Multi-level fusion building system sensor device data anomaly attribution method and system

Abstract

The application provides a multi-level fusion building system sensor device data anomaly attribution method and system, which comprises the following steps: obtaining current sensor data and preprocessed sensor data sequence of a sensor device in a building system; determining an extreme abnormal value according to whether the current sensor data exceeds a reasonable range of the preprocessed sensor data sequence; eliminating data exceeding the reasonable range in the preprocessed sensor data sequence to obtain a non-extreme sensor data sequence; calculating deviation of the current sensor data relative to data in the non-extreme sensor data sequence to obtain a trend abnormal value; performing reasoning on the current sensor data through a pre-trained structure anomaly model to obtain a structure abnormal value; obtaining a rule abnormal value based on a running rule model; and comparing each abnormal value with a preset abnormal threshold to obtain a sensor device data anomaly reason. The sensor device data anomaly reason in the application is explanatory, and the sensor device anomaly attribution determination is reliable.

Description

Technical Field

[0001] This invention relates to the fields of intelligent building operation and maintenance and IoT edge computing, specifically to a multi-layered integrated method and system for attributing data anomalies in building system sensing devices. Background Technology

[0002] With the widespread application of green buildings and smart buildings, various environmental and automatic control system sensors (such as temperature, humidity, air volume, and power) provide crucial data support for building control and energy efficiency optimization. Under the dual control policy for energy consumption, building energy consumption monitoring needs to have the ability to trace data and accurately locate faults. However, traditional data acquisition platforms cannot meet the needs of accurate judgment in complex scenarios. The core contradiction leading to this situation is that the root cause of abnormal sensor data is difficult to determine.

[0003] The root causes of sensor data anomalies are difficult to pinpoint, including: sensor performance gradually deteriorates due to aging and data drift over time, reducing data reliability. Traditional methods, relying on independent measurements from single sensors or data sources (such as temperature, humidity, and power at a specific moment), lack correlation analysis with other relevant sensor data and multi-level collaborative judgment mechanisms. This makes it difficult to trace controller erroneous commands or logical anomalies, resulting in a high false alarm rate. If only existing artificial intelligence models are relied upon, their lack of engineering rule interpretation capabilities makes anomaly detection results difficult for maintenance personnel to accept, leading to a situation where there is a model but no response.

[0004] Therefore, there is an urgent need for a comprehensive judgment method that integrates multi-level analysis and engineering mechanisms to achieve reliable determination of the cause of sensor device anomalies. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by providing a multi-layered fusion method and system for attributing data anomalies in building system sensing devices.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] A multi-layered fusion method for attributing data anomalies in building system sensor devices includes the following steps:

[0008] Acquire current sensing data from different sensing devices in the building system;

[0009] Historical data from different sensing devices in a building system are obtained based on time series data to obtain historical sensing data sequences. These historical sensing data sequences are then preprocessed to obtain preprocessed sensing data sequences.

[0010] Extreme outliers are determined based on whether the current sensor data exceeds a reasonable range; wherein, the reasonable range is obtained by dividing the preprocessed sensor data sequence based on quartiles;

[0011] Data exceeding the reasonable range in the preprocessed sensor data sequence is removed to obtain a non-extreme sensor data sequence.

[0012] Calculate the deviation of the current sensor data from the data in the non-extreme sensor data sequence to obtain the trend outlier;

[0013] The current sensor data is inferred by a pre-trained structural anomaly model to obtain the predicted value of the current data; the predicted value of the current data is compared with the current sensor data to obtain the structural anomaly value.

[0014] Based on the current sensor data and combined with the operation rule model, the total value of the rule anomaly weight that meets the anomaly conditions is compared with the total value of the rule anomaly weight that applies the anomaly conditions to obtain the rule anomaly value; wherein, the operation rule model is constructed based on the sensor device anomaly conditions and the rule anomaly weight that meets the anomaly conditions, and the anomaly conditions are set according to the operation rules of the sensor device in the building system.

[0015] Extreme outliers, trend outliers, structural outliers, and regular outliers are compared with preset outlier thresholds to determine the cause of data anomalies in the sensing device.

[0016] As one possible implementation method, determining extreme outliers based on whether the current sensing data exceeds a reasonable range is as follows:

[0017] If the current sensing data of the target sensing device exceeds the reasonable range of the corresponding sensing device data, then the target sensing device has an extreme anomaly, and the extreme anomaly value is a first preset extreme anomaly value; if the current sensing data of the target sensing device is within the reasonable range of the corresponding sensing device data, then the target sensing device does not have an extreme anomaly, and the extreme anomaly value is a second preset extreme anomaly value.

[0018] The reasonable range of the data from the sensing device is defined as follows:

[0019]

[0020] in, Indicates the first... The first quartile of the data from the sensor device Indicates the first... The third quartile of the data from the sensor device Indicates the first... The interquartile range of data from each sensor device , They represent the first The weights of the first and third quartiles of the data from each sensor device.

[0021] As one possible implementation, calculating the deviation of the current sensing data relative to data in a non-extreme sensing data sequence to obtain trend outliers includes the following steps:

[0022] Obtain the most recent data of the first preset duration in the non-extreme sensing data sequence to obtain the first duration sensing data sequence, and then obtain the mean and standard deviation of the first duration sensing data sequence;

[0023] Calculate the skewness of the non-extreme sensing data sequence to obtain the first sequence skewness;

[0024] When the skewness of the first sequence is less than the first preset skewness threshold, the deviation from the baseline value is the first preset deviation threshold.

[0025] When the skewness of the first sequence is greater than or equal to the first preset skewness threshold, the non-extreme sensing data sequence and the current sensing data are transformed to obtain the transformed data sequence and the transformed current sensing data; the skewness of the transformed data sequence is calculated to obtain the second sequence skewness; the deviation from the reference value is obtained by comparing the second sequence skewness with the first preset skewness threshold and the second preset skewness threshold.

[0026] The deviation value of the transformation trend is obtained based on the mean and standard deviation of the current sensing data and the first time-duration sensing data sequence.

[0027] Based on the deviation from the baseline value and the deviation from the transformed trend value, trend anomalies are obtained;

[0028] The deviation value of the transformation trend is represented as follows:

[0029]

[0030] The trend outliers are represented as follows:

[0031]

[0032] in, Indicates the first Deviation value of the changing trend of data from individual sensors Indicates the first Transformation of data from individual sensors to the current sensor data. Indicates the current timestamp. Indicates the first preset duration. Indicates the first The average value of the first time-long sensor data sequence from each sensor device. Indicates the first The standard deviation of the first duration sensor data sequence of a single sensor device Indicates the first Outliers in the trend of individual sensors This indicates that the smaller of the multiple values ​​will be returned. Indicates the first The deviation of the data from the reference value of each sensor device.

