A method and system for predicting faults in intelligent building equipment

By integrating a physical constraint encoder-generator architecture and a degradation trend field, health and operating condition characteristics are decoupled, solving the problems of scarce samples and lack of integration of physical laws in the fault prediction of intelligent building equipment, and realizing accurate prediction of early faults and multi-dimensional health status characterization.

CN122673642BActive Publication Date: 2026-11-10HUNAN UNIV OF TECH +1
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611161889.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-08-03
Publication Date
2026-11-10
Estimated Expiration
2046-08-03

AI Technical Summary

Technical Problem

Existing methods for predicting faults in intelligent building equipment rely on a large number of labeled fault samples, which are scarce, unevenly distributed, and costly to obtain, making it difficult to predict faults in the early stages. Furthermore, they do not fully incorporate the physical operating principles of the equipment, resulting in prediction results that do not conform to common sense physics, thus reducing the reliability and practicality of the prediction methods.

Method used

An encoder-generator architecture that integrates physical constraints is adopted. By constructing statistical soft boundaries and degradation trend fields, combined with anomaly feature extraction and hard boundary verification, health features and operating condition features are decoupled. Based on the degradation trend field, a multi-step potential space walk is performed to generate multiple effective walk paths for fault prediction.

Benefits of technology

It accurately characterizes the true health status of equipment, captures early subtle abnormal trends, ensures the rationality and robustness of prediction results, realizes quantitative assessment of fault risks and characterization of multi-dimensional health status, and assists maintenance personnel in formulating maintenance plans in advance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122673642B_ABST
    Figure CN122673642B_ABST
Patent Text Reader

Abstract

This invention provides a method and system for predicting faults in intelligent building equipment, comprising: constructing a statistical soft boundary based on a first set of potential vectors; generating a degradation trend field and a group reference step count by combining a preset fault degradation model; obtaining actual potential vectors and obtaining anchor vectors by extracting abnormal features; setting a target degradation trend based on the degradation trend field and the anchor vectors; performing a multi-step potential space walk along the target degradation trend; generating multiple effective walk paths and corresponding prediction sensor sequences through generator reconstruction, temporal discrimination, and physical hard boundary verification; determining the fault boundary step count of the effective walk path based on the hard boundary priority principle; obtaining a risk value and confidence weight based on the fault boundary step count and the group reference step count; constructing a fault susceptibility profile of the equipment based on the risk value and confidence weight; and performing fault warning in conjunction with a hierarchical early warning decision mechanism, thereby improving the accuracy of fault prediction for intelligent building equipment.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of operation and maintenance technology for intelligent building equipment, and more specifically, to a method and system for predicting faults in intelligent building equipment. Background Technology

[0002] Intelligent building equipment is a core component of modern building infrastructure, encompassing various intelligent control systems and terminal devices. These mainly include systems such as central air conditioning, water supply and drainage, ventilation and elevator operation monitoring, video surveillance, intrusion alarms, access control, electronic patrol systems, and parking management systems. However, as the level of building intelligence continues to increase, the complexity of these systems grows exponentially, and their operational stability directly affects building safety, energy consumption levels, and user experience.

[0003] Existing data-driven fault prediction methods mostly rely on a large number of labeled fault samples for supervised learning. However, fault samples of intelligent building equipment are generally characterized by scarcity, uneven distribution, and high acquisition costs. Most devices only experience a few faults during their lifespan, and these faults are often accompanied by data acquisition interruptions or data quality degradation, making it difficult to obtain a sufficient number and quality of labeled fault samples. This severely limits the generalization ability of traditional supervised learning methods, making it difficult to effectively predict early faults.

[0004] Meanwhile, most existing data-driven fault prediction methods only focus on the statistical characteristics of the data itself and do not fully incorporate the physical operating laws of the equipment, resulting in prediction results that do not conform to common sense physics, thus seriously reducing the credibility and practicality of the prediction methods. Summary of the Invention

[0005] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for predicting faults in intelligent building equipment, the method comprising: S1: Construct and train an encoder-generator architecture that integrates physical constraints, construct a statistical soft boundary based on the first latent vector set, and generate a degradation trend field and a group reference step number for each degradation trend by combining a preset fault degradation model. The encoder-generator architecture includes at least an encoder, a working condition classifier, and a generator. S2: Obtain the actual potential vectors corresponding to the current sensor window data and operating condition data, and use the abnormal feature extraction method to obtain the corresponding anchor vectors; S3: Based on the degradation trend field, the target degradation trend is set simultaneously in conjunction with the anchoring vector. Multi-step potential space walk is performed along the target degradation trend. Multiple effective walk paths and corresponding prediction sensor sequences are generated through generator reconstruction, temporal discrimination and physical hard boundary verification. S4: Determine the number of fault boundary steps for each effective walk path based on the hard boundary priority principle, and obtain the risk value and confidence weight in the target degradation trend based on the number of fault boundary steps and the group reference steps; S5: Based on the risk value and confidence weight of all degradation trends in the degradation trend field, construct a fault susceptibility profile of the equipment and combine it with a hierarchical early warning decision mechanism to provide fault early warning.

[0006] Furthermore, embodiments of the present invention also provide a fault prediction system for intelligent building equipment, comprising: The spatial construction module is used to construct and train an encoder-generator architecture that integrates physical constraints, and to construct a statistical soft boundary based on a first set of potential vectors. Simultaneously, it combines a preset fault degradation model to generate a degradation trend field and a group reference step number for each degradation trend. The state extraction module is used to generate the actual potential vector corresponding to the current sensor window data and the operating condition data, and to obtain the corresponding anchor vector by combining the abnormal feature extraction. The walk prediction module is used to set the target degradation trend based on the degradation trend field and simultaneously combine the anchor vector, and perform a multi-step walk in the potential space along the target degradation trend. Simultaneously, it combines generator reconstruction, temporal discrimination and physical hard boundary verification to obtain multiple effective walk paths and corresponding prediction sensor sequences. The risk quantification module is used to generate the number of fault boundary steps for each effective walk path according to the hard boundary priority principle, and to obtain the risk value and confidence weight in the target degradation trend based on the number of fault boundary steps and the group reference steps. The risk warning module is used to construct a fault susceptibility profile of equipment based on the risk value and confidence weight of all degradation trends in the degradation trend field, and simultaneously combine it with a hierarchical early warning decision mechanism to provide fault warnings.

