Intelligent building fault prediction method based on adaptive algorithm

Abstract

The invention discloses an intelligent building fault prediction method based on an adaptive algorithm, and the method comprises the steps: obtaining operation state data collected by a plurality of sensor nodes, carrying out the normalization preprocessing, inputting an improved adaptive prediction model, and obtaining the implicit state representation; an initial fuzzy rule set is generated through fuzzy rule induction, a real-time evolution strategy is introduced, fuzzy rules and neural connection weights are dynamically adjusted according to error feedback, and a self-adaptive prediction model updated through evolution is formed; and then performing multi-step prediction on the key operation parameter time sequence by using the updated model, comparing a prediction result with a dynamic fault threshold value to generate a candidate fault indication, and determining the position and category of a fault to be pre-warned in the building system through time sequence stability verification and space consistency analysis. The method can significantly improve the accuracy and real-time performance of intelligent building fault prediction, and has a good application prospect.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent buildings, and in particular to an intelligent building fault prediction method based on an adaptive algorithm. Background Art

[0002] In the field of smart buildings, achieving accurate perception of building operating status and early prediction of faults has always been a key issue in intelligent operation and maintenance systems. With the expansion of building scale and the increase in the number of smart devices, the dynamic complexity within building systems continues to increase. Traditional methods that rely on threshold setting, rule triggering, or simple statistical analysis are no longer able to meet the needs of real-time detection and prediction of system anomalies. Especially in scenarios where multi-source heterogeneous data frequently interacts and the building environment changes rapidly, existing fault prediction methods based on fixed models or static rules often suffer from poor adaptability, delayed predictions, and high false alarm and missed alarm rates, limiting the further advancement of intelligent building operation and maintenance.

[0003] To improve the accuracy and robustness of fault prediction, researchers have introduced data-driven methods such as machine learning and deep learning. By modeling the mapping relationship between historical data and future states, they attempt to overcome the limitations of traditional methods. However, existing data-driven methods generally face two challenges: First, the model's ability to adapt to changes in the building's operating environment is insufficient, making it difficult to cope with data distribution drift caused by factors such as equipment aging and load changes during long-term use; second, the model's updates lack real-time performance, making it impossible to adjust internal structures or parameters in a timely manner after detecting prediction errors, resulting in continued performance degradation.

[0004] In the area of adaptive prediction technology, some work has attempted to enhance the model's ability to handle complex dynamic environments by introducing adaptive adjustments to neural networks or fuzzy logic reasoning. For example, fuzzy neural networks combine the interpretability of fuzzy reasoning with the learning capabilities of neural networks, improving the flexibility and generalization of prediction systems to a certain extent. However, existing fuzzy neural network methods are mostly based on offline training. Once deployed, they struggle to dynamically evolve and adjust based on actual prediction errors, and thus remain unable to meet the dual requirements of real-time and adaptability for intelligent building fault prediction.

[0005] Furthermore, in the fault detection process, traditional methods are mostly based on single-step prediction residuals or simple threshold comparisons. They lack systematic modeling of the temporal and spatial correlation characteristics of prediction deviations, resulting in insufficient fault location and classification accuracy. Especially for complex anomalies such as hidden faults and progressive degradation, detection methods based solely on static models have difficulty promptly identifying and accurately defining the fault area and nature.

[0006] Therefore, how to provide an intelligent building fault prediction method based on an adaptive algorithm is an urgent problem that those skilled in the art need to solve. Summary of the Invention

[0007] One purpose of the present invention is to propose a smart building fault prediction method based on an adaptive algorithm. The present invention integrates the operating status data of multiple sensor nodes and the improved fuzzy neural network modeling technology. By introducing a real-time evolutionary strategy, the dynamic optimization adjustment of the fuzzy rule set and the neural connection weights is realized. The future state estimation of the key operating parameters of the building is generated based on multi-step time series prediction, and combined with the prediction deviation measurement, dynamic fault detection threshold comparison, time series stability verification and spatial consistency analysis, it finally realizes the accurate early warning of the potential fault location and category of the intelligent building system. It has the advantages of high prediction accuracy, strong real-time adaptability and high fault location accuracy.

[0008] According to an embodiment of the present invention, a method for predicting intelligent building faults based on an adaptive algorithm includes the following steps: S1. Obtain the operating status data collected by multiple sensor nodes in the target intelligent building, and perform normalization preprocessing on the operating status data to form a standardized data set; S2. Inputting the standardized data set into the improved adaptive prediction model to obtain a latent state representation; S3, performing fuzzy rule induction on the implicit state representation, constructing an initial fuzzy rule set based on the fuzzy membership function, and generating a first-stage prediction result; S4. Introducing a real-time evolutionary strategy to dynamically adjust the neural connection weights of the initial fuzzy rule set and the improved adaptive prediction model based on the error feedback of the prediction results of the first stage, thereby generating an evolved and updated adaptive prediction model; S5. Using the adaptive prediction model that has been updated through evolution, predict the time series consisting of key operating parameters of the target intelligent building and output multi-step prediction results for multiple time points in the future; S6. construct a prediction deviation metric sequence based on the multi-step prediction results, compare the prediction deviation metric sequence with a preset fault threshold, and generate a candidate fault indication; S7. Perform temporal stability verification and spatial consistency analysis on the candidate fault indications. If predefined consistency criteria are met, output the location of the fault to be warned and the fault category in the building system.

[0009] This paper proposes a fault prediction method for intelligent buildings based on an adaptive algorithm. It establishes a complete process from operational status data collection, standardization, feature encoding, fuzzy rule induction, real-time evolutionary optimization, multi-step prediction, deviation measurement, to fault indication output. This method effectively addresses the low fault prediction accuracy, delayed response, and early warning difficulties in existing intelligent building operations and maintenance. This method combines the nonlinear feature modeling capabilities of fuzzy neural networks with the dynamic adaptive optimization mechanism of real-time evolutionary strategies. By extracting latent state features and constructing a fuzzy rule set, it achieves in-depth modeling of the complex state evolution trends of building systems. Combined with error feedback-driven two-stage simultaneous evolution, it continuously optimizes fuzzy rules and neural connection weights to form a highly adaptable prediction model. Based on this adaptive prediction model, the present invention enables multi-step prediction of key operating parameters at multiple future time points on a continuous time scale. Combining sliding statistics with a dynamic fault threshold mechanism, it accurately identifies deviation-related anomalies and generates reliable candidate fault indications. Furthermore, through temporal stability verification and spatial consistency clustering analysis, the method comprehensively assesses the persistence and spatial connectivity of candidate faults, accurately determining the fault location and fault category, significantly improving the prediction and early warning capabilities and maintenance response efficiency of intelligent building systems. This method is real-time, adaptive, accurate and systematic, and can be efficiently deployed and applied in complex operation and maintenance environments such as large-scale intelligent buildings, complexes, and commercial buildings. It has important application value in improving the proactive operation and maintenance capabilities of intelligent buildings and reducing failure rates and maintenance costs.

[0010] Optionally, the S2 specifically includes: S21. Organize the standardized data set in chronological order to construct a time series input set consisting of multiple time series samples, wherein the time series samples include building operation status data of multiple sensor nodes in the standardized data set at various time points; S22, inputting each time series sample in the time series input set into the improved adaptive prediction model in sequence, passing the building operation status data of each sensor node into the fuzzification layer of the improved adaptive prediction model through the input layer structure of the improved adaptive prediction model, and performing fuzzy membership calculation on the building operation status data of each dimension to generate a corresponding fuzzy input representation; S23, performing a fuzzy rule matching operation based on the fuzzy input representation, calling a fuzzy rule set in the improved adaptive prediction model, calculating a matching degree between corresponding time series samples in the time series input set and the fuzzy rule set, and obtaining a fuzzy rule activation value set; S24. Passing the fuzzy rule activation value set as an input signal to the rule inference layer in the improved adaptive prediction model, and generating an implicit state representation by fusing the fuzzy rule activation value with the corresponding neural connection weight.