[0033] As one possible implementation, the transformation processing of the non-extreme sensing data sequence and the current sensing data to obtain the transformed data sequence and the transformed current sensing data includes the following steps:

[0034] When all data in the non-extreme sensing data sequence are positive, the non-extreme sensing data sequence is subjected to maximum likelihood estimation to obtain the first transformation parameter; based on the first transformation parameter, the non-extreme sensing data sequence and the current sensing data are subjected to the first transformation to obtain the first transformed data sequence and the first transformed current sensing data.

[0035] When there are negative numbers in the non-extreme sensing data sequence, the maximum likelihood estimation process is performed on the non-extreme sensing data sequence to obtain the second transformation parameter; based on the second transformation parameter, the second transformation process is performed on the non-extreme sensing data sequence and the current sensing data to obtain the second transformed data sequence and the second transformed current sensing data.

[0036] The first transformation process is represented as follows:

[0037]

[0038] The second transformation process is represented as follows:

[0039]

[0040] in, Represents the first transformed data sequence or the second transformed data sequence. The first data sequence of the sensor device One data point, Indicating the first non-extreme sensing data sequence The first data sequence of the sensor device One data point, Indicates the first The first transformation parameter of the data from the sensor device Indicates the first The second transformation parameter of the data from the sensor device.

[0041] As one possible implementation, obtaining the deviation from the reference value based on a comparison between the second sequence skewness and a first preset skewness threshold and a second preset skewness threshold includes the following steps:

[0042] When the skewness of the second sequence is less than the first preset skewness threshold, the deviation reference value is the first preset deviation threshold;

[0043] When the skewness of the second sequence is greater than or equal to the first preset skewness threshold and less than the second preset skewness threshold, the deviation benchmark value is the second preset deviation threshold.

[0044] When the second sequence skewness is greater than or equal to the second preset skewness threshold, the deviation benchmark value is the third preset deviation threshold; wherein, the first preset deviation threshold > the second preset deviation threshold > the third preset deviation threshold.

[0045] As one possible implementation, the step of reasoning about the current sensing data using a pre-trained structural anomaly model to obtain a predicted value for the current data, and comparing the predicted value with the current sensing data to obtain a structural anomaly value, includes the following steps:

[0046] By inferring the current sensing data of the target sensing device through a pre-trained structural anomaly model of the target sensing device, the predicted value of the current data of the target sensing device is obtained.

[0047] The current reconstruction error of the target sensor is obtained by squaring the difference between the current sensing data and the current data prediction value of the target sensor.

[0048] Based on the preset reconstruction error threshold and the current reconstruction error of the target sensing device, the structural anomaly value of the target sensing device is obtained;

[0049] The structural anomaly values ​​of other sensing devices were calculated;

[0050] The structural anomalies of the target sensing device are represented as follows:

[0051]

[0052] in, Indicates the first The structural anomaly value of each sensing device is the structural anomaly value of the target sensing device. Indicates the first Current timestamp of each sensor device The reconstruction error is the current reconstruction error of the target sensing device. , Indicates the first The current sensing data of the target sensing device is the data from the current sensing device. Indicates the first Current data prediction values ​​of each sensor device Indicates the first The preset reconstruction error threshold for each sensor device This indicates that the smaller of the multiple values ​​will be returned.

[0053] As one possible implementation, the step of comparing the total anomaly weight of rules that meet the abnormal conditions with the total anomaly weight of rules that apply the abnormal conditions, based on the current sensor data and combined with the running rule model, to obtain the rule anomaly value, includes the following steps:

[0054] If the current sensing data of the target sensing device meets the abnormal conditions, the corresponding rule abnormal weight is added to the target meets the rule abnormal weight, and then the total value of the target meets the rule abnormal weight is obtained.

[0055] If the current sensing data of the target sensing device is applicable to abnormal conditions, the corresponding rule abnormal weight is added to the target applicable rule abnormal weight, thereby obtaining the total value of the target applicable rule abnormal weight.

[0056] The rule anomaly value of the target sensing device is obtained by the ratio of the total weight of the target satisfying rule anomalies to the total weight of the target applying rule anomalies.

[0057] Calculate regular outliers from other sensors;

[0058] The rule-based outliers of the target sensing device are represented as follows:

[0059]

[0060] in, Indicates the first The regular abnormal values ​​of the sensor device Indicates the first The target of each sensing device satisfies the rule of the total weight of anomalies. Indicates the first The total weight of the abnormality applicable rules for the target of each sensing device.

[0061] As one possible implementation, comparing extreme outliers, trend outliers, structural outliers, and regular outliers with a preset outlier threshold to determine the cause of sensor data anomalies includes the following steps:

[0062] The extreme outliers, trend outliers, structural outliers, and regular outliers of the target sensing device are weighted and summed to obtain the total outlier value of the target sensing device.

[0063] When the total abnormal value of the target sensing device exceeds the preset total abnormal value threshold, the target sensing device in the building system is abnormal, and the reason for the abnormality is as follows:

[0064] When the rule anomaly value of the target sensing device is greater than the corresponding first preset rule anomaly threshold, the target sensing device has a physical logic conflict anomaly.

[0065] When the structural anomaly value of the target sensing device is greater than the corresponding first preset structural anomaly threshold and the regular anomaly value is less than the corresponding second preset regular anomaly threshold, the target sensing device has a structural anomaly.

[0066] When the trend anomaly value of the target sensing device is greater than the corresponding first trend anomaly threshold, there is no extreme anomaly, the structural anomaly value is less than the corresponding second preset structural anomaly threshold, and the regular anomaly value is less than the corresponding third preset regular anomaly threshold, then the target sensing device has a trend anomaly.

[0067] When the target sensing device exhibits an extreme anomaly, the cause of the anomaly is an extreme anomaly.

[0068] Among them, the first preset rule abnormal threshold > the second preset rule abnormal threshold > the third preset rule abnormal threshold, and the first preset structure abnormal threshold > the second preset structure abnormal threshold.

[0069] A multi-layered integrated data anomaly attribution system for building system sensors, capable of implementing the method described above, includes a current data module, a historical data module, an extreme anomaly module, a data removal module, a trend anomaly module, a structural anomaly module, a rule anomaly module, and an anomaly attribution module;

[0070] The current data module acquires current sensing data from different sensing devices in the building system;

[0071] The historical data module acquires historical data from different sensing devices in the building system based on time series, obtains historical sensing data sequences, and preprocesses the historical sensing data sequences to obtain preprocessed sensing data sequences.

[0072] The extreme anomaly module determines extreme anomaly values ​​based on whether the current sensing data exceeds a reasonable range; wherein, the reasonable range is obtained by dividing the preprocessed sensing data sequence based on quartiles;

[0073] The data removal module removes data that exceeds the reasonable range from the preprocessed sensor data sequence to obtain a non-extreme sensor data sequence.

[0074] The trend anomaly module calculates the deviation of the current sensing data from the data in the non-extreme sensing data sequence to obtain the trend anomaly value.

[0075] The structural anomaly module infers the current sensing data using a pre-trained structural anomaly model to obtain the predicted value of the current data; it then compares the predicted value of the current data with the current sensing data to obtain the structural anomaly value.