[0007] Compared with the prior art, the present invention has the following beneficial effects: An encoder-generator architecture is adopted to decouple health features from operating condition features. Physical constraints are injected into the training of the encoder-generator architecture to ensure that the latent space representation conforms to the equipment operation rules. At the same time, a degradation trend field is constructed by combining a preset fault degradation model to provide a degradation direction benchmark for subsequent fault prediction. The encoder mapping in the encoder-generator architecture is used to obtain the actual potential vector decoupled from the operating conditions, thereby accurately representing the current real health status of the equipment. Anchor vectors are obtained based on abnormal features, thereby capturing the evolution trend of early weak anomalies of the equipment and providing directional guidance for subsequent degradation trend correction and disturbance walk. By simulating the equipment failure and degradation process through a potential space perturbation walk, and combining generator reconstruction, temporal rationality judgment and physical hard boundary verification, invalid prediction results that violate temporal logic and physical limits are filtered out, thereby ensuring the rationality of the prediction sequence and the robustness of the final prediction results. The principle of prioritizing hard boundaries is adopted to ensure that the fault boundary determination conforms to the physical operating limits of the equipment. Based on the number of fault boundary steps and the group reference steps, normalized risk values ​​and confidence weights are obtained to achieve a quantitative assessment of fault risk. By aggregating risk and confidence information from multiple fault dimensions, a fault susceptibility profile of the equipment is constructed, enabling the representation of the equipment's multidimensional health status. Simultaneously, early warning thresholds are combined to achieve early prediction and graded warning of faults, thereby assisting maintenance personnel in formulating maintenance plans in advance. Attached Figure Description

[0008] Figure 1 This is a flowchart of the steps of a method for predicting faults in intelligent building equipment according to the present invention; Figure 2 This is a schematic diagram of a fault prediction system for intelligent building equipment according to the present invention. Detailed Implementation

[0009] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart of the steps of a method for predicting faults in intelligent building equipment according to the present invention. The following is a detailed description of this method for predicting faults in intelligent building equipment.

[0010] Step S1: Construct and train an encoder-generator architecture that integrates physical constraints, construct a statistical soft boundary based on the first latent vector set, and generate a degradation trend field and a group reference step number for each degradation trend by combining a preset fault degradation model.

[0011] Specifically, step S1 can be divided into the following sub-steps: Step S11: Construct and train an encoder-generator architecture that integrates physical constraints. This encoder-generator architecture includes at least an encoder, a condition classifier, and a generator. Specifically: First, obtain past sensor datasets and past operating condition datasets, and perform outlier removal, missing value imputation, and standardization on the past sensor datasets. For example, for the sensor dataset of a central air conditioning chiller unit, this dataset contains sensor data in at least seven dimensions: inlet water temperature, outlet water temperature, inlet water pressure, outlet water pressure, compressor current, compressor voltage, and lubricating oil temperature. For this dataset, outlier removal can be performed using a combination of the 3σ criterion and the isolated forest algorithm. First, the 3σ criterion is used to remove outliers that are significantly outside the normal range, such as inlet water temperatures exceeding 100℃ or falling below 0℃. Then, the isolated forest algorithm is used to remove statistically significant outliers. Missing value imputation is performed using a combination of linear interpolation and nearest neighbor imputation. For example, if the number of consecutive missing points is less than 5, linear interpolation is used for imputation; if the number of consecutive missing points is greater than or equal to 5, the average of the previous 7 days in the same period is used for imputation. Standardization is performed using Z-score standardization.

[0012] Discrete data is encoded using one-hot encoding, and continuous data is normalized using one-hot encoding. Continuing from the previous example, assume that the operating data of the central air conditioning chiller unit includes three dimensions: operating mode, load rate, and ambient temperature. The operating mode includes three modes: (cooling, heating, and standby). The load rate ranges from 0% to 100%, and the ambient temperature ranges from -10℃ to 40℃. The operating mode is discrete data, which can be converted into a 3-dimensional binary vector using one-hot encoding. For example, the cooling mode is encoded as [1,0,0], the heating mode as [0,1,0], and the standby mode as [0,0,1]. The load rate and ambient temperature are continuous data, which can be normalized using min-max normalization to map the data to the interval [0,1].

[0013] Then, training samples are constructed according to the preset window length and sliding step size. For example, the window length is set to 30 sampling points, corresponding to 30 minutes of device operation data; the sliding step size is set to 5 sampling points, corresponding to a 5-minute sliding interval. Through the above processing, one year of historical data can generate multiple time series samples. The dimension of each sample is (30,7), which corresponds to the measurement values ​​of 7 sensors within 30 minutes.

[0014] Then, an adversarial training strategy is used to alternately update the parameters of the encoder and the condition classifier. Specifically: The encoder adopts a hybrid architecture of one-dimensional convolutional neural network and long short-term memory network (LSTM). During the training phase, it takes preprocessed past sensor datasets and past working condition datasets as conditional inputs and outputs a latent vector.

[0015] In one possible embodiment, the encoder's intermediate layer employs a hybrid architecture of three convolutional layers and two LSTM layers. The first convolutional layer contains 32 convolutional kernels with a kernel size of 5 and a stride of 1, using the ReLU activation function. The second convolutional layer contains 64 convolutional kernels with a kernel size of 3 and a stride of 1, using the ReLU activation function. The third convolutional layer contains 128 convolutional kernels with a kernel size of 3 and a stride of 1, using the ReLU activation function. Assuming these three convolutional layers take the previous sensor dataset with dimensions (30, 7) as input, the output after three convolutional layers is a convolutional temporal feature of (30, 128). For the 5-dimensional operating condition data, i.e., the 3D one-hot encoded operating mode... The 2D normalized load rate and ambient temperature are copied 30 times along the time dimension by the temporal expansion module to expand into a working condition temporal feature of dimension (30,5). In the intermediate layer after the convolutional layer and before the LSTM layer, the convolutional temporal feature of dimension (30,128) and the working condition temporal feature of dimension (30,5) are concatenated in the channel dimension to obtain a joint temporal feature of dimension (30,133). After the convolutional layer, two LSTM layers are connected. Each LSTM layer contains 64 hidden units and uses the tanh activation function. The LSTM takes the output of the last step of the temporal sequence and fixes the output to a 64-dimensional temporal feature vector. After being mapped by a fully connected layer, the 64-dimensional LSTM feature vector is mapped to a 16-dimensional normalized latent vector.

[0016] The operating condition classifier adopts a multi-layer fully connected network architecture. Taking the latent vector output by the encoder as input, it outputs the probability distribution of various operating modes corresponding to the building's intelligent equipment. Continuing the previous example, since the central air conditioning chiller unit only includes three operating modes (cooling, heating, and standby), a three-layer fully connected network architecture can be adopted. The input layer of this architecture receives latent vectors with a dimension of 16; the first hidden layer contains 128 neurons and uses the ReLU activation function; the second hidden layer contains 64 neurons and uses the ReLU activation function; the output layer contains 3 neurons and uses the Softmax activation function, finally outputting the probability distribution of the above three operating modes.

[0017] The training process of the adversarial training strategy can be represented as follows: First, fix the parameters of the encoder, train the operating condition classifier to accurately distinguish the operating mode of the equipment based on the latent vector, and use the cross-entropy loss of the operating condition classification as the loss function; then fix the parameters of the operating condition classifier, train the encoder so that the operating condition classifier cannot correctly distinguish the operating mode; when the classification accuracy of the operating condition classifier converges to the level of random guessing, the training ends. For example, for the above-mentioned central air conditioning chiller unit, since it contains three operating modes, when the classification accuracy of the operating condition classifier converges to about 33%, it has converged to the level of random guessing, and the training can be ended at this time.