[0011] By carefully defining the construction of time series input sample sets, implicit state generation, and fuzzy rule matching and inference processes, this invention ensures strict alignment between input data and the internal structure of the prediction model, effectively avoiding the problems of data time series distortion, ambiguous state feature extraction, and increased inference errors in traditional intelligent building fault prediction. By employing a standardized dataset to construct multiple time series sample sets based on chronological order, this method ensures the coherence and consistency of operating state data within the temporal domain, providing an accurate sequential basis for implicit state generation. By sequentially inputting time series samples into an improved adaptive prediction model, and utilizing layer-by-layer transfer from the input layer to the hidden layer and feature encoding, the one-dimensional mapping rules for each sensor node data in the implicit state space are strictly defined, improving the accuracy and expressiveness of state feature extraction. In particular, through fuzzy rule attribution calculation and feature mapping, the raw operating state data is converted into a high-level implicit state representation, which not only reduces the redundancy of the original features but also enhances the sensitivity and distinguishability of the features to future state change trends. During the fuzzy rule inference stage, this invention designs a dynamic inference mechanism based on the degree of matching between time series samples and fuzzy rules, ensuring high matching accuracy and inference stability during fuzzy rule activation. In the overall process, the structural unification, data closure and state constraints of each link from input to reasoning have significantly improved the data consistency, model stability and prediction reliability from multi-time series input to the prediction and reasoning stage in intelligent building fault prediction, providing a high-quality state foundation for fault deviation detection and early warning output, and comprehensively improving the overall performance and application value of the system.

[0012] Optionally, the improved adaptive prediction model specifically includes: Constructing an initial fuzzy neural network structure, the structure comprising an input layer, a fuzzification layer, a fuzzy rule layer, a rule inference layer, and an output layer, wherein the fuzzy rule layer is composed of a plurality of initial fuzzy rules, each of which has an initial membership function parameter and an initial neural connection weight; Define the prediction error feedback indicator, calculate the real-time prediction error sequence, and calculate the real-time error sensitivity based on the real-time prediction error sequence : ; in, Indicates that at the current moment Next Fuzzy rules and the corresponding neural connection weights Real-time error sensitivity to forecast errors, Indicates that at the current moment The real-time prediction error is Indicates the current time Next The initial membership function parameters of the fuzzy rules, Indicates the current time Next The neural connection weights corresponding to the fuzzy rules; Based on the real-time error sensitivity, a unified evolutionary criterion is established for the fuzzy rule set and the neural connection weights. The real-time prediction error at the current moment is taken as the only optimization target, and the fuzzy rule set and the neural connection weights are simultaneously included in the synchronous iterative update range. A two-phase alternating update process is performed in real time: Rule evolution stage: based on real-time error sensitivity , adjust the membership function parameters of the fuzzy rule structure of the fuzzy rule layer: ; in, Indicates the first Fuzzy rules at the current moment The updated membership function parameters of Indicates the Fuzzy rules at the current moment Membership function parameters before updating, The step size coefficient for adjusting the fuzzy rule parameters; Weight evolution stage: Based on the membership function parameters of the updated fuzzy rule structure, the neural connection weights are further dynamically adjusted: ; in, Indicates the first The neural connection weights corresponding to the fuzzy rules are at time The updated value, Indicates the The neural connection weights corresponding to the fuzzy rules are at time The value before updating, The step size coefficient for adjusting the neural connection weights; By repeatedly executing the above two-stage alternating updating process in real time until the real-time prediction error converges to a predetermined threshold range, an improved adaptive prediction model after the two-stage alternating updating and adjustment is obtained.

[0013] By constructing an initial fuzzy neural network structure consisting of an input layer, a fuzzification layer, a fuzzy rule layer, and an inference output layer, and setting initial connection weights between the fuzzy rule layer and the inference layer, a two-stage alternating update mechanism driven by real-time error feedback is proposed, significantly improving the adaptability and prediction accuracy of the intelligent building fault prediction model. By defining a real-time prediction error sensitivity index to quantify the contribution of fuzzy rule parameters and neural connection weights to the overall prediction error, within each evolutionary cycle, the membership function parameters of the fuzzy rules are first dynamically modified based on the error sensitivity, with the adjustment amplitude controlled by the step size coefficient to adapt in real time to changes in the building's operating status. Subsequently, based on the updated fuzzy rules, the corresponding neural connection weights are further synchronously adjusted to ensure the consistency of the model inference process and error convergence. By continuously executing two-stage alternating optimization, this invention achieves rapid convergence of error sensitivity parameters and deep structural adaptation, avoiding the performance degradation caused by fixed parameters in traditional models and improving prediction robustness in changing environments. The overall optimization process forms a closed-loop dynamic update mechanism, which not only improves the response speed and stability of intelligent building fault prediction, but also enhances the model's ability to continuously learn from environmental changes, providing more forward-looking fault warning support for the building operation and maintenance system.

[0014] Optionally, the S3 specifically includes: S31, arranging the implicit state representations into an implicit state sequence in chronological order, and determining the maximum and minimum value ranges of the implicit state values of each dimension; S32. Divide the implicit state value of each dimension into a preset number of clusters using the K-means clustering algorithm within the maximum and minimum value intervals, using the sum of squared errors within the clusters reaching a local minimum as the division convergence criterion, determine the center value and the width value of each cluster, and set the center value and the width value of each cluster as the center parameter and the width parameter of the Gaussian membership function, respectively, to construct a Gaussian membership function family for the implicit state value of each dimension; S33, using the Gaussian membership function family, calculating the membership of the implicit state value of each dimension in the implicit state sequence, performing fuzzy label mapping on the membership of all dimensions at the same time point, and obtaining the fuzzy label combination corresponding to the time point; S34. Based on the fuzzy label combination at each time point, the fuzzy labels in all fuzzy label combinations are arranged and combined, and an initial fuzzy rule set is constructed according to the "if-and-then" semantic format, wherein the antecedent of each initial fuzzy rule in the initial fuzzy rule set is formed by the fuzzy label combination, and the consequent is the output label to be predicted; S35. For each initial fuzzy rule in the initial fuzzy rule set, use the product T-norm as a rule activation calculation operator to calculate the rule activation and generate a rule activation set; S36: Perform weighted normalization on the rule activation set to calculate the first-stage prediction result: ; in, For time point The first stage prediction results, For the The activation degree of the initial fuzzy rules, For the The output value corresponding to the consequent “output label to be predicted” of the initial fuzzy rule, is the total number of initial fuzzy rules.

[0015] By introducing normalization, clustering, Gaussian membership function modeling, and automatic fuzzy rule generation based on implicit state representation, this paper effectively addresses the problem of prediction failure in existing intelligent building fault prediction systems caused by rough state space partitioning, arbitrary rule generation, and insufficient reasoning accuracy. The K-means clustering algorithm is used to optimally cluster the implicit state data of each dimension within the range of maximum and minimum values. The clustering termination condition is when the sum of squared classification errors reaches a local minimum, ensuring the physical rationality and classification compactness of the cluster center values. On this basis, a family of Gaussian membership functions is constructed with cluster center values and cluster widths as parameters, ensuring the continuity and discrimination of fuzzy membership descriptions. By introducing a fuzzy label combination mechanism at the same time point, the initial fuzzy rule set is automatically generated according to the "if-and-then" semantic rule, strictly defining the mapping relationship between input features and predicted outputs, and avoiding the redundancy or conflict caused by the manual specification of traditional rules. During the fuzzy inference phase, this paper designs a weighted normalized fuzzy inference output mechanism. This weighted average of prediction results is performed based on rule activation, ensuring that the inference process precisely responds to the contribution of different rules, significantly improving the smoothness and accuracy of the prediction output. The overall process, from implicit state normalization and membership modeling to rule combination reasoning, forms a highly self-consistent, data-driven fuzzy inference system, providing a precise and robust input foundation for multi-step predictions in intelligent building systems.

[0016] Optionally, the S4 specifically includes: S41. Obtain the first-stage prediction result and the corresponding actual operating status value at each time point, calculate the real-time prediction error between the first-stage prediction result and the actual operating status value, and form a real-time prediction error sequence in chronological order; S42, introducing a real-time evolution strategy, calculating a real-time error sensitivity according to the real-time prediction error sequence, and performing an error sensitivity evaluation on the membership function parameters of each initial fuzzy rule in the initial fuzzy rule set; S43, in the rule evolution stage of the real-time evolution strategy, adaptively adjusting the parameters of each membership function according to the step coefficient of the fuzzy rule parameter adjustment to obtain an updated fuzzy rule set; S44. In the weight evolution phase of the real-time evolution strategy, adaptively adjust the neural connection weights of the improved adaptive prediction model according to the step coefficient of the neural connection weight adjustment and in combination with the real-time error sensitivity to obtain an updated neural connection weight set; S45, synchronously writing the updated fuzzy rule set and the updated neural connection weight set into the improved adaptive prediction model to form an evolutionary update iterative round, and repeatedly executing S42 to S44 until the real-time prediction error converges to a set threshold; S46. When the convergence criterion of the real-time evolution strategy is met, the adaptive prediction model updated through evolution is output.