[0076] The rule anomaly module, based on the current sensor data and combined with the operating rule model, compares the total value of rule anomaly weights that meet the anomaly conditions with the total value of rule anomaly weights that apply the anomaly conditions to obtain the rule anomaly value; wherein, the operating rule model is constructed based on the sensor device anomaly conditions and the rule anomaly weights that meet the anomaly conditions, and the anomaly conditions are set according to the operating rules of the sensor devices in the building system.

[0077] The anomaly attribution module compares extreme anomalies, trend anomalies, structural anomalies, and regular anomalies with preset anomaly thresholds to obtain the cause of the anomaly in the sensor data.

[0078] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described above.

[0079] A multi-layered fusion building system sensor data anomaly attribution device includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the method described above.

[0080] This invention, by adopting the above technical solutions, has significant technical effects:

[0081] This invention provides a multi-layered fusion method and system for attributing anomalies in building system sensor data. It employs a multi-layered fusion mechanism that first considers statistics, then structure, and finally physical logic. Based on historical data and current sensor data, it obtains extreme anomalies, trend anomalies, and structural anomalies from the sensor devices. It then executes a set of engineering rules to generate rule-based anomalies for targeted verification. The different types of anomalies are progressively coordinated at each level. Finally, the anomalies are fused and analyzed to form a multi-layered fusion, revealing the cause of the building system sensor anomalies. The extreme anomalies, trend anomalies, structural anomalies, and rule-based anomalies in this multi-layered fusion method possess dynamic synergistic characteristics of progressive enhancement, cross-verification, and result feedback. The resulting causes of sensor anomalies are explanatory, making the attribution reliable and trustworthy for maintenance personnel to take appropriate responses. Attached Figure Description

[0082] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0083] Figure 1 This is a flowchart illustrating an embodiment of the method of the present invention;

[0084] Figure 2 This is an overall schematic diagram of an embodiment of the system of the present invention. Detailed Implementation

[0085] The present invention will be further described in detail below with reference to embodiments. These embodiments are illustrative of the invention and the invention is not limited thereto. Unless otherwise specified, the features in the following embodiments can be combined with each other.

[0086] Example 1:

[0087] A multi-layered fusion method for attributing data anomalies in building system sensor devices, such as Figure 1 As shown, it includes the following steps:

[0088] S100: Acquire current sensing data from different sensing devices in the building system;

[0089] S200: Based on time series, acquire historical data of different sensing devices in the building system to obtain historical sensing data sequences, and preprocess the historical sensing data sequences to obtain preprocessed sensing data sequences.

[0090] S300: Determine extreme outliers based on whether the current sensor data exceeds a reasonable range; wherein, the reasonable range is obtained by dividing the preprocessed sensor data sequence based on quartiles;

[0091] S400: Remove data that exceeds the reasonable range from the preprocessed sensor data sequence to obtain a non-extreme sensor data sequence;

[0092] S500: Calculate the deviation of the current sensor data from the data in the non-extreme sensor data sequence to obtain the trend outlier;

[0093] S600: The current sensing data is inferred using a pre-trained structural anomaly model to obtain the predicted value of the current data; the predicted value of the current data is compared with the current sensing data to obtain the structural anomaly value;

[0094] S700: Based on the current sensor data and combined with the operation rule model, the total value of the rule anomaly weight that meets the anomaly conditions is compared with the total value of the rule anomaly weight that applies the anomaly conditions to obtain the rule anomaly value; wherein, the operation rule model is constructed based on the sensor device anomaly conditions and the rule anomaly weight that meets the anomaly conditions, and the anomaly conditions are set according to the operation rules of the sensor device in the building system.

[0095] S800: Compare extreme outliers, trend outliers, structural outliers, and regular outliers with preset outlier thresholds to obtain the cause of the sensor data anomaly.

[0096] In S200, historical data from different sensors in the building system are acquired based on time series data to obtain a historical sensor data sequence. This historical sensor data sequence is then preprocessed to obtain a preprocessed sensor data sequence, including the following steps:

[0097] S210: Obtain historical data of sensing devices in the building system based on time series data to obtain historical sensing data sequences.

[0098] There are many sensors in a building system. Among them, temperature sensors, humidity sensors, airflow sensors, and power sensors are relatively important. Therefore, the data from these sensors must include at least two of the following: temperature, humidity, airflow, and power.

[0099] S220: Preprocess the historical sensor data sequence to obtain a preprocessed sensor data sequence.

[0100] The invalid data of each sensor device data sequence in the historical sensor data sequence is removed and filled in to obtain the preprocessed sensor device data sequence; wherein, the invalid data includes at least duplicate data and missing data.

[0101] Invalid data in sensor data is usually caused by data quality issues or acquisition anomalies. Common invalid data types include, but are not limited to, the following: Missing data, caused by sensor malfunctions, communication interruptions, storage errors, etc., resulting in data points not being recorded or lost during transmission, manifested as null values, NULL, NaN, etc.; Duplicate data, caused by retransmission, storage logic errors, improper sampling frequency settings, etc., resulting in identical or highly similar data points being recorded multiple times; Fixed-value data, caused by sensor freezing, software crashes, exceeding measurement limits, etc., resulting in data remaining unchanged for a long time; Invalid format data, caused by transmission protocol errors, parsing failures, etc., resulting in data format errors; Timestamp errors, caused by clock missynchronization, storage anomalies, etc., resulting in missing or out-of-order timestamps, etc.

[0102] Data imputation can be achieved using one or more of the following methods: filling missing data with fixed values ​​(mean, median, etc.) or sliding window statistics; linear, polynomial, or time-weighted interpolation; utilizing the periodicity of the data to impute it with historical data from the same period; predicting missing values ​​using models such as ARIMA, LSTM, or random forest; or imputing data based on physical laws or business rules. The appropriate data imputation method should be selected based on the characteristics of the data from different sensing devices.

[0103] If the historical data of a certain sensor is all zero within a sufficiently long time window, and the physical quantity monitored by the sensor is theoretically impossible to be zero for a long time (such as temperature, current, gas concentration, etc.), it can be preliminarily determined that it is in an extreme abnormal state, and it is necessary to investigate whether there are problems such as power supply failure, sensor failure or communication interruption.

[0104] In S300, extreme outliers are determined based on whether the current sensor data exceeds a reasonable range, as follows:

[0105] In addition to invalid data with the aforementioned format errors, abnormal data from sensing devices also includes abrupt changes caused by sampling faults, bit errors, etc., requiring detection of such abrupt anomalies. The reasonable range is obtained by dividing the preprocessed sensing data sequence based on quartiles; therefore, the reasonable range for each sensing device's data is expressed as follows:

[0106]

[0107] in, Indicates the first... The first quartile of a sensor's data represents the 25th percentile of the data. Indicates the first... The third quartile of a sensor's data represents the 75th percentile of the data. Indicates the first... The interquartile range of data from each sensor reflects the range of data variation. . They represent the first The weights of the first and third quartiles of the data from each sensor device. All are positive numbers, and can be set according to the characteristics of the data from each sensor device. The value, for example, the commonly set It is 1.5. It is also 1.5.