[0018] It should be noted that the operating condition classifier is an auxiliary network specifically designed to counter the encoder during the training phase and is no longer used in the actual inference phase. Its function is to take the latent vector as input and attempt to identify the corresponding operating condition category. The amount of operating condition information remaining in the latent vector is measured by using the cross-entropy loss of the operating condition classification. Its design aims to form an adversarial game with the encoder, whereby the encoder is trained to maximize the classification error of the operating condition classifier, and the operating condition classifier is trained to minimize this error. Both sides alternately update their parameters. This mechanism forces the encoder to learn to completely separate the operating condition information from the latent vector when receiving operating condition data as input, retaining only the internal state representation of the equipment that is unrelated to the operating condition. The reason for this design is that sensor data is naturally coupled with the influence of the operating condition. If explicit decoupling is not performed, the subsequent perturbation walk along the degradation trend in the latent space will be affected by the operating condition factors, causing the same degree of degradation of the fault to be mapped to different locations in the latent space under different seasons or load conditions. This would disrupt the index consistency of the degradation trend field and make the risk assessment lose physical comparability.

[0019] Next, the generator is trained using a training strategy that combines reconstruction loss with physical constraints. The generator network consists of a hybrid architecture combining deconvolutional neural networks and LSTM.

[0020] Following the above embodiments, the generator in the training phase takes a 16-dimensional latent vector and corresponding 5-dimensional past working condition data as joint input. Its intermediate layers sequentially include a pre-connected mapping layer, a reshape layer, two LSTM layers, and three deconvolution layers. Specifically: first, the pre-connected mapping layer maps the 21-dimensional joint input vector into a 1920-dimensional feature vector (30×64); then, the reshape layer reshapes the 1920-dimensional feature vector into an initial temporal feature map with dimensions (30, 64); the initial temporal feature map is then fed into two stacked LSTM layers, each with 64 hidden units, a tanh activation function, and other relevant parameters. The LSTM layers process the temporal features step-by-step, outputting a global temporal feature map with dimensions maintained at (30, 64); then… Three one-dimensional deconvolutional layers are connected, with all layers having a stride of 1 and padding set to 'same' to ensure the time dimension remains constant at 30. The first deconvolutional layer contains 128 deconvolutional kernels of size 3, using the ReLU activation function, with an output dimension of (30, 128). The second deconvolutional layer contains 64 deconvolutional kernels of size 3, using the ReLU activation function, with an output dimension of (30, 64). The third deconvolutional layer contains 32 deconvolutional kernels of size 5, using the ReLU activation function, with an output dimension of (30, 32). The output layer uses a 1×1 one-dimensional convolutional layer to map the 32 feature channels to 7 sensor channels, ultimately outputting a reconstructed sensor data sequence with a dimension of (30, 7).

[0021] The joint loss function used by the generator during the training phase consists of two parts: a reconstruction loss term and a physical constraint term. The reconstruction loss term represents the mean square error between the reconstructed sensor data and the corresponding past sensor data in the reconstructed sensor data sequence. The physical constraint term is represented by the weighted sum of the residuals of the energy conservation equation and the equipment characteristic curve equation. The weighting coefficients for the two are set by domain experts according to actual needs. For example, for a central air conditioning chiller unit, the corresponding energy conservation equation can be simplified to evaporator heat exchange = compressor input power + condenser heat exchange. The residual value of the corresponding energy conservation equation is the absolute value of the difference between the evaporator heat exchange and (compressor input power + condenser heat exchange). The residual value of the equipment characteristic curve equation is the absolute value of the difference between the actual operating point and the chiller unit performance curves provided by the manufacturer, such as the cooling capacity-COP curve and the current-load rate curve.

[0022] The generator is trained with the goal of minimizing the joint loss function, and the generator network parameters are updated using the gradient backpropagation algorithm until the total loss function value is lower than the preset convergence threshold.

[0023] Step S12, constructing statistical soft boundaries based on the first potential vector set, specifically: All preprocessed past sensor data are input into the pre-trained encoder to obtain the first latent vector set. A single-class classification algorithm is used to fit the distribution of the first latent vector set, with minimizing the envelope volume as the training objective. The envelope boundary is defined by a preset quantile to generate a statistical soft boundary.

[0024] For example, using a single-class classification algorithm with Support Vector Data Description (SVDD), a minimum hypersphere that can contain the most historical latent vectors is constructed. Assuming the preset quantile is set to 99%, that is, the hypersphere must contain 99% of the historical latent vectors, then the surface of the hypersphere is the statistical soft boundary. The latent vectors inside the hypersphere correspond to the normal operating state of the device, and the latent vectors outside the hypersphere correspond to the abnormal operating state of the device.

[0025] Step S13: Combine the preset fault degradation model to generate a degradation trend field and a group reference step number for each degradation trend. Specifically: First, a preset set of fault degradation models is obtained, and a corresponding degradation direction curve is generated for each known fault in the preset set of fault degradation models. All degradation direction curves are collected and organized, and a degradation trend field is constructed using the potential space coordinates as the index key and the fault type label as the value. The degradation trend field is used to describe the direction and path of the device evolving into various fault states in the potential space.

[0026] For example, for a central air conditioning chiller unit, it is assumed that its corresponding preset fault degradation model set includes five common faults: condenser scaling, evaporator scaling, compressor wear, refrigerant leakage, and lubricating oil deterioration. Taking condenser scaling fault as an example, its physical degradation model can be described as follows: as the scaling thickness increases, the heat transfer coefficient of the condenser gradually decreases, leading to an increase in condensing temperature, an increase in compressor power consumption, and a decrease in cooling capacity. Its degradation parameter is the scaling thickness, which is assumed to be in the range of [0mm, 5mm], representing the degradation from a healthy state to a severe fault state. The discrete sampling points of the degradation parameter are set to 11 sampling points: 0mm, 0.5mm, 1mm, 1.5mm, 2mm, 2.5mm, 3mm, 3.5mm, 4mm, 4.5mm, and 5mm, thus covering the complete degradation process from a healthy state to a severe fault state. At each sampling point, a simulation model of the central air conditioning chiller unit calibrated with past operating data is run, and multi-dimensional sensor simulation data corresponding to the degradation stage is output. The simulation model can be constructed using software such as TRNSYS and its parameters are calibrated using past operating data. The sensor simulation data corresponding to the degradation stage of each discrete sampling point is input into the encoder to obtain the degradation potential vector sequence. This sequence contains 11 16-dimensional potential vectors, corresponding to 11 different degradation stages. Then, cubic spline interpolation or Bezier curve fitting algorithm is used to smoothly fit the degradation potential vector sequence to generate the degradation direction curve of condenser scaling failure. The above process is repeated for the remaining four types of faults to generate corresponding degradation direction curves. For all five types of fault degradation direction curves, the 16-dimensional potential space is divided into uniformly discrete grids of fixed size. Each degradation direction curve is mapped to all the grid cells it passes through. In each grid cell, the fault type label corresponding to the curve, the tangent vector at the grid center coordinates, and the curve parameterization expression are stored, forming a spatial mapping structure with the grid center coordinates as the index key and the associated degradation trend information as the value. This enables the constructed degradation trend field to support the retrieval of all degradation trend curves passing through any location in the potential space.