[0017] By introducing an evolutionary triggering mechanism based on real-time prediction error mean statistics within a fixed-length sliding time window and a dynamic step-size coefficient management strategy, an efficient and accurate real-time adaptive evolutionary control process is established for intelligent building fault prediction systems. This effectively addresses the issues of adjustment lag, fixed step-size, and high evolutionary volatility inherent in traditional adaptive models in dynamic environments. A single evolutionary triggering threshold and a single convergence threshold judgment strategy ensure that evolutionary cycles are triggered only when the mean prediction error increases significantly, avoiding model oscillation and resource waste caused by excessive evolution. An evolutionary parameter registration table is established to record the current values and allowable adjustment ranges of the step-size coefficients for fuzzy rule parameters and neural connection weights, enabling precise dynamic parameter control for each evolutionary cycle. During the evolutionary process, fuzzy rule attribution parameters and neural connection weight parameters are independently adjusted based on the real-time error sensitivity component, and each adjustment is immediately written back to the corresponding registration table to maintain the continuity and adaptability of the evolutionary path. After a two-stage evolutionary cycle, the system recalculates the sliding window error mean. If the mean drops below the convergence threshold, the current evolutionary cycle is terminated; otherwise, the optimization cycle continues, forming a closed-loop dynamic update mechanism. This design not only improves the real-time response speed of the intelligent building fault prediction model in a changing operating environment, but also greatly enhances the accuracy of model structure adjustment and the overall robustness of the system, significantly extending the system's failure-free continuous operation time.

[0018] Optionally, the real-time evolution strategy specifically includes: Setting a single evolution trigger threshold and a single convergence threshold, and establishing an evolution parameter registration table for the fuzzy rule set and the neural connection weight, wherein the evolution parameter registration table records the current values and adjustable intervals of the step coefficients for adjusting the fuzzy rule parameters and the neural connection weight; The window mean of the real-time prediction error sequence is continuously counted using a fixed-length sliding time window. When the window mean is greater than the single evolution trigger threshold, an evolution cycle is initiated. In the rule evolution stage, the membership function parameters are adaptively adjusted according to the real-time error sensitivity. The adjustment range is controlled by the step coefficient of the fuzzy rule parameter adjustment in the evolution parameter registration table. After the adjustment is completed, the updated fuzzy rule set is immediately written back. In the weight evolution stage, the neural connection weights are adaptively adjusted according to the real-time error sensitivity. The adjustment range is controlled by the step coefficient of the neural connection weight adjustment in the evolution parameter registration table. After the adjustment is completed, the updated neural connection weight set is immediately written back. After completing the rule evolution stage and the weight evolution stage, the window mean is recalculated and compared with the single convergence threshold. When the window mean is not higher than the single convergence threshold or the evolution cycle has reached the maximum number of allowed iterations, the evolution is terminated. Otherwise, return to step S63 to continue executing the next round of evolution cycle.

[0019] This invention establishes a complete, controllable, and efficient adaptive optimization process by designing a single evolution trigger threshold and a single convergence threshold within a real-time evolution strategy, and by introducing a dynamic registration mechanism for the step-size coefficients for adjusting fuzzy rule parameters and neural connection weights. This process effectively addresses the frequent loss of control, severe error fluctuations, and poor system stability inherent in existing intelligent building fault prediction systems. A fixed-length sliding time window is used to calculate the mean of the real-time prediction error sequence. When the window mean exceeds the evolution trigger threshold, a two-stage evolution process is initiated. During the rule evolution stage, fuzzy rule attribution parameters are adaptively adjusted based on real-time error sensitivity, and the adjusted step-size coefficients are simultaneously updated in a registration table. During the weight evolution stage, neural connection weights are synchronously adjusted based on error sensitivity, and the step-size coefficients are also updated, ensuring that each parameter update is historically traceable and controllable in real time. After each round of two-stage evolution, the system recalculates the window mean and compares it with the convergence threshold. If the mean falls below the set standard, the evolution round is terminated; otherwise, the process automatically returns to the next round. This mechanism dynamically links the prediction error trend and parameter step adjustment to form an error-aware driven intelligent evolutionary system. It not only avoids model instability caused by excessive parameter updates, but also significantly improves the continuous adaptability and anti-interference ability of the prediction system in complex building environments, providing a solid guarantee for achieving high-precision, low-false-alarm intelligent building fault warning.

[0020] Optionally, the S5 specifically includes: S51. Using the adaptive prediction model that has been updated through evolution, obtain a time series consisting of key operating parameters of the target intelligent building, and set a prediction start point and a prediction end point according to the time span and data density of the time series; S52. Based on the current moment, a time series consisting of key operating parameters within a fixed-length time window in the past is input into the adaptive prediction model that has been updated through evolution, and a prediction result for the first prediction moment is obtained through reasoning based on the fuzzy rules and neural network structure within the adaptive prediction model that has been updated through evolution; S53. At the next prediction moment, append the prediction result output at the previous prediction moment to the end of the input key operating parameter time series to form an updated key operating parameter time series, input the updated key operating parameter time series into the adaptive prediction model that has been updated through evolution, and generate a prediction result for the next prediction moment; S54, repeatedly executing step S53, gradually predicting the key operating parameters at multiple consecutive future prediction moments, and forming a multi-step prediction result sequence for multiple future time points; S55. Perform linear interpolation processing on the generated multi-step prediction result sequence to unify the time interval and data dimension of the multi-step prediction result sequence and generate a normalized multi-step prediction result.

[0021] By introducing a rolling window mechanism based on the current moment and combining it with an adaptive prediction model that has been updated through evolution, the time series of key operating parameters of the target intelligent building are gradually recursively predicted to form a continuous multi-step prediction result sequence, effectively solving the problems of short prediction time span and poor prediction consistency in existing intelligent building systems. Historical data within a fixed-length time window in the past is used as dynamic input, and the updated model is used to infer the prediction results for the first future prediction moment. On this basis, the prediction results are cyclically appended, gradually extending the prediction time span to construct a continuous future time series chain. By introducing interpolation processing and time interval standardization steps in the multi-step prediction sequence, the prediction results are ensured to have uniform sampling density and time series consistency in the time domain. This method significantly improves the stability and accuracy of multi-point, multi-parameter continuous predictions, providing reliable support for fault trend judgment, warning window extension, and maintenance strategy optimization in intelligent buildings.

[0022] Optionally, the S6 specifically includes: S61, pairing the normalized multi-step prediction result sequence with the actual operating status data of the target intelligent building at the corresponding time point, calculating the prediction residual point by point in time axis order to form a prediction residual sequence, where the prediction residual is the difference between the multi-step prediction result and the actual operating status data; S62. Based on the prediction residual sequence, a prediction deviation measurement sequence is constructed using a dual statistical method of sliding mean and sliding standard deviation, wherein the prediction deviation measurement value corresponding to each time point is obtained by weighted fusion of the residual mean and residual standard deviation in a fixed window before and after the time point; S63. Based on the prediction deviation measurement sequence, a unique dynamic fault detection threshold is set. The dynamic fault detection threshold is generated based on the residual mean and standard deviation statistics during the historical normal operation of the intelligent building, and is adaptively adjusted synchronously with the system operation load level. S64. Compare the predicted deviation metric value at each time point with the dynamic fault detection threshold at the corresponding moment. If the predicted deviation metric values at multiple consecutive time points exceed the corresponding dynamic fault detection threshold, extract the center position within the continuous exceeding limit interval as a candidate fault indication. The candidate fault indication includes the time point of fault occurrence, duration, maximum deviation amplitude and maximum deviation moment.

[0023] By matching the normalized multi-step prediction result sequence with the actual operating status data at different time points, a prediction residual sequence is formed, and on this basis, a prediction deviation measurement sequence is constructed using the dual statistical methods of sliding mean and sliding standard deviation, which effectively improves the ability to sensitively capture abnormal trend changes in the operating status of intelligent buildings. By dynamically setting a unique fault detection threshold based on the prediction deviation measurement sequence, and adaptively adjusting the detection sensitivity according to the deviation level during historical normal operation, the consistency and adaptability of fault detection under different load environments are achieved. Furthermore, in the candidate fault extraction process, the center position within the continuous over-limit interval is used as the representative point of the fault event, and the time of occurrence, duration, maximum deviation amplitude and maximum deviation moment of the fault are recorded to form structured fault indication data, which significantly improves the accuracy and stability of the fault detection results. This method effectively compensates for the high false alarm rate and serious omission of the traditional single-point threshold method, and is suitable for intelligent building fault prediction scenarios in complex dynamic environments.

