = 0.2 + 0.5 × (vibration intensity proportion).

[0085] For example, by time alignment, data standardization and dynamic weighted fusion, a feature vector that can accurately represent the device running state is constructed. On the time synchronization mechanism of multiple sources of data, the dynamic time warping algorithm can be used to align the timestamps of the voltage / current running data and the temperature / humidity environmental data. The alignment degree A is calculated by the S-shaped compensation function to ensure that the time deviation of the two types of data does not exceed the 500 ms threshold. For example, when the voltage data collection is delayed by 300 ms, the system realizes accurate alignment by stretching the environmental data time axis, avoiding state misjudgment caused by time misalignment. In the vibration feature standardization processing, the vibration intensity η extracted from the video stream is subjected to Gaussian normalization:

[0086]

[0087] where μ is the historical 30-day vibration intensity average, and σ is the historical data standard deviation. Assuming that the historical average vibration intensity of a certain device is 0.5 g, and the standard deviation is 0.1 g, when the real-time measured vibration is 0.7 g, the normalized value is 2.0σ, which is significantly higher than the normal threshold, triggering an abnormal warning; in the dynamic weight fusion strategy, the construction of the feature vector F uses variable weight coefficients:

[0088] F = α × [voltage, current] || β × temperature || γ × normalized vibration intensity

[0089] The weight α decreases as the voltage fluctuation rate increases (e.g., α decreases to 0.4 when the fluctuation rate exceeds 10%), and γ increases as the vibration intensity proportion increases (e.g., γ increases to 0.7 when the vibration contribution exceeds 30%). This dynamic adjustment mechanism makes the feature vector focus on voltage / current data during electrical abnormalities and emphasizes vibration characteristics during mechanical failures. Through a refined data processing process, high-dimensional and strongly correlated feature inputs are provided for subsequent anomaly detection models, significantly improving the recognition accuracy in complex fault scenarios.

[0090] The composite feature index calculation of step S3 includes:

[0091] S31, wavelet packet decomposition is performed on the voltage and current signals to extract the jth layer frequency band energy entropy: electrical anomaly index

[0092] wherein, represents the energy proportion of the jth layer kth frequency band, is the wavelet packet decomposition frequency band energy value, and the frequency band range is set to 0.5-1.5 Hz;

[0093] S32, the environmental stability coefficient is calculated based on the temperature and humidity gradient change:

[0094] Environmental stability coefficient S = exp(-|ΔT / Δt| - λ|ΔH / Δt|)

[0095] wherein, ΔT / Δt is the temperature change rate, ΔH / Δt is the humidity change rate, and λ = 0.33 represents the humidity weight factor;

[0096] S33, the device health degree is generated by fusing electrical and environmental features:

[0097] Device health degree H = 1 - [ ×(1-S) + ×E]

[0098] wherein, =0.4, =0.6 respectively represent the environmental and electrical weights, H ∈ [0, 1], and H <0.2 triggers step 4 of the anomaly detection.

[0099] For example, by constructing a device health evaluation system through multi-dimensional feature fusion, in step S31, for the electrical abnormality index extraction, the voltage / current signal can be analyzed in time and frequency based on wavelet packet decomposition technology, and the energy entropy is calculated in the frequency band of 0.5-1.5 Hz (corresponding to the characteristic frequency band when the device starts / operates). For example, when a certain frequency converter is running, the energy proportion of this frequency band decreases from 85% in the normal state to 40%, and the electrical abnormality index E=0.32 is calculated by formula, which is significantly lower than the normal threshold 0.6, triggering preliminary abnormal warning; in step S32, for the environment stability evaluation, the temperature and humidity data can be collected to calculate the environment stability coefficient, for example, the temperature in a certain computer room rises by 3°C (ΔT / Δt=1.5°C / h) in 2 hours, and the humidity decreases by 15% (ΔH / Δt=-7.5% / h), then S=exp(-1.5-0.33×7.5)=0.18, which indicates that the environmental mutation may aggravate the risk of device failure; in step S33, for the health degree fusion diagnosis, the electrical characteristics (E=0.32) and environmental characteristics (S=0.18) can be weighted and fused to calculate the device health degree H=1-(0.4×0.82+0.6×0.68)=1-0.73=0.27, when H<0.2 (the measured value 0.27 is close to the threshold), the system automatically reduces the preset threshold of the abnormal detection model to enhance the sensitivity to potential faults. By establishing an electrical-environmental double-factor evaluation model, the device health degree calculation error rate is further reduced compared with the traditional single-factor method.