[0027] Next, the group reference step count for each degradation trend in the degradation trend field is determined. The group reference step count represents the average number of walks for equipment of the same model as it evolves from a healthy state to a fault boundary. This is used to convert the absolute step count into a relative risk value. Specifically: Collect historical walk data of equipment of the same model under various degradation trends, extract the boundary steps from the healthy state to the fault boundary of each equipment, and take the median as the group reference steps for the corresponding degradation trend; if there is no group data, the domain experts calibrate the preset maximum number of steps as the reference number based on the equipment design life and operating stress level.

[0028] For example, historical walk data of 100 central air conditioning chiller units of the same model under the trend of condenser scaling degradation are collected, and the boundary steps from the healthy state to the fault boundary of each device are extracted. The median is calculated to be 200 steps. Then, the group reference step number for the condenser scaling degradation trend is set to 200 steps. For the evaporator scaling degradation trend, due to the lack of sufficient historical data, based on the equipment design life (15 years) and the operating stress level (average load rate of 70%), the preset maximum step number is calibrated to 180 steps as the reference step number (this is only an example, and the specific setting needs to be combined with the actual situation).

[0029] Step S2: Obtain the actual potential vectors corresponding to the current sensor window data and operating condition data, and obtain the corresponding anchoring vectors by using anomaly feature extraction.

[0030] Specifically, step S2 can be divided into the following sub-steps: Step S21: Obtain the actual potential vectors corresponding to the current sensor window data and operating condition data. Specifically: First, acquire the current sensor window data and operating condition data. Perform the same cleaning and standardization processing on the current sensor window data as on historical data. Perform the same encoding and normalization processing on the current operating condition data as on historical data. Verify the integrity of the current sensor window data. If the proportion of missing data exceeds a preset threshold, reacquire the data.

[0031] For example, the current sensor window data consists of sensor data from the central air conditioning chiller unit over the past 30 minutes, containing measurements across 7 dimensions. The sampling frequency is 1 minute / time. First, outliers are removed using the same 3σ criterion and isolated forest algorithm as historical data. Then, Z-score standardization is performed, similar to historical data. Assuming the current operating condition data is [operating mode: cooling, load rate: 75%, ambient temperature: 28℃], the operating mode is converted to [1,0,0] using one-thermal encoding. The load rate and ambient temperature are converted to their corresponding normalized values ​​using minimum-maximum normalization. Subsequently, the integrity of the current sensor window data is verified. If one sampling point is found to be missing, linear interpolation is used to fill it. Assuming the preset threshold for the missing data ratio is 10%, and the missing data ratio is 1 / 30 = 3.3%, which is lower than the preset threshold, the data in the current window is deemed valid.

[0032] Next, the preprocessed current sensor window data and current operating condition data are input into the pre-trained encoder for forward propagation to obtain the actual latent vector. The actual latent vector is then verified to see if it is located within the statistical soft boundary. If it is outside the boundary, it is marked as a state offset and the offset distance and offset time are recorded synchronously.

[0033] It should be noted that the normalized state offset distance is used as the quantitative basis for verifying whether the actual latent vector is located inside the statistical soft boundary. The normalized state offset distance is defined as the ratio of the Euclidean distance from the actual latent vector to the center of the SVDD hypersphere corresponding to the statistical soft boundary to the radius of the hypersphere.

[0034] For example, if the distance from an actual potential vector to the statistical soft boundary is 0.8 times the radius of the hypersphere, it is located inside the statistical soft boundary, and the device is determined to be in normal operation. If the distance from an actual potential vector to the statistical soft boundary is 1.2 times the radius of the hypersphere, it is outside the statistical soft boundary, and it is immediately marked as a state deviation. The deviation distance is recorded as 0.2 times the radius, and the deviation time is the current system time. If a state deviation is detected three times in a row, a level two warning is triggered directly, indicating that the device may have experienced an early anomaly.

[0035] Step S22: Obtain the current anchor vector using anomaly feature extraction. Specifically: An unsupervised anomaly detection algorithm is used to scan the current sensor window data dimension by dimension, and a set of statistically rare fragments with risk scores exceeding a preset threshold are extracted as anomaly features. The anomaly features are then input into a fully connected correction network, and anchor vectors are obtained through forward propagation.

[0036] The risk score is a quantitative score derived from a comprehensive analysis of the sensor's single-dimensional amplitude anomaly, temporal fluctuation anomaly, and multi-dimensional coupling anomaly. The higher the risk score, the greater the degree to which the corresponding sampling segment deviates from the normal operating distribution of the equipment.

[0037] In one possible embodiment, the isolated forest algorithm is used to obtain the global outlier degree of each sampling point in the current window. That is, the outlier score of each sampling point is obtained by using the outlier scoring formula of the isolated forest, and the 95th percentile of the outlier scores of all sampling points in the window is taken as the single-dimensional amplitude outlier value of that dimension. Obtain the difference sequence of adjacent sampling points within the current window, standardize the difference sequence using the mean and standard deviation of the difference of historical normal data using Z-score, and take the 95th percentile of the absolute value of all standardized differences within the window as the time series fluctuation anomaly value; Obtain the Pearson correlation coefficient matrix between sensors in all historical normal data and the Pearson correlation coefficient matrix between sensors in the current window. Use the Frobenius norm to quantify the distance between the two Pearson correlation coefficient matrices and use the normalized value of the distance as the multidimensional coupling anomaly value. The distance can be normalized using the max-min normalization method. The three indicators are weighted and combined using weighting coefficients set by domain experts based on actual needs to obtain the final risk score.

[0038] For example, using the Isolation Forest algorithm, with 100 trees and a subsample size of 256, the sensor window data for the current 30 minutes is scanned dimension by dimension. The window size is set to 5 sampling points, and the sliding step size is set to 1 sampling point. The risk score for each sliding window is calculated, assuming a preset risk score threshold of 0.6. The scan finds that the risk score reaches 0.72 in the compressor current dimension from the 20th to the 24th sampling points, exceeding the preset threshold. At this point, the compressor current data of these 5 sampling points are extracted as anomaly features.

[0039] The correction network adopts a 3-layer fully connected network architecture: the input layer receives the abnormal features; the first hidden layer contains 64 neurons and uses the ReLU activation function; the second hidden layer contains 32 neurons and uses the ReLU activation function; the output layer contains 16 neurons and uses the tanh activation function, outputting an anchor vector with a dimension of (16,1); the tanh activation function is used to restrict the output value to the range [-1,1] to ensure that the magnitude of the direction vector is 1. Continuing with the previous example, the above abnormal features are input into the correction network, and a 16-dimensional anchor vector is obtained through forward propagation. This vector points in the direction of the abnormal increase in compressor current.