[0024] Optionally, the S7 specifically includes: S71. Based on the fault occurrence time, duration, maximum deviation amplitude, and maximum deviation moment recorded in the candidate fault indication, a time continuity verification model is constructed within a fixed timing window. The time interval distribution of consecutive candidate fault indications is statistically analyzed. If the consecutive fault interval is less than a set continuity verification threshold and the duration is greater than a minimum stable duration, it is determined to be a timing stable fault event. S72. Under each temporally stable fault event, based on the spatial layout relationship in the intelligent building, call the operating status data of all sensor nodes that overlap with the fault occurrence time period and construct the state similarity between the nodes. The state similarity is calculated based on the synchronous change trend and change amplitude of the node data. S73. Based on the state similarity, perform spatial consistency cluster analysis to extract a node subset with significant spatial correlation. If the number of nodes in the node subset exceeds a set spatial consistency criterion and the nodes cover a continuous area of the building's physical space, then confirm that the spatial consistency condition is met. S74. Under the premise of simultaneously satisfying the temporal stability verification and spatial consistency analysis criteria, extract the node with the highest abnormality in the node subset as the fault center node, and combine the fault center node category and the monitored operating parameter type to calibrate the location and fault category information of the fault to be warned in the building system; S75. Output the location and fault category information of the fault to be warned that has been calibrated.

[0025] By introducing a dual verification mechanism of temporal continuity verification and spatial consistency analysis based on candidate fault indications, the accuracy and stability of fault identification in the intelligent building fault prediction system have been significantly improved. A fixed time series window is used to perform statistical analysis on the distribution of fault time intervals. If the interval between consecutive fault indications is less than the set threshold and the duration exceeds the stability standard, it is determined to be a temporally stable fault event, effectively eliminating the influence of isolated noise points. In the spatial consistency analysis, based on the node operation status data, a node status similarity matrix is constructed and clustered based on the position relationship. Node groups with physical spatial connectivity and consistent state change trends are screened out, thereby improving the physical rationality of fault location. Further combining the results of temporal verification and spatial clustering, the fault center node is extracted and the fault location and category information is generated to form a structured warning output. This method effectively avoids the problem of misjudgment of a single abnormal point and achieves high-precision confirmation of the fault area and type in a complex building environment.

[0026] Optionally, the spatial consistency analysis specifically includes: Normalizing the state similarity and mapping the state similarities between all nodes to a unified numerical range; Based on the normalized state similarity between nodes, a density peak clustering algorithm is used to automatically identify areas with abnormally high local state density and extract the node subset corresponding to each density peak. In each node subset, the spatial physical distance matrix between nodes is calculated, and node pairs with continuous physical distances and state similarity higher than the set spatial correlation threshold are screened out to construct spatially continuous sub-regions; In each spatial continuous sub-region, the node abnormality distribution is statistically analyzed, and the spatial sub-regions whose average abnormality is greater than the normal distribution threshold are selected as the fault-related node set; If the number of nodes in the final fault-associated node set reaches the spatial consistency criterion and the physical distribution of the nodes constitutes a connected area inside the building, the spatial consistency condition is determined to be met.

[0027] By normalizing the state similarity between nodes and combining it with a density peak clustering algorithm to automatically identify areas of high local anomaly density, a refined spatial consistency analysis process was constructed for intelligent building systems. This effectively addresses the issues of sparse node distribution and large clustering errors in traditional anomaly detection. Based on the node subset extracted based on the standardized state similarity, the physical distance matrix between nodes was further calculated to identify groups of nodes with high spatial physical connectivity and consistent state change trends, thereby establishing spatially connected subregions. Within these subregions, the distribution of node anomalies was statistically analyzed to identify areas with a mean anomaly value higher than the normal distribution as the set of fault-associated nodes. Finally, the spatial consistency criterion was determined based on the number of nodes and the spatial physical connectivity conditions. This method fully integrates state change trends with physical spatial relationships, improving fault detection accuracy while significantly reducing false alarm rates. This provides a reliable foundation for accurately determining fault locations and developing early warning strategies in intelligent buildings.

[0028] The beneficial effects of the present invention are: (1) This invention introduces an improved adaptive prediction model that integrates fuzzy neural networks and real-time evolution strategies. After acquiring building operation status data, it can dynamically adjust the fuzzy rule set and neural connection weights based on real-time prediction errors, forming a prediction system with self-evolutionary capabilities. This method effectively overcomes the problems of poor adaptability and update lag of traditional static modeling methods, significantly improves the real-time response capability and long-term stability of building fault prediction, and ensures that the system can maintain a high level of prediction accuracy in complex and dynamic building environments.

[0029] (2) This paper proposes a fault detection mechanism that combines multi-step time series prediction with dynamic fault detection threshold comparison. This mechanism can adaptively adjust the deviation detection standard according to the different operating loads and environmental conditions of the building. By constructing a prediction deviation measurement sequence and linking it with a dynamic threshold, the ability to identify potential fault trends early is greatly improved. This avoids the false alarm and missed alarm problems caused by environmental disturbances in traditional fixed threshold methods, and achieves a more sensitive and robust fault detection function for intelligent building systems.

[0030] (3) This invention designs a temporal stability verification and spatial consistency analysis process for fault confirmation. By performing stability statistics on consecutive abnormal time periods and performing density clustering and spatial connectivity tests on spatially related nodes, it can effectively eliminate occasional noise and isolated anomalies and accurately identify the actual source of the fault. Compared with existing methods based solely on single-point data judgment, this invention significantly improves the reliability, accuracy, and spatial resolution of fault judgment, supporting more accurate local fault location and early warning for building systems.

[0031] (4) In the process of fuzzy rule generation and evolutionary update, the present invention uses a Gaussian membership function based on implicit state feature extraction to construct the initial fuzzy rule set, and combines a real-time error sensitivity two-stage update mechanism to synchronously adjust the rule parameters and neural weights. This solution not only improves the model's ability to model the complex dynamic characteristics of building operations, but also achieves continuous self-optimization of the fault prediction model through sliding window dynamic monitoring and adaptive evolutionary control, effectively balancing the relationship between prediction accuracy, model complexity, and computational overhead, ensuring that the method can operate efficiently and stably in actual building operation and maintenance scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 This is a schematic diagram of the overall process of an intelligent building fault prediction method based on an adaptive algorithm proposed by the present invention; Figure 2 This is a schematic diagram of the structural composition of an improved adaptive prediction model of an intelligent building fault prediction method based on an adaptive algorithm proposed in the present invention; Figure 3 This is a fault confirmation flow chart based on temporal stability verification and spatial consistency analysis of an intelligent building fault prediction method based on an adaptive algorithm proposed by the present invention. DETAILED DESCRIPTION

[0033] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0034] refer to Figure 1-Figure 3 , a method for predicting intelligent building faults based on an adaptive algorithm, comprising the following steps: S1. Obtain the operating status data collected by multiple sensor nodes in the target intelligent building, and perform normalization preprocessing on the operating status data to form a standardized data set; During the specific implementation process, real-time operating status data is first obtained from multiple sensor nodes distributed in the target intelligent building. The sensor nodes include but are not limited to various types of collection equipment installed in air-conditioning systems, elevator systems, power distribution systems, security systems, fire protection systems, and environmental monitoring systems. The collected operating status data covers a variety of parameters such as temperature, humidity, wind speed, carbon dioxide concentration, voltage, current, power, vibration amplitude, and equipment start-stop status. Each sensor node collects data according to a set sampling frequency, preferably set to sample once per minute. The collected data is aggregated to the data processing center in real time via a local area network, Internet of Things communication protocol, or other wireless data transmission methods. In order to ensure the consistency and stability of the model input data, the operating status data of each dimension is normalized and preprocessed. Preferably, the range normalization method is used to map the raw data of each dimension to an interval based on its historical maximum and minimum values. At the same time, the sliding window statistical method is combined with the normalization process to detect and eliminate outliers, shield or smooth out outliers that exceed three times the standard deviation of the mean, and use adjacent time point interpolation to repair small-scale missing data caused by signal loss. The operating status data after normalization and outlier cleaning are arranged in chronological order to form a standardized data set, which serves as the basic input data source for feature encoding, implicit state extraction, fuzzy rule generation and fault prediction reasoning processes, ensuring the consistency, temporality and numerical stability of the overall system operating status modeling.