[0100] The abnormal detection model of step S4 adopts a structure combining time sequence convolution network TCN and self-attention mechanism, and its processing flow includes:

[0101] S41, input the composite feature index matrix of N continuous time windows , wherein N=20 is the number of time windows, D=5 is the feature dimension, including device health degree H, electrical abnormality index E, environment stability coefficient S, vibration intensity η, and temperature and humidity gradient;

[0102] S42, extract local time sequence features through TCN layer, and the dilated convolution output is:

[0103]

[0104] , wherein, is the layer index, is the dilated factor, K=3 is the convolution kernel size, σ is the ReLU activation function, is the learnable weight;

[0105] S43, input the TCN output into the self-attention layer to calculate the attention weight:

[0106]

[0107] wherein Q, K, V are query, key, value matrix, is feature dimension;

[0108] S44, finally output abnormal confidence C∈[0,1] through the full connection layer:

[0109]

[0110] When C>0.85, trigger fault positioning process.

[0111] For example, the abnormality detection model adopts the structure of combining time series convolution network TCN with self-attention mechanism. The model realizes accurate identification of device abnormalities through four-layer architecture of multi-dimensional data input layer, local time series feature extraction layer, global dependence modeling layer and confidence decision layer. For the multi-dimensional data input layer, it can receive a composite feature index matrix of 20 continuous time windows, each window containing 5 core feature dimensions, including device health degree H (reflecting device aging state), electrical abnormality index E (monitoring circuit stability), environmental stability coefficient S (evaluating environmental factors such as temperature and humidity), vibration intensity η (capturing mechanical vibration abnormalities) and temperature and humidity gradient (detecting environmental mutations), so as to ensure comprehensive monitoring of the device state. For the local time series feature extraction layer, the expansion convolution structure of TCN can be used to capture the local dependence relationship of time series data through layer-by-layer expansion of the receptive field. For example, the first layer of convolution kernel size is 3 and the expansion factor is 1, which can capture the short-term changes of adjacent 3 time windows. The expansion factor of the next layer is increased to 2, which expands the time series coverage range while maintaining the parameter efficiency. The ReLU activation function and the learning mechanism of nonlinear weight enable the model to adaptively extract periodic fluctuations and sudden abnormal patterns in device operation. For the global dependence modeling layer, the time series features output by TCN can be input into the self-attention mechanism to generate attention weight distribution through query-key-value matrix operation. The global dependence modeling layer can dynamically focus on key time windows. For example, when the vibration intensity abnormally rises in a certain time period, the self-attention mechanism will give higher weight to that time period, thereby strengthening the expression of abnormal signals in the overall features. For the confidence decision layer, the fully connected layer maps the fused time-space features to the abnormal confidence in the interval [0, 1]. When the confidence exceeds the threshold of 0.85, the system determines that the device has significant abnormalities that need to be intervened, and triggers the subsequent fault positioning process. The threshold setting balances the false positive rate and the false negative rate, ensuring efficient allocation of operation and maintenance resources. By combining the local time series modeling capability of TCN with the global dependence capturing capability of self-attention, the sensitivity of convolutional network to periodic patterns is retained, and the key abnormal period is highlighted through the attention mechanism, significantly improving the accuracy of device abnormality detection under complex working conditions.

[0112] The device association graph construction of the step S5 includes:

[0113] S51, parse the building device topology configuration file to generate a physical connection matrix:

[0114] ,

[0115] Where M is the total number of devices, and if there is a physical connection between devices i and j ;

[0116] S52, calculate the device data coupling weight by mutual information algorithm:

[0117]

[0118] Where is the historical operation data of devices i, j, and p(x, y) is the joint probability distribution;

[0119] S53, fuse the physical and data correlation to construct a weighted adjacency matrix:

[0120] ,

[0121] The graph node represents the device entity, and the edge weight A[i, j] represents the correlation strength.