[0040] Step S3: Based on the degradation trend field, the target degradation trend is set simultaneously with the anchoring vector. A multi-step potential space walk is performed along the target degradation trend. Multiple effective walk paths and corresponding predictive sensor sequences are generated through generator reconstruction, temporal discrimination and physical hard boundary verification.

[0041] The specific step S3 includes steps S31-S33: Step S31: Based on the degradation trend field, simultaneously combine the anchoring vector to determine the target degradation trend, specifically: The target degradation trend represents the evolution direction corresponding to the possible failure types of the equipment. Since the theoretical degradation trend in the degradation trend field is generated based on a general physical degradation model, it may deviate from the actual situation of specific equipment. Therefore, it needs to be corrected by combining anchor vectors. The specific operation process can be expressed as follows: Select the target degradation trend curve from the degradation trend field and extract the tangent vector of the curve at the actual potential vector position; perform a weighted summation of the tangent vector and the anchor vector, with the weighting coefficient determined by the magnitude of the anchor vector through a saturated nonlinear scaling function, and restrict the weighting coefficient to a preset correction interval to obtain the corrected target degradation trend.

[0042] For example, the degradation trend curve of condenser scaling is selected from the degradation trend field. This curve is a parametric curve in a 16-dimensional potential space, with the parameter being the degree of degradation. The tangent vector t at the actual potential vector position of this curve is obtained. The direction of the tangent vector represents the theoretical evolution direction of condenser scaling failure. Assuming that the magnitude of the corresponding anchor vector a is 0.8, it indicates that the equipment currently has a certain abnormal trend. The saturated nonlinear scaling function adopts the sigmoid function. Assuming that the calculated weighting coefficient w is 0.73 and the preset correction interval is [0,1], and the current weighting coefficient is within this interval, then the corrected target degradation trend d = (1-w)×t + w×a = 0.27×t + 0.73×a. At this time, the target correction direction combines the theoretical degradation trend and the actual abnormal trend, thus better reflecting the actual evolution of the equipment.

[0043] Step S32: Starting from the corresponding position of the actual latent vector, perform a multi-step walk operation along the corrected target degradation trend to generate a perturbed latent sequence. Specifically: Starting from the actual potential vector, perform a multi-step walk operation along the modified target degradation trend; store the potential vectors of each step in sequence to form a perturbation potential sequence; the walk terminates when the number of walk steps reaches the preset maximum number of walk steps or when a physical hard boundary check is triggered.

[0044] For example, suppose the maximum number of steps for walking is set to 300. Starting from the actual potential vector, a 300-step walk operation is performed to generate a perturbed potential sequence of 301 potential vectors, including the starting point. Suppose that during the walk, a physical hard boundary check is triggered at the 150th step, at which point the subsequent walk process will be terminated immediately, and the perturbed potential sequence will only contain the first 151 potential vectors.

[0045] It should be noted that the single-step walk operation is expressed as follows: the actual potential vector of the current step is multiplied by the corrected target degradation trend by the perturbation step size and the product is added to obtain the direction perturbation term; random noise is sampled from a preset Gaussian distribution with a mean of 0. The variance of this Gaussian distribution is determined by the distance of the current step potential vector from the statistical soft boundary through proportional mapping. The larger the deviation distance, the larger the noise variance; the direction perturbation term is added to the random noise to obtain the next potential vector.

[0046] For example, the perturbation step size can be set to 0.02 times the latent spatial feature scale, where the latent spatial feature scale is defined as the average Euclidean distance between latent vectors in the first latent vector set. Assuming this average is 0.5, the perturbation step size can be set to 0.01. Assuming the current step latent vector is... The corrected target degradation trend is Then the direction disturbance term is Assuming the current step's latent vector deviates from the statistical soft boundary by 0.8 times the radius, then the variance of the Gaussian distribution is... ,in If the variance in this example is a pre-defined scaling factor set by domain experts based on historical data, then... Therefore, this example will sample random noise from a Gaussian distribution with a mean of 0 and a variance of 0.004. Then the next potential vector is .

[0047] Step S33 involves performing generator reconstruction, temporal discrimination, and physical hard boundary verification based on the perturbation latent sequence to generate multiple valid walk paths and corresponding predicted sensor sequences. Specifically: First, the potential vectors at each step in the perturbation potential sequence are paired with the current operating condition data and input point by point into the pre-trained generator for forward propagation to generate the predicted sensor sequence. For example, the perturbation potential sequence contains 301 16-dimensional potential vectors. Each potential vector is concatenated with the current operating condition data to form a 21-dimensional joint input vector, which is input into the generator and outputs 301 sensor data windows with a dimension of (30,7), which are then arranged in order to form the predicted sensor sequence.

[0048] Next, the predicted sensor sequence is input into a pre-trained time discriminator for time series rationality verification. The time series discriminator is used to distinguish between real device transient processes and generated unreasonable transient processes, thereby ensuring the temporal continuity and rationality of the predicted sensor sequence. The time series discriminator adopts a one-dimensional convolutional neural network architecture, with the sensor data window whose input layer receives the same dimension as the generator output vector dimension.

[0049] The training set of the time-series discriminator includes at least real sensor data of various transient processes such as normal device startup, shutdown, and load changes, as well as unreasonable transient data generated by the simulation model corresponding to each device. The training process adopts the binary cross-entropy loss function and uses the Adam optimizer for iterative optimization. The training terminates when the discrimination accuracy of the time-series discriminator reaches a preset threshold or a preset maximum number of training steps.

[0050] Continuing the previous example, the time-series discriminator adopts a one-dimensional convolutional neural network architecture. The input layer receives sensor data windows with a dimension of (30,7). The intermediate layer can be set to contain four convolutional layers, with 32, 64, 128, and 64 convolutional kernels respectively, and kernel sizes of 5, 3, 3, and 3 respectively. The stride is set to 1, and the ReLU activation function is used. After the global average pooling layer, two fully connected layers are connected, containing 64 and 32 neurons respectively, and the ReLU activation function is used. The output layer contains one neuron, uses the Sigmoid activation function, and outputs a time-series rationality score. The score range is [0,1]. The higher the time-series rationality score, the better the time-series rationality.

[0051] Each 30-minute window in the predicted sensor sequence is input into the time series discriminator. If the average time series rationality score is 0.65, which is higher than the preset score threshold of 0.6, the time series rationality check is passed, and the path that passes the time series rationality check is defined as a valid walk path. If the average time series rationality score is 0.5, which is lower than the preset threshold of 0.6, the perturbation step size is reduced by a preset decay factor, and the random noise variance is increased by a preset growth factor, and the walk operation is re-executed. The maximum number of retries is set to 3. If the time series rationality score still fails to meet the standard after 3 retries, the walk path is marked as an invalid path.

[0052] The cumulative number of invalid paths is calculated. When the proportion of invalid paths to the total number of paths exceeds a preset threshold under this degradation trend, a confidence reduction penalty factor that is negatively correlated with the proportion of invalid paths is applied to the confidence weights calculated subsequently. For example, assuming there are 50 total paths and 20 invalid paths, the proportion of invalid paths is 0.4, which exceeds the preset threshold of 0.3. In this case, the confidence reduction penalty factor p = 1 - 0.5 × (r - 0.3) can be set, where r is the proportion of invalid paths. In this example, the calculated confidence reduction penalty factor p = 1 - 0.5 × (0.4 - 0.3) = 0.95.