[0108] If the current sensing data of the target sensing device exceeds the reasonable range of the corresponding sensing device data, then the target sensing device has an extreme anomaly, and the extreme anomaly value is assigned to the first preset extreme anomaly value; if the current sensing data of the target sensing device is within the reasonable range of the corresponding sensing device data, then the target sensing device does not have an extreme anomaly, and the extreme anomaly value is assigned to the second preset extreme anomaly value.

[0109] The extreme outliers are represented as follows:

[0110]

[0111] in, Indicates the first Extreme anomalies of individual sensors, Indicates the first The first preset extreme anomaly value for each sensor device Indicates the first The second preset extreme anomaly value of each sensing device Indicates the first Current sensing data of each sensing device This represents the current timestamp. Typically, an extreme exception is represented by a value of 1, and an exception is represented by a value of 0. Therefore, the first preset extreme exception value is set to 1, and the second preset extreme exception value is set to 0.

[0112] The IQR value is unaffected by the data distribution, is suitable for removing extreme values ​​from normally and skewed data, and is easy to detect sampling faults and bit errors in sensing devices.

[0113] In S400, data that exceeds the reasonable range in the preprocessed sensor data sequence is removed to obtain a non-extreme sensor data sequence.

[0114] Mutant data is a strong interference point affecting the model training in this invention and must be removed because: (1) The subsequent trend anomaly detection analysis relies on the mean and standard deviation of the preprocessed sensor data sequence. If there are extreme outliers in the preprocessed sensor data sequence, it will cause the mean to shift and the standard deviation to increase, thereby affecting the accuracy of trend anomaly detection. Therefore, before calculating the mean and standard deviation, extreme outliers (data outside the reasonable range) in the preprocessed sensor data sequence need to be removed so that the calculation of trend anomalies is based on more stable data. (2) The preprocessed sensor data sequence is also used for training the pre-trained artificial intelligence model in structural anomaly detection. Extreme outliers also need to be removed in order to learn the relationship between normal data from different sensing devices, thereby improving the accuracy of the trained model.

[0115] In S500, the deviation of the current sensing data relative to data in the non-extreme sensing data sequence is calculated to obtain trend anomalies. Anomalous data from the sensing device also includes drift over time due to factors such as sensor aging, calibration failure, and cumulative environmental effects, causing the data to slowly deviate from the true value. This data drift reflects a trend-based anomaly in the sensing device data, requiring further analysis to identify long-term, slowly changing problems that cannot be captured by extreme anomalies. This includes the following steps:

[0116] S510: Obtain the most recent data within a first preset duration from the non-extreme sensing data sequence to obtain a first-duration sensing data sequence, and then obtain the mean and standard deviation of the first-duration sensing data sequence. The first preset duration can be set according to the sampling frequency and anomaly detection accuracy of the sensing device. Selecting several recent data points as the basis for trend anomaly detection allows focusing on the latest data, quickly reflecting the current state, reducing historical noise interference, and more sensitively capturing short-term anomalies. It eliminates the need for long-term storage of large amounts of historical data, reducing processing complexity and storage costs.

[0117] S520: The skewness of the non-extreme sensing data sequence is calculated using conventional skewness calculation methods and is known as the first sequence skewness.

[0118] Based on the current sensing data, the mean and standard deviation of the first time-duration sensing data sequence, the baseline trend deviation value is obtained. The baseline trend deviation value is expressed as follows:

[0119]

[0120] in, Indicates the first The baseline trend deviation of data from a sensor device represents the degree of standardization to which the current data deviates from the historical average. Indicates the first Current sensor data from each sensor device. Indicates the current timestamp. Indicates the first preset duration. Indicates the first The average value of the first time-long sensor data sequence from each sensor device. Indicates the first The standard deviation of the first duration sensor data sequence of the sensor device data.

[0121] The calculation method for baseline trend deviation is applicable to data following a normal distribution. Current technology generally uses a baseline trend deviation greater than the baseline deviation (typically 3) as the criterion for anomaly detection. However, when the data is skewed, using a baseline deviation of 3 as the criterion can easily lead to false positives or false negatives. Furthermore, some sensor data in building systems (such as current, voltage, and cooling load) are positively skewed, which can widen the standard deviation of the sensor data, masking trend anomalies and making them undetectable. It can also shift the mean of the sensor data, resulting in a distorted baseline trend deviation. Moreover, the probability of a baseline trend deviation of 3 in skewed data is far greater than 0.13% (the standard anomaly probability under a normal distribution), leading to false positives or false negatives.

[0122] Therefore, this invention introduces skewness checking before existing trend anomaly judgment, and performs basic trend deviation value determination or dynamic threshold setting based on skewness or after data transformation.

[0123] S530: When the first sequence skewness of a certain sensing device is less than the first preset skewness threshold, it indicates that the data of this sensing device follows a normal distribution. Existing technology can be used to determine whether there is an abnormal trend. The deviation from the reference value is set as the first preset deviation threshold. The first preset deviation threshold is preferably 3, but it can also be adjusted according to the actual situation.

[0124] S540: When the first sequence skewness of a certain sensing device is greater than or equal to a first preset skewness threshold (indicating that the data of this sensing device is positively skewed, and judging the trend abnormality based on existing technology will cause false alarms), then the non-extreme sensing data sequence and the current sensing data are transformed to obtain a transformed data sequence and transformed current sensing data; the skewness of the transformed data sequence is calculated to obtain a second sequence skewness; based on the comparison between the second sequence skewness and the first preset skewness threshold and the second preset skewness threshold, the deviation from the reference value is obtained. This includes the following steps:

[0125] (1) When all data in the non-extreme sensing data sequence are positive, the non-extreme sensing data sequence is subjected to maximum likelihood estimation to obtain the first transformation parameter; based on the first transformation parameter, the non-extreme sensing data sequence and the current sensing data are subjected to the first transformation to obtain the first transformed data sequence and the first transformed current sensing data. When there are negative data in the non-extreme sensing data sequence, the non-extreme sensing data sequence is subjected to maximum likelihood estimation to obtain the second transformation parameter; based on the second transformation parameter, the non-extreme sensing data sequence and the current sensing data are subjected to the second transformation to obtain the second transformed data sequence and the second transformed current sensing data. The first transformation is expressed as follows:

[0126]

[0127] The second transformation process is represented as follows:

[0128]

[0129] in, Represents the first transformed data sequence or the second transformed data sequence. The first data sequence of the sensor device One data point, Indicating the first non-extreme sensing data sequence The first data sequence of the sensor device One data point, Indicates the first The first transformation parameter of the data from the sensor device Indicates the first The second transformation parameter of the data from the first sensor device. Current sensing data of each sensing device Perform either the first transformation process or the second transformation process to obtain the current sensing data after the first transformation or the current sensing data after the second transformation. The formulas for the first transformation process and the second transformation process described above can be used, and will not be repeated here.