[0122] For example, the device association graph realizes quantitative expression of the association relationship between devices through a three-layer architecture of a physical topology analysis layer, a data coupling analysis layer, and an association fusion modeling layer. For the physical topology analysis layer, a physical connection matrix can be generated based on a building device configuration file. For example, if a layer includes an air conditioning unit (device A), a water pump (device B), and a frequency converter (device C), and A and B are connected through a pipeline and B and C are connected through a cable, then the A-B and B-C positions in the matrix are marked as 1, and the positions between other devices that are not directly connected are marked as 0. The binary matrix clearly depicts the physical layout relationship between devices. For the data coupling analysis layer, the statistical association of device operation data can be calculated using a mutual information algorithm. For example, the compressor current of the air conditioning unit (device A) and the outlet pressure of the water pump (device B) can have a coordinated change mode. When the compressor load increases, the water pump pressure also increases. The higher the mutual information value, the closer the data association between the two devices. The quantitative relationship can reveal the implicit dependence that is not reflected in the physical connection. For the association fusion modeling layer, the physical connection matrix and the data mutual information value can be weighted and fused to generate a weighted adjacency matrix. For example, device A and B have a physical connection (physical weight = 1) and a mutual information value of 0.8 (data weight = 0.8), so the final edge weight A[A,B] = 1 x 0.8 = 0.8. The edge weight A[B,C] of device B and C is 0.3, although they are physically connected (weight = 1). This fusion mechanism ensures that the graph reflects the actual wiring relationship and highlights the coordinated features of the operation data. The present application combines the “hard connection” of the physical topology with the “soft association” of the operation data, and expresses the association strength between devices through quantifiable edge weights. Compared with the traditional graph that only relies on physical connection, this model can more accurately depict the actual influence relationship between devices in a complex system and provide a more reliable path search basis for subsequent fault location.

[0123] The step S6 of fault point positioning includes:

[0124] S61, taking a device with an abnormal confidence C > 0.85 as a starting node, and performing a random walk algorithm on the device association graph A, with a state transition probability of:

[0125] ;

[0126] S62, calculating the fault propagation probability of each node after T = 100 steps of walking:

[0127]

[0128] wherein is the one-hot vector of the initial abnormal node;

[0129] S63, screening nodes constitute a candidate fault point set Ω;

[0130] S64, calculate the matching degree of the candidate node and the historical fault feature library:

[0131]

[0132] Select sim>0.9 and The largest node as the root fault point.

[0133] For example, the fault point positioning mechanism realizes accurate tracing of the root fault point through four stages of abnormal node initialization stage, fault propagation path simulation stage, candidate fault point screening stage and historical experience verification stage. For the abnormal node initialization stage, the device with an abnormal confidence exceeding 0.85 can be taken as the starting node, and the initial fault source is marked in the pre-constructed device association graph. The edge weight of the association graph reflects the physical connection or data flow dependence relationship between devices, for example, the pipeline connection relationship or sensor data sharing path of the air conditioning system and the fresh air system. For the fault propagation path simulation stage, a random walk algorithm can be used to perform 100-step state transition on the graph, and the transition probability of each step is determined by the association strength of the current node and the adjacent node. For example, if a valve (node A) has a direct control relationship with a pipeline (node B), the transition probability from A to B is higher, which simulates the characteristics of the possible spread of the fault along the actual physical path. For the candidate fault point screening stage, the fault propagation probability of each node can be calculated by accumulating the node access frequency after 100-step walking, and the nodes with probability values higher than the threshold (such as 0.7) are screened out to form the candidate fault point set Ω. This process is equivalent to marking the area most likely to be affected by the initial fault in the network. For the historical experience verification stage, the matching degree of the candidate node and the historical fault feature library can be calculated, and the node with a matching degree exceeding 0.9 and the highest score is determined as the root fault point. For example, if a sensor node has been abnormally caused by power module failure in historical data multiple times, and the characteristics (such as voltage fluctuation pattern) of the current candidate node are similar to the historical case with a similarity of 0.95, the root fault source is determined. The present application combines graph reasoning and historical experience verification, dynamically simulates the fault propagation path through random walk, avoids relying on fixed rules, and uses empirical data of the historical database to improve the positioning accuracy, effectively solving the technical problem of difficult tracing of the root cause of fault in complex systems.

[0134] The hierarchical alarm signal of step S7 includes:

[0135] Primary alarm: triggered when the device health degree H<0.15 or the fault propagation probability >0.95, and the critical device fault instruction is pushed;

[0136] Secondary alarm: when 0.6≤ ≤0.95 and the number of associated devices ≥3, trigger, prompt the risk of chain failure;

[0137] Third-level alarm: trigger performance degradation warning when the device health degree δH>10% for K=5 consecutive periods.