[0035] S2. Inputting the standardized data set into the improved adaptive prediction model to obtain a latent state representation; In this embodiment, S2 specifically includes: S21. Organize the standardized data set in chronological order to construct a time series input set consisting of multiple time series samples, wherein the time series samples include building operation status data of multiple sensor nodes in the standardized data set at various time points; S22, inputting each time series sample in the time series input set into the improved adaptive prediction model in sequence, passing the building operation status data of each sensor node into the fuzzification layer of the improved adaptive prediction model through the input layer structure of the improved adaptive prediction model, and performing fuzzy membership calculation on the building operation status data of each dimension to generate a corresponding fuzzy input representation; S23, performing a fuzzy rule matching operation based on the fuzzy input representation, calling a fuzzy rule set in the improved adaptive prediction model, calculating a matching degree between corresponding time series samples in the time series input set and the fuzzy rule set, and obtaining a fuzzy rule activation value set; S24. Passing the fuzzy rule activation value set as an input signal to the rule inference layer in the improved adaptive prediction model, and generating an implicit state representation by fusing the fuzzy rule activation value with the corresponding neural connection weight.

[0036] The improved adaptive prediction model specifically includes: Constructing an initial fuzzy neural network structure, the structure comprising an input layer, a fuzzification layer, a fuzzy rule layer, a rule inference layer, and an output layer, wherein the fuzzy rule layer is composed of a plurality of initial fuzzy rules, each of which has an initial membership function parameter and an initial neural connection weight; Define the prediction error feedback indicator, calculate the real-time prediction error sequence, and calculate the real-time error sensitivity based on the real-time prediction error sequence : ; in, Indicates that at the current moment Next Fuzzy rules and the corresponding neural connection weights Real-time error sensitivity to forecast errors, Indicates that at the current moment The real-time prediction error is Indicates the current time Next The initial membership function parameters of the fuzzy rules, Indicates the current time Next The neural connection weights corresponding to the fuzzy rules; Specifically, the prediction error refers to the difference between the model's output value at the current moment and its actual operating state. The fuzzy rule parameters are the initial membership function parameters corresponding to the fuzzy rule at the current moment, and the neural connection weights are the connection weights corresponding to the fuzzy rule at the current moment. By calculating this real-time error sensitivity, the fuzzy rule parameters and neural connection weights can be dynamically adjusted during the optimization process based on the sensitivity, achieving continuous optimization and adaptive evolution of the model's prediction performance.

[0037] Based on the real-time error sensitivity, a unified evolutionary criterion is established for the fuzzy rule set and the neural connection weights. The real-time prediction error at the current moment is taken as the only optimization target, and the fuzzy rule set and the neural connection weights are simultaneously included in the synchronous iterative update range. A two-phase alternating update process is performed in real time: Rule evolution stage: based on real-time error sensitivity , adjust the membership function parameters of the fuzzy rule structure of the fuzzy rule layer: ; in, Indicates the first Fuzzy rules at the current moment The updated membership function parameters of Indicates the Fuzzy rules at the current moment Membership function parameters before updating, The step size coefficient for adjusting the fuzzy rule parameters; Specifically, during the rule evolution phase, for each fuzzy rule's current membership function parameters, the sensitivity calculation results of the real-time prediction error are first used to determine the current impact of the fuzzy rule on the overall prediction error change. Based on this, the fuzzy rule membership function parameters at the previous moment are used as the baseline value and multiplied by the real-time error sensitivity value corresponding to the fuzzy rule according to a certain adjustment step coefficient. The resulting product is used to correct the fuzzy rule parameters at the previous moment to generate the adjusted membership function parameters. The adjustment step coefficient is a pre-set positive real number used to control the amplitude of the fuzzy rule parameter change and ensure the smoothness and convergence of the parameter adjustment process. Through this process, the membership function parameters of each fuzzy rule are dynamically updated based on its actual current contribution to the prediction error, allowing the fuzzy rule set as a whole to adapt more quickly to changes in the building system's operating status, thereby improving the accuracy and adaptability of the fault prediction model.

[0038] Weight evolution stage: Based on the membership function parameters of the updated fuzzy rule structure, the neural connection weights are further dynamically adjusted: ; in, Indicates the first The neural connection weights corresponding to the fuzzy rules are at time The updated value, Indicates the The neural connection weights corresponding to the fuzzy rules are at time The value before updating, The step size coefficient for adjusting the neural connection weights; Specifically, for each fuzzy rule, the neural connection weight of the fuzzy rule at the current moment is used as the baseline value, and the adjustment amount is calculated based on the updated fuzzy rule parameters and the current prediction error sensitivity. The adjustment amount is determined by the product of the neural connection weight adjustment step coefficient and the real-time error sensitivity. The above adjustment amount is deducted from the original neural connection weight value to obtain the updated neural connection weight value. The adjustment step coefficient is a preset value, which controls the amplitude of each weight change and ensures the convergence stability and adjustment sensitivity during the connection weight update process. Through the above method, the neural connection weight can be optimized and adjusted in real time according to the system operation status, further reducing the prediction error between the model output and the actual state, and improving the adaptability and prediction accuracy of the overall model in complex building operation environments.

[0039] By repeatedly executing the above two-stage alternating updating process in real time until the real-time prediction error converges to a predetermined threshold range, an improved adaptive prediction model after the two-stage alternating updating and adjustment is obtained.

[0040] S3, performing fuzzy rule induction on the implicit state representation, constructing an initial fuzzy rule set based on the fuzzy membership function, and generating a first-stage prediction result; In this embodiment, S3 specifically includes: S31, arranging the implicit state representations into an implicit state sequence in chronological order, and determining the maximum and minimum value ranges of the implicit state values of each dimension; S32. Divide the implicit state value of each dimension into a preset number of clusters using the K-means clustering algorithm within the maximum and minimum value intervals, using the sum of squared errors within the clusters reaching a local minimum as the division convergence criterion, determine the center value and the width value of each cluster, and set the center value and the width value of each cluster as the center parameter and the width parameter of the Gaussian membership function, respectively, to construct a Gaussian membership function family for the implicit state value of each dimension; S33, using the Gaussian membership function family, calculating the membership of the implicit state value of each dimension in the implicit state sequence, performing fuzzy label mapping on the membership of all dimensions at the same time point, and obtaining the fuzzy label combination corresponding to the time point; S34. Based on the fuzzy label combination at each time point, the fuzzy labels in all fuzzy label combinations are arranged and combined, and an initial fuzzy rule set is constructed according to the "if-and-then" semantic format, wherein the antecedent of each initial fuzzy rule in the initial fuzzy rule set is formed by the fuzzy label combination, and the consequent is the output label to be predicted; S35. For each initial fuzzy rule in the initial fuzzy rule set, use the product T-norm as a rule activation calculation operator to calculate the rule activation and generate a rule activation set; S36: Perform weighted normalization on the rule activation set to calculate the first-stage prediction result: ; in, For time point The first stage prediction results, For the The activation degree of the initial fuzzy rules, For the The output value corresponding to the consequent “output label to be predicted” of the initial fuzzy rule, is the total number of initial fuzzy rules.

[0041] Specifically, the activation of each initial fuzzy rule at the current time point is multiplied by the predicted output label value corresponding to the rule's consequent to obtain each rule's weighted contribution to the final predicted output. Subsequently, the weighted contribution values of all rules are accumulated to obtain a weighted sum. The activation values of all rules are also accumulated to form a normalized denominator. Finally, the weighted sum is divided by the accumulated activation value to obtain the first-stage prediction output at the current time point. The first-stage prediction output represents a preliminary prediction of the system's future state change trend at the current moment. The activation value reflects the degree of match between each fuzzy rule and the input implicit state at the current moment. The predicted output label value is the specific predicted value associated with the consequent of each initial fuzzy rule. The total number of initial fuzzy rules is the number of rules involved in the inference. This weighted normalized inference mechanism fully integrates the responsiveness of different rules to the current state, effectively improving the continuity, smoothness, and accuracy of the prediction output.

[0042] S4. Introducing a real-time evolutionary strategy to dynamically adjust the neural connection weights of the initial fuzzy rule set and the improved adaptive prediction model based on the error feedback of the prediction results of the first stage, thereby generating an evolved and updated adaptive prediction model; In this embodiment, the S4 specifically includes: S41. Obtain the first-stage prediction result and the corresponding actual operating status value at each time point, calculate the real-time prediction error between the first-stage prediction result and the actual operating status value, and form a real-time prediction error sequence in chronological order; S42, introducing a real-time evolution strategy, calculating a real-time error sensitivity according to the real-time prediction error sequence, and performing an error sensitivity evaluation on the membership function parameters of each initial fuzzy rule in the initial fuzzy rule set; S43, in the rule evolution stage of the real-time evolution strategy, adaptively adjusting the parameters of each membership function according to the step coefficient of the fuzzy rule parameter adjustment to obtain an updated fuzzy rule set; S44. In the weight evolution phase of the real-time evolution strategy, adaptively adjust the neural connection weights of the improved adaptive prediction model according to the step coefficient of the neural connection weight adjustment and in combination with the real-time error sensitivity to obtain an updated neural connection weight set; S45, synchronously writing the updated fuzzy rule set and the updated neural connection weight set into the improved adaptive prediction model to form an evolutionary update iterative round, and repeatedly executing S42 to S44 until the real-time prediction error converges to a set threshold; S46. When the convergence criterion of the real-time evolution strategy is met, the adaptive prediction model updated through evolution is output.