[0053] Next, physical hard boundary checks are performed step-by-step on the valid walk path. These physical hard boundaries represent the physical insurmountable limits defined by the equipment nameplate parameters and the constraints of the equipment's operating logic. Once these limits are reached, it indicates a equipment malfunction. The specific process for physical hard boundary checks includes: The parameter values ​​of each dimension in the predicted sensor sequence corresponding to the effective walk path are compared with the rated operating range and equipment operating logic constraints extracted from the equipment nameplate parameters. If any parameter exceeds the rated range or violates the operating logic, it is directly determined that the step touches the physical hard boundary, the step is recorded as a hard boundary trigger step, and the generation of subsequent steps of the current path is terminated.

[0054] For example, suppose the nameplate parameters of a central air conditioning chiller unit specify: the compressor rated current is 100A, the maximum operating current is 110A; the evaporator outlet water temperature range is [5℃, 12℃]; the condenser outlet water temperature range is [30℃, 40℃]; the equipment operation logic constraints include: the oil pump must run when the compressor is running; the chilled water pump and cooling water pump must start before the compressor starts and stop after the compressor stops; suppose that at step 180 of the prediction sensor sequence, the compressor current prediction value reaches 112A, exceeding the above-mentioned maximum operating current of 110A, at this time it is directly determined that the step has touched the physical hard boundary, the hard boundary trigger step is recorded as 180, and the generation of subsequent steps of the current path is terminated.

[0055] Finally, the above process is repeated several times for the same target degradation trend. Random noise is sampled independently each time to obtain multiple valid walk paths and corresponding prediction sensor sequences. For example, 50 independent walk paths are generated for the condenser scaling degradation trend. Different random noise sequences are sampled independently each time. After time series rationality test and physical hard boundary verification, 42 valid walk paths and 8 invalid paths are obtained. The proportion of invalid paths is 0.16, which is lower than the preset threshold and does not trigger the signal reduction penalty.

[0056] Step S4: Determine the number of fault boundary steps for each effective walk path based on the hard boundary priority principle, and obtain the risk value and confidence weight of the target degradation trend based on the number of fault boundary steps and the group reference steps.

[0057] Specifically, step S4 can be divided into the following sub-steps: Step S41: Determine the number of fault boundary steps for each valid walk path based on the hard boundary priority principle. Specifically: For each valid walk path, first check if there are any hard boundary trigger step records; If it exists, the hard boundary trigger step is used as the fault boundary step number of the path; If there is no hard boundary trigger step record, the predicted values ​​of sensor parameters in each dimension are extracted step by step along the walking step sequence. The predicted values ​​of each dimension are compared with the preset safety threshold of the corresponding dimension. The step in which the predicted value of any dimension first exceeds the preset safety threshold is recorded as the fault boundary step of the path. If the path has no hard boundary trigger step records and the preset safety threshold is not triggered in all steps, the path is marked as not touching the fault boundary, and the fault boundary step count is recorded as the preset maximum walking step count.

[0058] For example, assuming the preset maximum walk step count is 300 steps, if a valid walk path is found to have a hard boundary trigger step record with 180 hard boundary trigger steps, then the fault boundary step count for this path is set to 180 steps; if a valid walk path has no hard boundary trigger step record, then the predicted values ​​of each dimension parameter are checked sequentially along the walk step sequence, and it is found that the predicted value of the condenser outlet water temperature reaches 38.2℃ at step 220, exceeding the warning threshold of 38℃ for the first time, then the fault boundary step count for this path is set to 220 steps; if a valid walk path does not trigger either the hard boundary or the preset safety threshold in all 300 steps, then this path is marked as not touching the fault boundary, and its fault boundary step count is set to the preset maximum walk step count of 300 steps.

[0059] Step S42, based on the number of fault boundary steps and the group reference steps, obtain the risk value and confidence weight of the target degradation trend, specifically: For the set of fault boundary steps of all effective paths under the same degradation trend, calculate the arithmetic mean and standard deviation of the boundary steps; The arithmetic mean is compared with the population reference steps for the degradation trend, and a truncation function is used to constrain the corresponding ratio to [0,1] to obtain the initial risk value. The truncation function means that when the ratio is less than 0, the risk value is 0; when the ratio is greater than 1, the risk value is 1; otherwise, the risk value is the ratio itself. The standard deviation is converted into the initial confidence weight using an inverse proportional mapping function, which can be expressed as: Initial confidence weight = 1 / (1 + standard deviation / reference standard deviation).

[0060] The reference standard deviation is the average standard deviation of the boundary steps of similar historical devices. If there is insufficient historical data, the reference standard deviation can be set by domain experts based on the group reference steps.

[0061] For example, set the reference standard deviation to 15% of the population reference steps.

[0062] The determination of whether a confidence reduction penalty is triggered is based on the proportion of invalid paths under this degradation trend. If a confidence reduction penalty is triggered, a corresponding confidence reduction penalty factor is generated to penalize the initial confidence weight, resulting in the final confidence weight. If a confidence reduction penalty is not triggered, the initial confidence weight is directly used as the final confidence weight.

[0063] Step S5: Based on the risk values ​​and confidence weights of all degradation trends in the degradation trend field, construct a fault susceptibility profile for the equipment, and combine it with a hierarchical early warning decision-making mechanism to issue fault warnings. Specifically: First, traverse all degradation trends in the degradation trend field, and execute the corresponding operation procedures in steps S3 and S4 for each degradation trend in sequence to obtain the set of fault risk values ​​and the set of final confidence weights for faults. Then, using the fault type label as the dimension identifier, organize the set of fault risk values ​​and the set of final confidence weights for faults into a risk aggregation matrix.

[0064] For example, the degradation trend field of a central air conditioning chiller unit includes five types of faults: condenser scaling, evaporator scaling, compressor wear, refrigerant leakage, and lubricating oil deterioration. For each of these five types of faults, the corresponding operation procedures in steps S3 and S4 are executed to obtain a risk aggregation matrix. In this example, the risk aggregation matrix is ​​represented as a 5x3 matrix, where each row corresponds to one of the five fault types, and each column corresponds to the risk value, confidence weight, and proportion of invalid paths, respectively.

[0065] Next, a radar chart mapping mechanism is used to generate a fault susceptibility profile, specifically: The fault type labels in the risk aggregation matrix are mapped to the radiation axes of the radar chart, the risk values ​​corresponding to each fault type are mapped to the numerical scales of the radiation axes, and the final confidence weights corresponding to each fault type are attached to the corresponding radiation axes in the form of numerical labels or node radii to generate a fault susceptibility radar chart for the current device. The fault susceptibility radar chart is a fault susceptibility profile.

[0066] For example, the fault susceptibility radar chart in the above case contains 5 radial axes, corresponding to condenser scaling, evaporator scaling, compressor wear, refrigerant leakage, and lubricating oil deterioration, respectively. The value range of each radial axis is 0 to 1.