[0130] The first transformation process applies a power transformation to positive numbers, thereby reducing the skewness of non-extreme sensor data sequences. The second transformation applies a power transformation to negative and zero values, further reducing the skewness of non-extreme sensor data sequences. Through either the first or second transformation process, the skewness of non-extreme sensor data sequences can be reduced, thus ensuring they conform to a normal distribution and improving the accuracy of trend anomaly detection. Data in non-extreme sensor data sequences typically do not contain all zeros. Even if such a situation occurs, it indicates an extreme anomaly in the sensing device, which can ultimately be identified and therefore does not affect the anomaly attribution result.

[0131] In the maximum likelihood estimation process, both the first and second transformation parameters are obtained by performing maximum likelihood estimation on the non-extreme sensor data sequences. Statistical software such as Python's `scipy.stats.boxcox` or R's `car::powerTransform()` can be used to automatically calculate the optimal first or second transformation parameters. It is important to note that the first and second transformation parameters are not the same because the underlying non-extreme sensor data sequences used to obtain them are different.

[0132] (2) The skewness of the transformed data sequence is calculated by conventional skewness calculation method, which is the second sequence skewness.

[0133] (3) Based on the comparison between the second sequence skewness and the first preset skewness threshold and the second preset skewness threshold, the deviation from the benchmark value is obtained, specifically including:

[0134] When the skewness of the second sequence is less than the first preset skewness threshold, it indicates that the transformed data sequence is approximately normally distributed. Existing technology can be used to determine whether there is an abnormal trend. The deviation benchmark value is the first preset deviation threshold.

[0135] When the skewness of the second sequence is greater than or equal to the first preset skewness threshold and less than the second preset skewness threshold, it indicates that the transformed data sequence is moderately skewed. It is necessary to reduce the deviation from the baseline value before existing technology can be used to determine whether there is an abnormal trend. That is, the deviation from the baseline value is the second preset deviation threshold.

[0136] When the skewness of the second sequence is greater than or equal to the second preset skewness threshold, it indicates that the transformed data sequence is obviously skewed or has a long tail. It is necessary to further reduce the deviation from the benchmark value before existing technology can be used to determine whether there is an abnormal trend. The deviation from the benchmark value is the third preset deviation threshold.

[0137] Therefore, the first preset deviation threshold > the second preset deviation threshold > the third preset deviation threshold.

[0138] The deviation from the reference value is expressed as follows;

[0139]

[0140] in, Indicates the first The deviation of each sensor from the reference value Indicates the first The first preset deviation threshold of each sensor device Indicates the first The second preset deviation threshold of the sensor device Indicates the first The third preset deviation threshold of each sensing device, and , Indicates the first The second sequence skewness of the data sequence from the sensor device Indicates the first The first preset skewness threshold of each sensing device Indicates the first The second preset bias threshold of the sensor.

[0141] S550: Based on the mean and standard deviation of the current sensing data and the first time-duration sensing data sequence, the transformation trend deviation value is obtained; the transformation trend deviation value is expressed as follows:

[0142]

[0143] in, Indicates the first The transformation of data from individual sensors to the current sensor data. Indicates the first Deviation of the trend of data transformation from individual sensors.

[0144] S560: Based on the deviation from the baseline value and the transformed trend deviation value, trend outliers are obtained; the trend outliers are represented as follows:

[0145]

[0146] in, Indicates the first Outliers in the trend of individual sensors This indicates that the smaller of the multiple values ​​will be returned. Indicates the first The deviation of the data from the reference value of each sensor device.

[0147] Trend outliers are mainly used to capture the statistical deviation behavior of a single variable within a certain time window. They are especially suitable for discovering drift-type anomalies that rise or fall slowly, periodic deviations or unstable control phenomena, and slight errors caused by aging or positional deviation of sensing devices.

[0148] Trend anomalies can trigger subsequent verification of physical operation mechanism rules, and can also trigger collaborative verification of system anomalies: it can determine whether the trend anomaly of this sensor device has caused anomalies in other sensors in the system (i.e., structural anomalies), thereby distinguishing whether it is a problem of offset of a single sensor device or a problem of related devices or the entire system. This embodiment supports sliding window dynamic analysis: the first preset duration can be set to 1 to 3 hours, suitable for real-time edge deployment. In addition, trend anomalies can also form a complementary and collaborative mechanism with other anomalies. When a trend anomaly is confirmed by structural anomalies or rule anomalies, the building system anomaly can be upgraded to a medium / high level warning; otherwise, it can be downgraded to an observation state to avoid false alarms.

[0149] The trend anomaly detection method provided by this invention has the following advantages: (1) It has a high sensitivity to detect trend anomalies / deviations, and can effectively identify anomalies that accumulate slowly and gradually shift in the data of sensing devices. It is suitable for early detection of drift-type faults, aging errors and control system response lag problems, and is an important component of the "early warning" mechanism. (2) It can adapt to different distributions with an adaptive threshold mechanism. By combining skewness calculation with the first transformation / second transformation, it realizes the normalization of skewed data, and then dynamically adjusts the deviation from the benchmark value according to the degree of skewness to avoid false alarms caused by non-normal distribution. (3) It has low computational cost and is easy to deploy at the edge. Compared with neural networks or complex graph models, the trend anomaly detection method provided by this invention only relies on basic statistics such as mean and standard deviation. The computational resource consumption is extremely low, and it is suitable for real-time anomaly prediction in low-power devices such as AI Edge Box. (4) Standardized data output facilitates fusion analysis. The output trend anomaly value is a continuous normalized score, which can be seamlessly integrated into the subsequent total anomaly value model (weighted summation with mutation anomaly value, structural anomaly value, and rule anomaly value). It can also be used as a "pre-screening module" for other algorithms. (5) Supports sliding window and trend analysis extension. The first preset duration (as needed) can be extended to support moving average, volatility judgment, piecewise regression and other analysis methods to further model and track complex trends such as periodic shifts and control instability. (6) Can be linked with physical rules to locate the root cause. After an anomaly is triggered in the building system, the trend anomaly value can drive the directional rule screening mechanism to check only the engineering rules involved in the relevant variables, which greatly improves the accuracy and interpretability of anomaly judgment.

[0150] The complementarity between mutational anomalies and trend anomalies can be illustrated by the following example:

[0151] (1) Continuous data of the air supply temperature of the terminal air conditioning in the building system were collected to form a time series sample of 30 points. The first 20 points were basically stable at around 22°C, and the last 10 points slowly rose to 27°C, simulating the abnormal trend of sensor drift or control accuracy decline.

[0152] The preprocessed sensor data sequence is as follows: 22.0, 22.1, 22.2, 22.0, 22.3, 22.2, 22.1, 22.4, 22.2, 22.0, 22.1, 22.0, 22.3, 22.1, 22.0, 22.2, 22.4, 22.3, 22.2, 22.1, 22.6, 23.1, 23.8, 24.2, 24.8, 25.3, 25.9, 26.4, 26.7, 27.0.