[0043] The real-time evolution strategy specifically includes: Setting a single evolution trigger threshold and a single convergence threshold, and establishing an evolution parameter registration table for the fuzzy rule set and the neural connection weight, wherein the evolution parameter registration table records the current values and adjustable intervals of the step coefficients for adjusting the fuzzy rule parameters and the neural connection weight; The window mean of the real-time prediction error sequence is continuously counted using a fixed-length sliding time window. When the window mean is greater than the single evolution trigger threshold, an evolution cycle is initiated. In the rule evolution stage, the membership function parameters are adaptively adjusted according to the real-time error sensitivity. The adjustment range is controlled by the step coefficient of the fuzzy rule parameter adjustment in the evolution parameter registration table. After the adjustment is completed, the updated fuzzy rule set is immediately written back. In the weight evolution stage, the neural connection weights are adaptively adjusted according to the real-time error sensitivity. The adjustment range is controlled by the step coefficient of the neural connection weight adjustment in the evolution parameter registration table. After the adjustment is completed, the updated neural connection weight set is immediately written back. After completing the rule evolution stage and the weight evolution stage, the window mean is recalculated and compared with the single convergence threshold. When the window mean is not higher than the single convergence threshold or the evolution cycle has reached the maximum number of allowed iterations, the evolution is terminated. Otherwise, return to step S63 to continue executing the next round of evolution cycle.

[0044] S5. Using the adaptive prediction model that has been updated through evolution, predict the time series consisting of key operating parameters of the target intelligent building and output multi-step prediction results for multiple time points in the future; In this embodiment, the S5 specifically includes: S51. Using the adaptive prediction model that has been updated through evolution, obtain a time series consisting of key operating parameters of the target intelligent building, and set a prediction start point and a prediction end point according to the time span and data density of the time series; S52. Based on the current moment, a time series consisting of key operating parameters within a fixed-length time window in the past is input into the adaptive prediction model that has been updated through evolution, and a prediction result for the first prediction moment is obtained through reasoning based on the fuzzy rules and neural network structure within the adaptive prediction model that has been updated through evolution; S53. At the next prediction moment, append the prediction result output at the previous prediction moment to the end of the input key operating parameter time series to form an updated key operating parameter time series, input the updated key operating parameter time series into the adaptive prediction model that has been updated through evolution, and generate a prediction result for the next prediction moment; S54, repeatedly executing step S53, gradually predicting the key operating parameters at multiple consecutive future prediction moments, and forming a multi-step prediction result sequence for multiple future time points; S55. Perform linear interpolation processing on the generated multi-step prediction result sequence to unify the time interval and data dimension of the multi-step prediction result sequence and generate a normalized multi-step prediction result.

[0045] S6. construct a prediction deviation metric sequence based on the multi-step prediction results, compare the prediction deviation metric sequence with a preset fault threshold, and generate a candidate fault indication; In this embodiment, S6 specifically includes: S61, pairing the normalized multi-step prediction result sequence with the actual operating status data of the target intelligent building at the corresponding time point, calculating the prediction residual point by point in time axis order to form a prediction residual sequence, where the prediction residual is the difference between the multi-step prediction result and the actual operating status data; S62. Based on the prediction residual sequence, a prediction deviation measurement sequence is constructed using a dual statistical method of sliding mean and sliding standard deviation, wherein the prediction deviation measurement value corresponding to each time point is obtained by weighted fusion of the residual mean and residual standard deviation in a fixed window before and after the time point; S63. Based on the prediction deviation measurement sequence, a unique dynamic fault detection threshold is set. The dynamic fault detection threshold is generated based on the residual mean and standard deviation statistics during the historical normal operation of the intelligent building, and is adaptively adjusted synchronously with the system operation load level. S64. Compare the predicted deviation metric value at each time point with the dynamic fault detection threshold at the corresponding moment. If the predicted deviation metric values at multiple consecutive time points exceed the corresponding dynamic fault detection threshold, extract the center position within the continuous exceeding limit interval as a candidate fault indication. The candidate fault indication includes the time point of fault occurrence, duration, maximum deviation amplitude and maximum deviation moment.

[0046] S7. Perform temporal stability verification and spatial consistency analysis on the candidate fault indications. If predefined consistency criteria are met, output the location of the fault to be warned and the fault category in the building system.

[0047] In this embodiment, the S7 specifically includes: S71. Based on the fault occurrence time, duration, maximum deviation amplitude, and maximum deviation moment recorded in the candidate fault indication, a time continuity verification model is constructed within a fixed timing window. The time interval distribution of consecutive candidate fault indications is statistically analyzed. If the consecutive fault interval is less than a set continuity verification threshold and the duration is greater than a minimum stable duration, it is determined to be a timing stable fault event. S72. Under each temporally stable fault event, based on the spatial layout relationship in the intelligent building, call the operating status data of all sensor nodes that overlap with the fault occurrence time period and construct the state similarity between the nodes. The state similarity is calculated based on the synchronous change trend and change amplitude of the node data. S73. Based on the state similarity, perform spatial consistency cluster analysis to extract a node subset with significant spatial correlation. If the number of nodes in the node subset exceeds a set spatial consistency criterion and the nodes cover a continuous area of the building's physical space, then confirm that the spatial consistency condition is met. S74. Under the premise of simultaneously satisfying the temporal stability verification and spatial consistency analysis criteria, extract the node with the highest abnormality in the node subset as the fault center node, and combine the fault center node category and the monitored operating parameter type to calibrate the location and fault category information of the fault to be warned in the building system; S75. Output the location and fault category information of the fault to be warned that has been calibrated.

[0048] The spatial consistency analysis specifically includes: Normalizing the state similarity and mapping the state similarities between all nodes to a unified numerical range; Based on the normalized state similarity between nodes, a density peak clustering algorithm is used to automatically identify areas with abnormally high local state density and extract the node subset corresponding to each density peak. In each node subset, the spatial physical distance matrix between nodes is calculated, and node pairs with continuous physical distances and state similarity higher than the set spatial correlation threshold are screened out to construct spatially continuous sub-regions; In each spatial continuous sub-region, the node abnormality distribution is statistically analyzed, and the spatial sub-regions whose average abnormality is greater than the normal distribution threshold are selected as the fault-related node set; If the number of nodes in the final fault-associated node set reaches the spatial consistency criterion and the physical distribution of the nodes constitutes a connected area inside the building, the spatial consistency condition is determined to be met.

[0049] Example 1: In order to verify the feasibility and superiority of the intelligent building fault prediction method based on the adaptive algorithm proposed in the present invention, the present invention is applied to the intelligent operation and maintenance system of a large-scale comprehensive commercial building, Center B, in South China. The building area of Center B exceeds 120,000 square meters, covering shopping malls, office buildings, hotels and large parking lots. More than 4,300 sensor nodes are installed to monitor equipment operation, energy consumption, security systems, environmental control and other aspects. Due to the huge size of the building and the complexity of the system, Center B has long faced problems in daily operation and maintenance, such as difficulty in early warning of equipment failures, delayed response to sudden failures, and high cost of manual inspections. Especially in energy consumption management and air-conditioning unit operation monitoring, there are frequent cases of delayed fault detection, high false alarm rate and difficulty in positioning, which directly affect tenant experience and operational safety.

[0050] To address the aforementioned issues, Center B decided to introduce the method of the present invention and integrate it into its intelligent operation and maintenance platform. The application deployment process begins with data acquisition, connecting the operating parameters of all air-conditioning units, power distribution cabinets, elevator systems, and environmental monitoring nodes to the data processing center, with a real-time acquisition frequency set to once per minute. After standardized preprocessing, the key operating parameters of each major subsystem (such as supply air temperature, current load, start-stop frequency, etc.) are feature-encoded based on the improved adaptive prediction model proposed in this invention to form an implicit state representation. Subsequently, by integrating fuzzy neural networks with real-time evolution strategies, the initial fuzzy rule set and neural connection weights are dynamically evolved, continuously improving the adaptability and accuracy of the prediction model under actual operating conditions.