[0067] Taking condenser scaling failure as an example, assuming its corresponding risk value is 1.0, its node on the fault susceptibility radar map is located at the outermost end of the radiation axis; the node radius is proportional to the confidence weight. Assuming its corresponding confidence weight is 0.461, the node radius of condenser scaling failure is 0.461 times the maximum radius.

[0068] It should be noted that the fault susceptibility profile also includes at least the basic equipment information (equipment number, model, installation location, and operating time), current operating status (operating mode, load rate, and ambient temperature), and historical risk change trends. The historical risk change trends are represented by the degradation trend risk value change curves for the most recent 7 days, 30 days, and 90 days, so that maintenance personnel can observe the evolution of the equipment's health status.

[0069] Finally, a threshold comparison and hierarchical early warning decision-making mechanism is adopted, triggering different levels of early warning signals based on the magnitude of the risk value and confidence weight. Specifically: A warning threshold is set, which is based on the historical statistical distribution of failures of similar devices. The risk value of each degradation trend is compared with the warning threshold one by one. When any risk value exceeds the warning threshold and the corresponding final confidence weight meets the requirements, a warning signal of the corresponding level is generated. The warning signal includes at least the fault type label that triggers the warning, the risk value and final confidence weight corresponding to the fault type, the effective perturbation walk path data corresponding to the degradation trend, the predicted sensor sequence data, and the proportion of invalid paths.

[0070] In one possible embodiment, the device's warning level is divided into the following three levels (this is only an example of one possible embodiment; the specific parameters need to be set according to the actual situation): Level 1 warning: Risk value ≥ 0.8 and final confidence weight ≥ 0.6, indicating that the equipment is about to fail and should be stopped and repaired immediately; Level 2 warning: 0.6 ≤ risk value < 0.8 and final confidence weight ≥ 0.5, indicating that the equipment has a high risk of failure and should be repaired within 24 hours; Level 3 warning: 0.4 ≤ risk value < 0.6 and final confidence weight ≥ 0.4. This indicates that the equipment has a potential risk of failure and an inspection should be arranged within 7 days.

[0071] Figure 2 This is a schematic diagram of a fault prediction system for intelligent building equipment according to the present invention.

[0072] Specifically, a fault prediction system for intelligent building equipment includes: The spatial construction module is used to construct and train an encoder-generator architecture that integrates physical constraints, and to construct a statistical soft boundary based on a first set of potential vectors. Simultaneously, it combines a preset fault degradation model to generate a degradation trend field and a group reference step number for each degradation trend. The state extraction module is used to generate the actual potential vector corresponding to the current sensor window data and the operating condition data, and to obtain the corresponding anchor vector by combining the abnormal feature extraction. The walk prediction module is used to set the target degradation trend based on the degradation trend field and simultaneously combine the anchor vector, and perform a multi-step walk in the potential space along the target degradation trend. Simultaneously, it combines generator reconstruction, temporal discrimination and physical hard boundary verification to obtain multiple effective walk paths and corresponding prediction sensor sequences. The risk quantification module is used to generate the number of fault boundary steps for each effective walk path according to the hard boundary priority principle, and to obtain the risk value and confidence weight in the target degradation trend based on the number of fault boundary steps and the group reference steps. The risk warning module is used to construct a fault susceptibility profile of equipment based on the risk value and confidence weight of all degradation trends in the degradation trend field, and simultaneously combine it with a hierarchical early warning decision mechanism to provide fault warnings.

[0073] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for predicting faults in intelligent building equipment, characterized in that, It includes the following steps: Construct and train an encoder-generator architecture that integrates physical constraints, construct a statistical soft boundary based on a first set of potential vectors, and generate a degradation trend field and a group reference step number for each degradation trend by combining a preset fault degradation model. The encoder-generator architecture includes at least an encoder, a working condition classifier, and a generator. Obtain the actual potential vectors corresponding to the current sensor window data and operating condition data, and use anomaly feature extraction to obtain the corresponding anchor vectors, specifically including: Acquire and preprocess the current sensor window data and operating condition data. The data preprocessing operations include at least outlier removal, missing value imputation, one-hot encoding of discrete data, and normalization of continuous data. The preprocessed current sensor window data and current operating condition data are input into the encoder for forward propagation to generate the actual potential vector. Based on preset rules, it is verified whether the actual potential vector is located inside the statistical soft boundary. If it exceeds the boundary, it is marked as a state offset and the corresponding offset distance and offset time are recorded. The abnormal feature extraction means using an unsupervised anomaly detection algorithm to scan the current sensor window data dimension by dimension and extract a set of statistically rare fragments with risk scores exceeding a preset threshold as abnormal features. The risk score is a quantitative score obtained by comprehensively analyzing the sensor's single-dimensional amplitude anomaly, temporal fluctuation anomaly, and multi-dimensional coupling anomaly. The higher the risk score, the greater the degree to which the corresponding sampling segment deviates from the normal operating distribution of the equipment. The abnormal features are input into a correction network based on a fully connected architecture for forward propagation to obtain anchor vectors; Based on the degradation trend field, the target degradation trend is set simultaneously with the anchoring vector. Multi-step potential space walks are performed along the target degradation trend. Multiple effective walk paths and corresponding predictive sensor sequences are generated through generator reconstruction, temporal discrimination and physical hard boundary verification. Based on the hard boundary priority principle, the number of fault boundary steps for each effective walk path is determined. Based on the number of fault boundary steps and the group reference steps, the risk value and confidence weight of the target degradation trend are obtained. Based on the risk value and confidence weight of all degradation trends in the degradation trend field, a fault susceptibility profile of equipment is constructed, and fault warning is carried out in combination with a hierarchical early warning decision mechanism.

2. The method for predicting faults in intelligent building equipment according to claim 1, characterized in that, A physical constraint-integrated encoder-generator architecture is constructed and trained. A statistical soft boundary is built based on the first latent vector set. A degradation trend field and a group reference step count for each degradation trend are generated by combining a pre-defined fault degradation model. This includes: Obtain past sensor datasets and past working condition datasets, and construct training samples according to the preset window length and sliding step size; Based on the training samples, train the encoder-generator architecture, input the previous working condition dataset into the pre-trained encoder, and generate the first latent vector set. A single-class classification algorithm is used to fit the distribution of the first potential vector set, with minimizing the envelope volume as the training objective, and the envelope boundary is defined by a preset quantile to generate a statistical soft boundary. Obtain a preset set of fault degradation models, and generate a corresponding degradation direction curve for each known fault in the preset set of fault degradation models. Construct a degradation trend field based on the degradation direction curve. Historical walk data of each building intelligent device under various degradation trends are obtained, and the historical walk data is extracted to generate corresponding group reference steps.