[0153] (2) Mutation anomaly determination: The calculated values ​​are Q1=22.1, Q3=24.35, and IQR=2.25. The upper limit of the reasonable range is Q3+1.5×IQR=27.725. Since the maximum value 27.0<27.725, no mutation anomaly was identified.

[0154] (3) Remove data that exceeds the reasonable range in the preprocessed sensor data sequence to obtain a non-extreme sensor data sequence.

[0155] (4) Skewness analysis: The first sequence skewness of the non-extreme sensing data sequence is calculated to be approximately 1.01, which is significantly right-biased, triggering the first transformation process.

[0156] (5) Perform maximum likelihood estimation on the non-extreme sensing data sequence to obtain the first transformation parameter of approximately 0.2.

[0157] (6) The first transformation parameter performs the first transformation process on the non-extreme sensing data sequence and the current sensing data respectively to obtain the transformed data sequence and the transformed current sensing data. The skewness (second sequence skewness) of the transformed data sequence is about 0.65, which is moderately skewed / close to normal. The deviation from the reference value is the second preset deviation threshold (2.5).

[0158] (7) Trend anomaly determination: The (second) trend deviation value calculated based on the current sensor data is approximately 2.76. Based on this, the trend anomaly value is calculated. Since the trend deviation value is greater than the deviation from the baseline value, the building system successfully identifies that the sensor device has a trend anomaly and triggers an alarm.

[0159] This example demonstrates how, in situations where abrupt changes in a building system cannot detect trend anomalies, further processing can be used to successfully detect trend anomalies, showcasing the complete mechanism chain of "statistics → transformation → standardization → judgment" within the system.

[0160] In S600, a pre-trained structural anomaly model is used to infer the current sensing data to obtain the predicted value of the current data; the predicted value of the current data is compared with the current sensing data to obtain the structural anomaly value, including the following steps:

[0161] The pre-trained structural anomaly model can be an existing, pre-trained model, or a model that is self-constructed and trained based on the specific characteristics of the building system. This embodiment provides a method for self-construction and training, as detailed below:

[0162] (1) Based on the time series, data of different sensing devices in the building system during normal operation are obtained to obtain the original normal sensing data sequence; based on the unified timestamp, the data sequences of different sensing devices in the original normal sensing data sequence are aligned to obtain the aligned sensing data sequence; the aligned sensing data sequence is normalized to obtain the normalized aligned sensing data sequence; the target sensing device data is set, and the target training sample dataset is obtained by combining the normalized aligned sensing data sequence.

[0163] Time alignment unifies the time base of different sensors, resolves differences in sampling frequency and latency, ensures data synchronization (such as matching camera and radar data), avoids temporal discrepancies, and meets the requirements of artificial intelligence models for regular temporal input. Normalization eliminates dimensional differences between sensors (such as temperature and voltage), scaling all features to a similar range (such as [0,1]), preventing features with large values ​​from dominating model training, while accelerating convergence and improving generalization ability. The normalized and aligned sensor data sequence obtained after time alignment and normalization is high-quality, spatiotemporally consistent data. As training data, it enables pre-trained models to learn the correlation features across sensors more accurately, improving robustness. Ignoring these steps will lead to a significant decrease in model performance.

[0164] In real-world physical systems, selective correlations often exist between data from multiple sensors; that is, structural anomalies in the target sensor are only related to data from some sensors. Traditional methods require manual screening of correlated sensors, while this invention automatically learns these correlations through a pre-trained artificial intelligence model. The model autonomously identifies sensor data related to the target sensor and dynamically assigns weights, giving higher weights to highly correlated sensor data and lowering or reducing the weights of low-correlation data. In use, only the target sensor needs to be specified, and the trained model can automatically weight and process multi-source sensor data based on the learned correlation patterns, without requiring manual intervention in correlation determination.

[0165] (2) Based on the characteristics of specific building systems, construct a pre-trained model of structural anomalies using artificial intelligence models.

[0166] In this embodiment, a pre-trained model of structural anomalies is constructed based on AutoEncoder (an unsupervised learning neural network model) to illustrate how to construct a pre-trained model of structural anomalies. However, it is not limited to the pre-trained model of structural anomalies in this embodiment. In practice, different types of pre-trained models of structural anomalies can be constructed as needed.

[0167] The structural anomaly pre-trained model in this embodiment comprises an encoder and a decoder. The encoder includes an input layer, several hidden layers, and a bottleneck layer. The decoder includes a bottleneck layer, several hidden layers, and an output layer. In the encoder: the input layer receives the input data and converts it into a format that the neural network can process (such as a vector or tensor). The hidden layers progressively extract high-order features from the input data and reduce the dimensionality of the data through nonlinear transformations (such as ReLU and Sigmoid). The number of neurons in each hidden layer typically decreases layer by layer, forming a "compressed" structure. The bottleneck layer forces the network to compress the input data into a low-dimensional latent space with a dimension much lower than that of the input layer, generating a compact representation (i.e., encoding) of the data. The output of the encoder is also the input of the decoder. In the decoder: the bottleneck layer is shared with the encoder and receives the latent encoding generated by the encoder, serving as the starting point for the decoder's input. The hidden layers progressively reconstruct the low-dimensional latent encoding into a high-dimensional representation of the original data, restoring data details through operations such as upsampling or deconvolution. The number of neurons in each layer typically increases layer by layer, forming an "expanded" structure. The output layer generates reconstructed data with the same dimension as the input layer.

[0168] There are weight matrices between the layers of the encoder and decoder. The goal of the pre-trained structural anomaly model is to minimize the reconstruction error by continuously adjusting the weight matrices, that is, to minimize the error between the model's output value and the input value. The reconstruction error model is represented as follows:

[0169]

[0170] in, Indicates the first The reconstruction error of the pre-trained model for structural anomalies of a single sensing device Indicates the first The first sensing device One data point, Indicates the first The pre-trained model of structural anomalies of the first sensing device is used for the first... The first of the sensing devices Predicted values ​​for each data point. Indicates the first The total amount of data used for model training from each sensor device.

[0171] (3) The structural anomaly pre-training model is trained and tested using the target training sample dataset to obtain the structural anomaly model.

[0172] The task of model training is to determine the corresponding weight matrices based on the principle of minimizing reconstruction error. After training, testing and validation, the trained model, namely the structural anomaly model, is obtained. The structural anomaly model can be used to: predict the target sensor data based on the input data of multiple sensor devices (excluding the target sensor data), and obtain the predicted value of the target sensor data. This predicted value is used to compare with the actual data of the target sensor device to determine whether there is a structural anomaly in the sensor device.

[0173] S610: The current sensing data of the target sensing device is inferred by using the pre-trained structural anomaly model of the target sensing device to obtain the predicted value of the current data of the target sensing device.

[0174] Since the structural anomaly model is trained based on data from normally operating sensors, the predicted values ​​of the target sensors obtained from the model are reasonable. The reconstruction error of normally operating sensors is small, while the reconstruction error under abnormal conditions is relatively large. If the error between the actual value and the predicted value (reasonable value) of the target sensor exceeds a preset level, it indicates that the target sensor has a structural anomaly.