[0051] During the fault prediction phase, the system performs multi-step predictions on the operating status of each subsystem in real time based on an adaptive prediction model that is updated through evolution. The prediction duration covers the state change trend within the next 90 minutes. The prediction output uses a rolling window update mechanism to ensure that the system generates a new chain of prediction results every minute. The multi-step prediction results are dynamically compared with the actual observation data. The system constructs a prediction deviation measurement sequence within the sliding window and introduces a dynamic fault detection threshold system to achieve keen capture of deviation anomaly trends. When an abnormal deviation is detected within a continuous time period, the system automatically generates a candidate fault indication and further performs time series stability verification and spatial consistency analysis. Ultimately, the specific equipment location and corresponding fault category of the fault to be warned are accurately output, greatly improving the speed of fault location and warning response.

[0052] During this implementation, the system was fully validated using Center B's data from four consecutive months, from September to December 2024, as a test period. The following are partial statistics and performance comparisons from the test period. The data is sourced from Center B's intelligent operation and maintenance system logs and fault event records.

[0053] Table 1: Application effect data of intelligent building fault prediction system in Center B ; The data in Table 1 shows that after implementing the method of this invention, Center B's fault prediction accuracy was significantly improved, and the missed alarm rate was significantly reduced. The system was able to issue early warnings long before faults occurred, significantly extending the operation and maintenance response window. In particular, in detecting energy consumption anomalies in the air conditioning system, the system successfully captured 12 early energy leakage events by continuously monitoring the prediction deviation metric sequence, thus preventing equipment overload damage caused by abnormal energy consumption.

[0054] For example, at 3:45 AM on November 23, 2024, the supply air temperature of unit 1 of unit B2 in the west section of the B Center experienced abnormal fluctuations. While traditional threshold monitoring systems failed to detect the problem immediately, the proposed method, through sliding standard deviation analysis of multi-step prediction results, issued a fault warning six hours and 45 minutes in advance. On-site engineers subsequently confirmed that insufficient cooling capacity was caused by valve mechanism wear. Emergency repairs were ultimately completed before the actual failure occurred, avoiding over 300,000 yuan in direct repair and downtime losses.

[0055] Another typical case involved an elevator motor on the sixth floor of the East Office Building in Center B, which experienced abnormal vibration on October 14, 2024. The proposed method, based on spatial consistency cluster analysis, comprehensively analyzed the operating status of the elevators on the sixth floor and the adjacent elevators on the fifth and seventh floors, identifying this elevator as the central node of the fault. The system issued an early warning at the onset of the vibration anomaly, pinpointing the fault location 48 hours earlier than traditional manual inspections, significantly improving response speed and accuracy.

[0056] Table 2: Comparison of detection and processing time for typical cases ; The aforementioned application examples and data comparisons fully demonstrate that the adaptive algorithm-based intelligent building fault prediction method proposed in this invention can effectively improve fault prediction accuracy, extend warning time, and reduce fault response costs in complex building operation and maintenance environments, demonstrating its high practical value and potential for widespread application. In particular, in core technical areas such as multi-step prediction, dynamic deviation detection, and spatial consistency analysis, this invention overcomes the problems of insufficient fault prediction accuracy, delayed response, and frequent false alarms and missed alarms in existing intelligent building systems. This method provides reliable and efficient technical support for intelligent building operation and maintenance, aligning with the development trend of intelligent, automated, and preventive maintenance.

[0057] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for predicting intelligent building faults based on an adaptive algorithm, characterized in that: The steps include: S1. Obtain the operating status data collected by multiple sensor nodes in the target intelligent building, and perform normalization preprocessing on the operating status data to form a standardized data set; S2. Inputting the standardized data set into the improved adaptive prediction model to obtain a latent state representation; S3, performing fuzzy rule induction on the implicit state representation, constructing an initial fuzzy rule set based on the fuzzy membership function, and generating a first-stage prediction result; S4. Introducing a real-time evolutionary strategy to dynamically adjust the neural connection weights of the initial fuzzy rule set and the improved adaptive prediction model based on the error feedback of the prediction results of the first stage, thereby generating an evolved and updated adaptive prediction model; S5. Using the adaptive prediction model that has been updated through evolution, predict the time series consisting of key operating parameters of the target intelligent building and output multi-step prediction results for multiple time points in the future; S6. construct a prediction deviation metric sequence based on the multi-step prediction results, compare the prediction deviation metric sequence with a preset fault threshold, and generate a candidate fault indication; S7. Perform temporal stability verification and spatial consistency analysis on the candidate fault indications. If predefined consistency criteria are met, output the location of the fault to be warned and the fault category in the building system.

2. The intelligent building fault prediction method based on adaptive algorithm according to claim 1 is characterized in that: The S2 specifically includes: S21. Organize the standardized data set in chronological order to construct a time series input set consisting of multiple time series samples, wherein the time series samples include building operation status data of multiple sensor nodes in the standardized data set at various time points; S22, inputting each time series sample in the time series input set into the improved adaptive prediction model in sequence, passing the building operation status data of each sensor node into the fuzzification layer of the improved adaptive prediction model through the input layer structure of the improved adaptive prediction model, and performing fuzzy membership calculation on the building operation status data of each dimension to generate a corresponding fuzzy input representation; S23, performing a fuzzy rule matching operation based on the fuzzy input representation, calling a fuzzy rule set in the improved adaptive prediction model, calculating a matching degree between corresponding time series samples in the time series input set and the fuzzy rule set, and obtaining a fuzzy rule activation value set; S24. Passing the fuzzy rule activation value set as an input signal to the rule inference layer in the improved adaptive prediction model, and generating an implicit state representation by fusing the fuzzy rule activation value with the corresponding neural connection weight.

3. The intelligent building fault prediction method based on adaptive algorithm according to claim 1 is characterized in that: The improved adaptive prediction model specifically includes: Constructing an initial fuzzy neural network structure, the structure comprising an input layer, a fuzzification layer, a fuzzy rule layer, a rule inference layer, and an output layer, wherein the fuzzy rule layer is composed of a plurality of initial fuzzy rules, each of which has an initial membership function parameter and an initial neural connection weight; Define the prediction error feedback indicator, calculate the real-time prediction error sequence, and calculate the real-time error sensitivity based on the real-time prediction error sequence : ； in, Indicates that at the current moment Next Fuzzy rules and the corresponding neural connection weights Real-time error sensitivity to forecast errors, Indicates that at the current moment The real-time prediction error is Indicates the current time Next The initial membership function parameters of the fuzzy rules, Indicates the current time Next The neural connection weights corresponding to the fuzzy rules; Based on the real-time error sensitivity, a unified evolutionary criterion is established for the fuzzy rule set and the neural connection weights. The real-time prediction error at the current moment is taken as the only optimization target, and the fuzzy rule set and the neural connection weights are simultaneously included in the synchronous iterative update range. A two-phase alternating update process is performed in real time: Rule evolution stage: based on real-time error sensitivity , adjust the membership function parameters of the fuzzy rule structure of the fuzzy rule layer: ； in, Indicates the first Fuzzy rules at the current moment The updated membership function parameters of Indicates the Fuzzy rules at the current moment Membership function parameters before updating, The step size coefficient for adjusting the fuzzy rule parameters; Weight evolution stage: Based on the membership function parameters of the updated fuzzy rule structure, the neural connection weights are further dynamically adjusted: ； in, Indicates the first The neural connection weights corresponding to the fuzzy rules are at time The updated value, Indicates the The neural connection weights corresponding to the fuzzy rules are at time The value before updating, The step size coefficient for adjusting the neural connection weights; By repeatedly executing the above two-stage alternating updating process in real time until the real-time prediction error converges to a predetermined threshold range, an improved adaptive prediction model after the two-stage alternating updating and adjustment is obtained.