3. The method for predicting faults in intelligent building equipment according to claim 2, characterized in that, The encoder-generator architecture is trained based on the training samples, including: The encoder takes the previous sensor dataset and the previous working condition dataset as input conditions, maps the previous sensor dataset to the latent space, and generates a latent vector; The working condition classifier takes the latent vector as input, uses the cross-entropy loss of working condition classification as the loss function, and alternately updates the parameters of the encoder and the working condition classifier until the preset convergence condition is reached. The generator takes the latent vector and the previous working condition dataset as joint input and reconstructs the sensor data sequence through deconvolution or upsampling operations; The mean square error between the reconstructed sensor data sequence and the previous sensor dataset is used as the reconstruction loss term, and a joint loss function is constructed by combining it with a preset physical constraint term; The generator is trained with the goal of minimizing the joint loss function, and the generator network parameters are updated using the gradient backpropagation algorithm until the total loss function value is lower than the preset convergence threshold.

4. The method for predicting faults in intelligent building equipment according to claim 2, characterized in that, Obtain a preset set of fault degradation models, and generate a corresponding degradation direction curve for each known fault type in the preset set of fault degradation models. Construct a degradation trend field based on the degradation direction curves, including: For each known fault type, discrete sampling points of degradation parameters covering the complete degradation process are set according to the preset physical degradation model of the corresponding fault type; Run the preset equipment simulation model calibrated with past operating data at each sampling point and output multi-dimensional simulation data corresponding to the degradation stage; The multidimensional simulation data corresponding to each degradation stage is input into the pre-trained encoder to generate a degradation potential vector sequence. The degradation potential vector sequence is smoothly fitted using cubic spline interpolation or Bézier curve fitting algorithms to generate degradation direction curves corresponding to the fault types. Based on the degradation direction curves of all fault types, a degradation trend field is constructed using the potential spatial coordinates as the index key and the fault type label as the value.

5. The method for predicting faults in intelligent building equipment according to claim 1, characterized in that, Based on the degradation trend field, the target degradation trend is set simultaneously with the anchoring vector. A multi-step latent space walk is performed along the target degradation trend, and multiple valid walk paths and corresponding predicted sensor sequences are generated through generator reconstruction, temporal discrimination, and physical hard boundary verification, including: Select a target degradation trend curve from the degradation trend field, and extract the tangent vector of the target degradation trend curve at the actual potential vector position; The tangent vector and the anchoring vector are weighted and fused to obtain the corrected target degradation trend; Starting from the position corresponding to the actual potential vector, a multi-step walk operation is performed along the corrected target degradation trend to generate a perturbation potential sequence; The potential vectors at each step of the perturbation potential sequence are paired with the current operating condition data and input point by point into the generator for forward propagation to generate a predictive sensor sequence. The predicted sensor sequence is checked and corrected for time sequence rationality based on a preset time sequence discriminator, and the path that passes the time sequence rationality check and correction is defined as a valid walk path. Physical hard boundary verification is performed on the valid walking path, and whether the preset physical hard boundary is triggered is recorded synchronously. If the preset physical hard boundary is triggered, the generation of subsequent steps of the current path is terminated. If the preset physical hard boundary is not triggered, continue to perform multiple potential space walks until the preset maximum number of walks is reached, and obtain multiple valid walk paths and corresponding prediction sensor sequences.

6. The method for predicting faults in intelligent building equipment according to claim 5, characterized in that, Multi-step walk operation, including: A directional perturbation term is generated based on the actual potential vector of the current step and the corrected target degradation trend, and random noise is sampled from a preset Gaussian distribution; The directional perturbation term is superimposed with random noise to generate the next potential vector, and the potential vectors of each step are stored sequentially until the number of walking steps reaches the preset maximum number of walking steps or the physical hard boundary verification is triggered, thus generating a perturbation potential sequence.

7. The method for predicting faults in intelligent building equipment according to claim 1, characterized in that, Based on the hard boundary priority principle, the number of fault boundary steps for each effective walk path is determined. Then, based on the number of fault boundary steps and the group reference steps, the risk value and confidence weight for the target degradation trend are obtained, including: For each valid walk path, check whether there are hard boundary trigger step records; If a hard boundary trigger step record exists, then the hard boundary trigger step is used as the fault boundary step number for that path; If there is no hard boundary trigger step record, the predicted values ​​of sensor parameters in each dimension are extracted step by step along the walking step sequence, and the predicted values ​​are compared with the corresponding preset safety threshold. The first step in which the predicted value of any dimension exceeds the preset safety threshold is recorded as the fault boundary step of the corresponding path; If no hard boundary trigger steps are recorded on the valid walking path, and the preset safety threshold is not triggered in all steps, then the path is marked as not touching the fault boundary, and the number of fault boundary steps is recorded as the preset maximum walking steps. Statistical features are extracted from the set of fault boundary steps corresponding to all valid walk paths to obtain a statistical feature set. Based on the statistical feature set and the group reference steps, the risk value and confidence weight of the target degradation trend are generated simultaneously by combining the preset mapping rules.

8. The method for predicting faults in intelligent building equipment according to claim 1, characterized in that, Based on the risk values ​​and confidence weights of all degradation trends in the degradation trend field, a fault susceptibility profile of the equipment is constructed, and fault warnings are issued in conjunction with a hierarchical early warning decision-making mechanism, including: Traverse all degradation trends in the degradation trend field, and execute the operation process of steps S3 and S4 in sequence for each degradation trend to obtain the risk value and confidence weight of all degradation trends. A risk aggregation matrix is ​​constructed based on the risk values ​​and confidence weights of all the aforementioned degradation trends, and a fault susceptibility profile is generated simultaneously by combining a preset radar chart mapping mechanism. Based on the aforementioned fault susceptibility profile and hierarchical early warning decision-making mechanism, fault early warning is conducted.

9. A fault prediction system for intelligent building equipment, used to implement the method described in any one of claims 1 to 8, characterized in that, include: The spatial construction module is used to construct and train an encoder-generator architecture that integrates physical constraints, and to construct a statistical soft boundary based on a first set of potential vectors. Simultaneously, it combines a preset fault degradation model to generate a degradation trend field and a group reference step number for each degradation trend. The state extraction module is used to generate the actual potential vector corresponding to the current sensor window data and the operating condition data, and to obtain the corresponding anchor vector by combining the abnormal feature extraction. The walk prediction module is used to set the target degradation trend based on the degradation trend field and simultaneously combine the anchor vector, and perform a multi-step walk in the potential space along the target degradation trend. Simultaneously, it combines generator reconstruction, temporal discrimination and physical hard boundary verification to obtain multiple effective walk paths and corresponding prediction sensor sequences. The risk quantification module is used to generate the number of fault boundary steps for each effective walk path according to the hard boundary priority principle, and to obtain the risk value and confidence weight in the target degradation trend based on the number of fault boundary steps and the group reference steps. The risk warning module is used to construct a fault susceptibility profile of equipment based on the risk value and confidence weight of all degradation trends in the degradation trend field, and simultaneously combine it with a hierarchical early warning decision mechanism to provide fault warnings.

Citation Information

Patent Citations

  • Hydraulic power plant edge calculation method and system with intelligent diagnosis function

    CN120725646A

  • Fault trend prediction method and device based on operation data of communication equipment

    CN122226633A