[0175] S620: The difference and square of the current sensing data and the current data prediction value of the target sensing device are calculated to obtain the current reconstruction error of the target sensing device. The current reconstruction error of the target sensing device is expressed as follows;

[0176]

[0177] in, Indicates the first The current reconstruction error of each sensor device Indicates the first Current sensing data of each sensing device Indicates the first The structural anomaly model of the first sensing device predicts the first... The current data prediction value of each sensor device.

[0178] S630: Based on a preset reconstruction error threshold and the current reconstruction error of the target sensing device, the structural anomaly value of the target sensing device is obtained; the structural anomaly value of the target sensing device is expressed as follows:

[0179]

[0180] in, Indicates the first The structural anomaly value of each sensing device is the structural anomaly value of the target sensing device. Indicates the first Current timestamp of each sensor device The reconstruction error is the current reconstruction error of the target sensing device. Indicates the first The preset reconstruction error threshold for each sensing device can be set by multiplying the maximum or minimum reconstruction error generated during the training of the structural anomaly pre-training model by a preset coefficient. The coefficient corresponding to the maximum reconstruction error should be less than 1, and the coefficient corresponding to the minimum reconstruction error should be greater than 1. This indicates that the smaller of the multiple values ​​will be returned.

[0181] S640: By obtaining the structural anomaly values ​​of other sensing devices using the same method described above, it is only necessary to set other sensing devices in the normalized and aligned sensing data sequence as the target sensing device one by one.

[0182] The structural anomaly values ​​provided by this invention are suitable for identifying implicit anomalies caused by imbalances in the relationships between multiple sensing devices, playing a role in system-level consistency judgment within the overall architecture. The structural anomaly detection method provided by this invention eliminates the need for manually setting rules, is highly adaptable, can handle complex collaborative relationship anomalies between sensing devices, can be embedded in edge computing gateways, and can be jointly inferred with anomalies from other sensing devices to enhance system anomaly identification capabilities. In subsequent anomaly attribution methods, structural anomalies can be linked with trend anomalies to determine whether it is a systemic imbalance or a local drift. Structural anomalies can also be jointly verified with rule anomalies to determine whether the structural anomaly violates the device's operating logic.

[0183] The following example illustrates the linkage between structural anomalies and other anomalies: In an office area, the supply air temperature is maintained at 22.0°C and the return air temperature at 26.0°C during midday, but the ambient temperature fails to decrease. Trend anomalies indicate the supply air temperature is normal, and neither sudden nor regular anomalies trigger any anomalies. However, the detected airflow during this period is only 180 CMH, far below the normal cooling requirement of 300 CMH. Simultaneously, the main unit power is normal, and the set temperature remains unchanged. The system calculates the current multivariate structural reconstruction error to be 1.8, with a preset reconstruction error threshold of 1.2, resulting in a structural anomaly value of 1.0. The building system classifies this anomaly as "insufficient supply airflow leading to abnormal cooling efficiency" and indicates a medium-level structural anomaly.

[0184] In S700, based on the current sensor data and combined with the operating rule model, the total value of rule anomaly weights that meet the anomaly conditions is compared with the total value of rule anomaly weights that apply the anomaly conditions to obtain the rule anomaly value, including the following steps:

[0185] The operational rule model is constructed based on abnormal conditions of the sensing devices and the abnormal weights of the rules that satisfy these abnormal conditions. These abnormal conditions are set according to the operational rules of the sensing devices in the building system. Corresponding abnormal conditions are set according to the operational rule mechanism of a single sensing device or different sensing devices in the building system, and the rule weights corresponding to a particular abnormal condition are set according to the importance of that abnormal condition within the operational rule mechanism. Table 1 shows an example of the operational rule model. Those skilled in the art can adjust it as needed, including but not limited to adding or removing abnormal conditions and adjusting the size of the rule weights corresponding to the abnormal conditions.

[0186] Table 1: Example of Operation Rule Model

[0187]

[0188] Among them, the first preset rule is the abnormal weight. (Corresponding strong rule) > Second preset rule anomaly weight (Corresponding rule) > Third preset rule anomaly weight (Corresponding to weak rules), among which exceptions 1, 3, and 11 are critical exceptions.

[0189] S710: If the current sensing data of the target sensing device meets the abnormal conditions, the corresponding rule abnormal weight is added to the target satisfying rule abnormal weight, thus obtaining the total value of the target satisfying rule abnormal weight. The sensing data is considered to meet the abnormal conditions when it simultaneously meets the following two conditions: the device and sensing device match the device and sensing device involved in the abnormal conditions; the sensing data value conforms to the threshold range or specification of the abnormal conditions. Iterate through the abnormal conditions; if the current sensing data meets the abnormal condition, the rule abnormal weight corresponding to that abnormal condition is accumulated into the target satisfying rule abnormal weight (initially 0). The final target satisfying rule abnormal weight is the total value of the target satisfying rule abnormal weight.

[0190] S720: If the current sensing data of the target sensing device is applicable to the abnormal condition, the corresponding rule abnormal weight is added to the target applicable rule abnormal weight, thus obtaining the total value of the target applicable rule abnormal weight. When the sensing data only meets the matching requirements of the device and sensing device, but the value does not meet the abnormal condition, it is considered an applicable (but not met) abnormal condition. Iterate through the abnormal conditions. If the current sensing data is applicable (but not met) to the abnormal condition, the rule abnormal weight corresponding to the abnormal condition is accumulated into the target applicable rule abnormal weight (initially 0). The final target applicable rule abnormal weight is the total value of the target applicable rule abnormal weight.

[0191] S730: Based on the ratio of the total weight of the target satisfying rule anomalies to the total weight of the target applying rule anomalies, the rule anomaly value of the target sensing device is obtained. The rule anomaly value of the target sensing device is expressed as follows:

[0192]

[0193] in, Indicates the first The rule outliers of each sensor device are the rule outliers of the target sensor device. Indicates the first The target of each sensing device satisfies the rule of the total weight of anomalies. Indicates the first The total weight of the abnormality applicable rules for the target of each sensing device.

[0194] In the multi-level fusion sensing device data anomaly attribution method proposed in this invention, rule-based outliers are an outlier type that needs to be judged at all times, used to continuously monitor whether the sensing data violates engineering logic or physical constraints. During the calculation of rule-based outliers, the system defaults to calculating rule-based outliers at a fixed running cycle (e.g., every minute), verifying the engineering consistency of all sensing devices point by point.