4. The intelligent building fault prediction method based on adaptive algorithm according to claim 1 is characterized in that: The S3 specifically includes: S31, arranging the implicit state representations into an implicit state sequence in chronological order, and determining the maximum and minimum value ranges of the implicit state values of each dimension; S32. Divide the implicit state value of each dimension into a preset number of clusters using the K-means clustering algorithm within the maximum and minimum value intervals, using the sum of squared errors within the clusters reaching a local minimum as the division convergence criterion, determine the center value and the width value of each cluster, and set the center value and the width value of each cluster as the center parameter and the width parameter of the Gaussian membership function, respectively, to construct a Gaussian membership function family for the implicit state value of each dimension; S33, using the Gaussian membership function family, calculating the membership of the implicit state value of each dimension in the implicit state sequence, performing fuzzy label mapping on the membership of all dimensions at the same time point, and obtaining the fuzzy label combination corresponding to the time point; S34. Based on the fuzzy label combination at each time point, the fuzzy labels in all fuzzy label combinations are arranged and combined, and an initial fuzzy rule set is constructed according to the "if-and-then" semantic format, wherein the antecedent of each initial fuzzy rule in the initial fuzzy rule set is formed by the fuzzy label combination, and the consequent is the output label to be predicted; S35. For each initial fuzzy rule in the initial fuzzy rule set, use the product T-norm as a rule activation calculation operator to calculate the rule activation and generate a rule activation set; S36: Perform weighted normalization on the rule activation set to calculate the first-stage prediction result: ； in, For time point The first stage prediction results, For the The activation degree of the initial fuzzy rules, For the The output value corresponding to the consequent "output label to be predicted" of the initial fuzzy rule, is the total number of initial fuzzy rules.

5. The intelligent building fault prediction method based on adaptive algorithm according to claim 1 is characterized in that: The S4 specifically includes: S41. Obtain the first-stage prediction result and the corresponding actual operating status value at each time point, calculate the real-time prediction error between the first-stage prediction result and the actual operating status value, and form a real-time prediction error sequence in chronological order; S42, introducing a real-time evolution strategy, calculating a real-time error sensitivity according to the real-time prediction error sequence, and performing an error sensitivity evaluation on the membership function parameters of each initial fuzzy rule in the initial fuzzy rule set; S43, in the rule evolution stage of the real-time evolution strategy, adaptively adjusting the parameters of each membership function according to the step coefficient of the fuzzy rule parameter adjustment to obtain an updated fuzzy rule set; S44. In the weight evolution phase of the real-time evolution strategy, adaptively adjust the neural connection weights of the improved adaptive prediction model according to the step coefficient of the neural connection weight adjustment and in combination with the real-time error sensitivity to obtain an updated neural connection weight set; S45, synchronously writing the updated fuzzy rule set and the updated neural connection weight set into the improved adaptive prediction model to form an evolutionary update iterative round, and repeatedly executing S42 to S44 until the real-time prediction error converges to a set threshold; S46. When the convergence criterion of the real-time evolution strategy is met, the adaptive prediction model updated through evolution is output.

6. The intelligent building fault prediction method based on adaptive algorithm according to claim 5, characterized in that: The real-time evolution strategy specifically includes: Setting a single evolution trigger threshold and a single convergence threshold, and establishing an evolution parameter registration table for the fuzzy rule set and the neural connection weight, wherein the evolution parameter registration table records the current values and adjustable intervals of the step coefficients for adjusting the fuzzy rule parameters and the neural connection weight; The window mean of the real-time prediction error sequence is continuously counted using a fixed-length sliding time window. When the window mean is greater than the single evolution trigger threshold, an evolution cycle is initiated. In the rule evolution stage, the membership function parameters are adaptively adjusted according to the real-time error sensitivity. The adjustment range is controlled by the step coefficient of the fuzzy rule parameter adjustment in the evolution parameter registration table. After the adjustment is completed, the updated fuzzy rule set is immediately written back. In the weight evolution stage, the neural connection weights are adaptively adjusted according to the real-time error sensitivity. The adjustment range is controlled by the step coefficient of the neural connection weight adjustment in the evolution parameter registration table. After the adjustment is completed, the updated neural connection weight set is immediately written back. After completing the rule evolution stage and the weight evolution stage, the window mean is recalculated and compared with the single convergence threshold. When the window mean is not higher than the single convergence threshold or the evolution cycle has reached the maximum number of allowed iterations, the evolution is terminated. Otherwise, return to step S63 to continue executing the next round of evolution cycle.

7. The intelligent building fault prediction method based on adaptive algorithm according to claim 1 is characterized in that: The S5 specifically includes: S51. Using the adaptive prediction model that has been updated through evolution, obtain a time series consisting of key operating parameters of the target intelligent building, and set a prediction start point and a prediction end point according to the time span and data density of the time series; S52. Based on the current moment, a time series consisting of key operating parameters within a fixed-length time window in the past is input into the adaptive prediction model that has been updated through evolution, and a prediction result for the first prediction moment is obtained through reasoning based on the fuzzy rules and neural network structure within the adaptive prediction model that has been updated through evolution; S53. At the next prediction moment, append the prediction result output at the previous prediction moment to the end of the input key operating parameter time series to form an updated key operating parameter time series, input the updated key operating parameter time series into the adaptive prediction model that has been updated through evolution, and generate a prediction result for the next prediction moment; S54, repeatedly executing step S53, gradually predicting the key operating parameters at multiple consecutive future prediction moments, and forming a multi-step prediction result sequence for multiple future time points; S55. Perform linear interpolation processing on the generated multi-step prediction result sequence to unify the time interval and data dimension of the multi-step prediction result sequence to generate a normalized multi-step prediction result.

8. The intelligent building fault prediction method based on adaptive algorithm according to claim 1, characterized in that: The S6 specifically includes: S61, pairing the normalized multi-step prediction result sequence with the actual operating status data of the target intelligent building at the corresponding time point, calculating the prediction residual point by point in time axis order to form a prediction residual sequence, where the prediction residual is the difference between the multi-step prediction result and the actual operating status data; S62. Based on the prediction residual sequence, a prediction deviation measurement sequence is constructed using a dual statistical method of sliding mean and sliding standard deviation, wherein the prediction deviation measurement value corresponding to each time point is obtained by weighted fusion of the residual mean and residual standard deviation in a fixed window before and after the time point; S63. Based on the prediction deviation measurement sequence, a unique dynamic fault detection threshold is set. The dynamic fault detection threshold is generated based on the residual mean and standard deviation statistics during the historical normal operation of the intelligent building, and is adaptively adjusted synchronously with the system operation load level. S64. Compare the predicted deviation metric value at each time point with the dynamic fault detection threshold at the corresponding moment. If the predicted deviation metric values at multiple consecutive time points exceed the corresponding dynamic fault detection threshold, extract the center position within the continuous exceeding limit interval as a candidate fault indication. The candidate fault indication includes the time point of fault occurrence, duration, maximum deviation amplitude and maximum deviation moment.

9. The intelligent building fault prediction method based on adaptive algorithm according to claim 1, characterized in that: The S7 specifically includes: S71. Based on the fault occurrence time, duration, maximum deviation amplitude, and maximum deviation moment recorded in the candidate fault indication, a time continuity verification model is constructed within a fixed timing window. The time interval distribution of consecutive candidate fault indications is statistically analyzed. If the consecutive fault interval is less than a set continuity verification threshold and the duration is greater than a minimum stable duration, it is determined to be a timing stable fault event. S72. Under each temporally stable fault event, based on the spatial layout relationship in the intelligent building, call the operating status data of all sensor nodes that overlap with the fault occurrence time period and construct the state similarity between the nodes. The state similarity is calculated based on the synchronous change trend and change amplitude of the node data. S73. Based on the state similarity, perform spatial consistency cluster analysis to extract a node subset with significant spatial correlation. If the number of nodes in the node subset exceeds a set spatial consistency criterion and the nodes cover a continuous area of the building's physical space, then confirm that the spatial consistency condition is met. S74. Under the premise of simultaneously satisfying the temporal stability verification and spatial consistency analysis criteria, extract the node with the highest abnormality in the node subset as the fault center node, and combine the fault center node category and the monitored operating parameter type to calibrate the location and fault category information of the fault to be warned in the building system; S75. Output the location and fault category information of the fault to be warned that has been calibrated.

10. The intelligent building fault prediction method based on adaptive algorithm according to claim 9, characterized in that: The spatial consistency analysis specifically includes: Normalizing the state similarity and mapping the state similarities between all nodes to a unified numerical range; Based on the normalized state similarity between nodes, a density peak clustering algorithm is used to automatically identify areas with abnormally high local state density and extract the node subset corresponding to each density peak. In each node subset, the spatial physical distance matrix between nodes is calculated, and node pairs with continuous physical distances and state similarity higher than the set spatial correlation threshold are screened out to construct spatially continuous sub-regions; In each spatial continuous sub-region, the node abnormality distribution is statistically analyzed, and the spatial sub-regions whose average abnormality is greater than the normal distribution threshold are selected as the fault-related node set; If the number of nodes in the final fault-associated node set reaches the spatial consistency criterion and the physical distribution of the nodes constitutes a connected area inside the building, the spatial consistency condition is determined to be met.

Images