[0138] For example, the present application realizes precise grading response of fault risk through a three-level threshold system of first-level alarm, second-level alarm and third-level alarm. For first-level alarm (critical device failure), when the device health degree H is lower than 0.15 or the fault propagation probability exceeds 0.95, the threshold is set for the scene that the device is in a critical failure state or the fault may spread rapidly, for example, the risk of chain shutdown of surrounding devices caused by abnormal core power supply device or pipeline leakage. The system will directly push "critical device failure instruction" to require the operation and maintenance personnel to intervene immediately, and priority is given to the operation of the core function of the building. For second-level alarm (chain failure risk), when the fault propagation probability is between 0.6 and 0.95 and the number of associated devices reaches or exceeds 3, the condition is suitable for potential spreading failure, for example, abnormal air conditioning unit in a certain area causes overload of adjacent devices. The system prompts "chain failure risk" to guide the operation and maintenance team to take regional inspection to block the propagation path before the fault expands, and balance the response efficiency and resource allocation. For third-level alarm (performance degradation warning), when the decline rate δH of device health degree for K=5 consecutive monitoring periods exceeds 10%, the mechanism captures progressive degradation through long-term trend analysis, for example, the slow decrease of efficiency caused by bearing wear of motor. The system issues "performance degradation warning" to prompt the need for preventive maintenance, avoiding the accumulation of small problems to cause sudden failure, so as to realize the change from passive repair to active operation and maintenance. In the present application, the grading strategy combines static threshold with dynamic trend analysis, which not only ensures the rapid response of emergency failure through high threshold, but also identifies potential risks in advance through multi-device association and long-term degradation indicators, forming an alarm system covering the whole cycle of "immediate failure - spread risk - performance degradation", which significantly improves the fine management level of building equipment operation and maintenance.

[0139] The updating operation of the step S8 comprises:

[0140] S81, when receiving the operation and maintenance feedback confirmation as a false alarm, adding the feature vector at this moment to the negative sample library, and retraining the abnormality detection model, the loss function is:

[0141]

[0142] Wherein, y∈{0,1} is the true label, and λ=0.01 is the regularization coefficient;

[0143] S82, correcting the edge weight of the device association graph according to the maintenance record:

[0144]

[0145] wherein γ = 0.1 is a decay factor, Δt is the time since last maintenance, μ = 0.3 is a maintenance boost coefficient, is a maintenance event indicator function, which is 1 when maintenance occurs.

[0146] For example, in step S81, when the operation and maintenance personnel confirms a system false alarm (such as a device is normal but triggers an alarm), the device feature vector (such as voltage, current, temperature and humidity, vibration intensity, etc.) at the false alarm time is marked as a negative sample, stored in the negative sample library, and the abnormal detection model is retrained. The loss function is through the cross-entropy loss combined with L2 regularization, which not only improves the model's ability to distinguish false alarms, but also prevents overfitting. In step S82, according to the maintenance record, the weight of the edge in the device association graph is adjusted, which can reflect the natural decay characteristics of the device association strength over time, is a maintenance event indicator function, which is 1 when maintenance occurs, otherwise 0, indicating that the maintenance behavior will significantly enhance the association weight between devices, for example,

[0147] If device A is recently maintained due to sensor failure, the data coupling relationship between device A and device B after maintenance may be weakened (because the data of the faulty sensor has been corrected), at this time Δt is small, the γ.Δt term has weak influence, but = 1 will directly enhance the edge weight between device A and device B through μ, ensuring that the graph reflects the association changes after actual maintenance.

[0148] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for identifying fault points in building monitoring equipment, characterized in that, The method includes the following steps: Step S1: Real-time acquisition of building equipment operation data, environmental data, and video stream data at the edge computing node. The operation data includes voltage and current signals, and the environmental data includes temperature and humidity. Step S2: Perform multimodal fusion processing on the operating data, environmental data, and video stream data to generate a device status feature vector; Step S3: Extract the dynamic change trend of the feature vector based on the sliding time window, and calculate the composite feature index. The composite feature index includes equipment health H, electrical anomaly index E, environmental stability coefficient S, vibration intensity η, and temperature and humidity gradient. Perform wavelet packet decomposition on the voltage and current signals to extract the energy distribution within the specified frequency band range of 0.5-1.5Hz, and calculate the electrical anomaly index E. Calculate the environmental stability coefficient S based on the temperature and humidity change rate. Fuse the environmental stability coefficient S and the electrical anomaly index E to generate the equipment health H. When the equipment health H is lower than 0.2, trigger anomaly detection. Step S4: Analyze the composite feature indicators using a pre-trained anomaly detection model to identify abnormal device behavior and generate anomaly confidence scores; Step S5: When the anomaly confidence level exceeds a preset threshold, construct a device association map, which includes the physical connection relationship and data coupling relationship of the devices; Step S6: Combine the historical fault database to traverse the device association map and locate the root fault point causing the anomaly. Fault point location includes: S61. Starting with devices with anomaly confidence exceeding 0.85, perform a random walk algorithm on the association graph, with the state transition probability determined based on edge weights. S62. Calculate the failure propagation probability of each node after T=100 steps of walking; S63. Select nodes with a failure propagation probability exceeding 0.7 to form a candidate failure point set; S64. Calculate the matching degree between the candidate nodes and the historical fault feature database, and select the node with a matching degree of more than 0.9 and the highest fault propagation probability as the root fault point. Step S7: Generate a graded alarm signal based on the fault location result and push it to the operation and maintenance terminal; Step S8: Update the anomaly detection model and correlation graph based on operation and maintenance feedback data.

2. The method for identifying fault points in building monitoring equipment according to claim 1, characterized in that, The video stream data processing in step S1 includes: S11. Define a dynamic monitoring area in the working area of ​​the equipment and extract continuous grayscale images at a fixed frame rate; S12. Perform optical flow field analysis on adjacent frame images to obtain the displacement vector of each pixel; S13. Calculate the vibration intensity index of the equipment based on the displacement vector, specifically including: taking the average value of the displacement vector's modulus as the basic vibration amount; The final vibration intensity is generated by superimposing a compensation factor on the rate of change of the monitored area. S14. Align the vibration intensity with the corresponding timestamp of the running data in time and space to form a four-element synchronous data group: vibration intensity value, instantaneous voltage value, effective current value, and ambient temperature value.

3. The method for identifying fault points in building monitoring equipment according to claim 1, characterized in that, The multimodal fusion processing in step S2 specifically includes: S21. The timestamps of the running data and the environmental data are aligned by the dynamic time warping algorithm. The alignment degree A is calculated to meet the preset conditions. The maximum allowable time delay threshold between the sensor data timestamp and the environmental data timestamp is 500ms. Time deviation compensation is performed by the S-shaped compensation function. S22. The vibration intensity extracted from the video stream data is Gaussian normalized, where the normalization parameter is determined based on the historical 30-day mean and standard deviation of vibration intensity. S23. The voltage and current data, temperature data, and normalized vibration intensity are fused in the latent space to generate the equipment state feature vector F. Vector splicing is used during fusion, and the weight coefficients of each data source are dynamically adjusted based on the voltage fluctuation rate and the proportion of vibration intensity.

4. The method for identifying fault points in building monitoring equipment according to claim 1, characterized in that, The anomaly detection model in step S4 adopts a structure combining temporal convolutional networks and self-attention mechanisms, and its processing flow includes: S41. Input a composite feature index matrix of N consecutive time windows, where the number of time windows N=20 and the feature dimension D=5; S42. Extract local temporal features through a temporal convolutional network, using a dilated convolutional structure and the ReLU activation function; S43. Input the output of the temporal convolutional network into the attention layer and calculate the attention weights for each time window; S44. Output the anomaly confidence level C through the fully connected layer. When C is greater than 0.85, trigger the fault location process.

5. A method for identifying fault points in building monitoring equipment according to claim 1 or 4, characterized in that, The device association map construction in step S5 includes: S51. Parse the building equipment topology configuration file and generate a matrix representing the physical connection relationships between the equipment; S52. Calculate the coupling weight of historical operating data between devices using the mutual information algorithm; S53. By integrating physical connection relationships and data coupling weights, a weighted adjacency matrix is ​​constructed, where the edge weights in the graph represent the strength of the association between devices.

6. The method for identifying fault points in building monitoring equipment according to claim 1, characterized in that, The graded alarm signals in step S7 include: Level 1 Alarm: Triggered when the equipment health H is below 0.15 or the fault propagation probability is greater than 0.95, pushing out fault instructions for critical equipment; Level 2 alarm: Triggered when the fault propagation probability is between 0.6 and 0.95 and the number of associated devices is not less than 3, indicating the risk of cascading failures; Level 3 alarm: When the device health rate decreases by more than 10% for 5 consecutive monitoring cycles, a performance degradation warning is triggered.

7. The method for identifying fault points in building monitoring equipment according to claim 1, characterized in that, The update operation in step S8 includes: S81. When the operation and maintenance feedback confirms that it is a false alarm, add the feature vector at that moment to the negative sample library, and retrain the anomaly detection model based on the cross-entropy loss function and regularization term. S82. Dynamically adjust the edge weights in the association graph based on maintenance records, and introduce a time decay factor and a maintenance enhancement coefficient to reflect changes in equipment association relationships.

Images