[0195] However, considering the practical application needs of large-scale field sensor data or limited edge computing resources, the calculation frequency of rule anomalies can be flexibly configured to a lower cycle (e.g., once every 5 minutes), i.e., a low-frequency rule scanning strategy, to balance resource consumption and system load. In this low-frequency operation mode, the calculation of trend anomalies and structural anomalies will still be performed every minute. However, when the trend anomaly of the target sensor is within a preset weak trend anomaly range or the structural anomaly of the target sensor is within a preset weak structural anomaly range, even if it is not currently within the fixed operation cycle of rule anomalies, the physical rule set corresponding to that sensor will be triggered to perform targeted verification and calculate rule anomalies to help confirm potential engineering logic errors and avoid missed detections. The preset weak trend anomaly range is [second trend anomaly threshold, first trend anomaly threshold], and the preset weak structural anomaly range is [second preset structural anomaly threshold, first preset structural anomaly threshold]. For example, the preset weak trend anomaly range is [0.4, 0.7], and the preset weak structural anomaly range is [0.3, 0.8].

[0196] The method of this invention retains a periodic rule scanning mechanism, performing a complete physical rule verification on all sensing devices within a set period to improve anomaly coverage and the completeness of logical consistency judgment. If the system operation mode prioritizes economy and energy efficiency, a low-frequency rule scanning strategy can be adopted, with rule judgments only executed locally in the event of weak anomalies. If the system emphasizes the real-time and comprehensiveness of anomaly identification, the execution frequency of the physical rule module can be kept consistent with that of the trend / structure module, thereby achieving full real-time engineering logic verification. This application does not limit the specific method used, but a low-frequency rule scanning strategy is used to illustrate this in the embodiments.

[0197] S740: Using the above method, calculate the regular outlier values ​​of other sensing devices.

[0198] The rule anomaly mechanism embodies the multi-layered fusion judgment path of this invention, from data-driven to physical logic constraints. It supports flexible configuration of computational intensity according to scenario requirements, achieving dynamic balance and control of system resource consumption while ensuring accuracy and interpretability. The main function of rule anomalies is to link the execution of the engineering rule set corresponding to the sensing device for targeted verification. This provides a physical constraint supplement for anomaly detection in the entire building system, which is particularly important in scenarios where the sensing device does not detect anomalies but the system logic is flawed. Furthermore, a periodic rule scanning mechanism can be set in the low-frequency rule scanning strategy to ensure that data from unmarked sensing devices are fully covered by physical logic rules / constraints. Anomaly levels and corresponding response behaviors can also be set by comparing rule anomalies with corresponding preset thresholds. In the low-frequency rule scanning strategy, when the trend anomaly of a certain sensing device is weak (not obvious), the engineering rules involved in that sensing device are detected for targeted rule screening, avoiding computational redundancy and false alarms caused by global judgment.

[0199] The linkage between trend anomalies and rule anomalies can be illustrated by the following example: When the supply air temperature is marked as a weak anomaly, the system will verify relevant rules such as Rule 1 (cooling supply air temperature is less than return air temperature), Rule 6 (the temperature difference between the set temperature and the feedback temperature is less than the preset temperature difference threshold), and Rule 9 (temperature trend is consistent with air volume). Once a rule is detected to be triggered, the original weak anomaly can be upgraded to a valid alarm, realizing a closed-loop logic of trend anomaly → engineering rule confirmation → hierarchical output, improving diagnostic efficiency and reducing building system resource consumption.

[0200] When a weak structural anomaly is detected, the building anomaly detection system will locate the possible anomaly sensor based on the sensor with the largest reconstruction error in the anomaly data, and then perform directional logic verification in conjunction with the associated engineering physical rules. The linkage between structural anomalies and rule-based anomalies can be illustrated by the following example: A structural anomaly is detected in a cooling water machine room (reconstruction error of 2.2, preset reconstruction error threshold of 1.5), with the largest errors in flow rate, return water temperature, and main unit power variables. The system triggers rule 11 (water system energy balance) and rule 8 (main unit operating power exceeds minimum operating power). It finds that the power is normal but the return water temperature is unchanged, and determines the risk of equipment water circuit blockage due to insufficient water flow, escalating it to a serious anomaly and triggering an alarm. This mechanism forms a closed-loop linkage chain: structural anomaly → anomaly main variable identification → linkage with physical rules → engineering confirmation → tiered output.

[0201] The method for calculating rule outliers provided by this invention originates from the actual physical logic and operating mechanism of the building system. All judgments can be traced back to explicit equipment status and operating rules, facilitating understanding, verification, and tracking by maintenance personnel. Different rule bases and weight systems can be configured according to different building system scenarios, and it is compatible with various air conditioning system structures (such as fan coil units + fresh air systems, water-cooled central air conditioning, VRV systems, etc.). In scenarios where sudden changes, trend anomalies, and structural anomalies cannot be identified or are easily misjudged (such as skewed changes and soft logic errors), rule outliers can provide physical constraints as a last line of defense. Certain strong rules can also be set as independent alarm channels, enabling immediate response even if rule outliers do not exceed the threshold.

[0202] In S800, extreme outliers, trend outliers, structural outliers, and regular outliers are compared with preset outlier thresholds to determine the cause of sensor data anomalies, including the following steps:

[0203] S810: The extreme outliers, trend outliers, structural outliers, and regular outliers of the target sensing device are weighted and summed to obtain the total outlier value of the target sensing device. The total outlier value of the target sensing device is expressed as follows:

[0204]

[0205] in, Indicates the first The total abnormal value of each sensor device Indicates the first Extreme anomalies of individual sensors, Indicates the first Outliers in the trend of individual sensors Indicates the first The structural anomaly value of each sensing device is the structural anomaly value of the target sensing device. Indicates the first The regular abnormal values ​​of the sensor device They represent the first The weights of extreme outliers, trend outliers, structural outliers, and regular outliers for each sensing device, and Extreme outliers represent extreme value jumps caused by sampling errors or other reasons. Extreme outliers have two fixed values: a first preset extreme outlier (1) when an extreme outlier exists, and a second preset extreme outlier (0) when no extreme outlier exists. Trend outliers represent trend-type drift anomalies caused by factors such as sensor aging. Structural outliers represent anomalies caused by imbalances in the relationship between different sensors. Regular outliers represent anomalies caused by violations of engineering physics rules between different sensors under weak trend or weak structural anomalies. Trend, structural, and regular outliers have been normalized during calculation to be continuous values ​​between [0,1] (inclusive), ensuring comparability between different outliers.

[0206] The weights of extreme outliers, trend outliers, structural outliers, and regular outliers from sensing devices can be learned and dynamically adjusted based on labeled data to adapt to the distribution of anomaly patterns under different sensing devices, buildings, seasons, and control strategies. Table 2 shows examples of weight setting strategies, but the weight settings for different strategy types in Table 2 only illustrate the trend of the values; the specific values ​​are not limited. Among them, the model-driven strategy is suitable for mature systems with sufficient samples and well-trained models. In the early stages, it is recommended to use a balanced or empirical strategy.

[0207] Table 2 Outlier Weight Settings for Different Strategy Types

[0208]

[0209] If an outlier is missing, its weight is automatically reset to zero, and the weights of the remaining outliers are normalized.

[0210] S820: When the total abnormal value of the target sensor exceeds the preset total abnormal value threshold, the target sensor in the building system is abnormal. The order of judgment for the cause of the abnormality is as follows:

[0211] (1) First, determine whether there is a rule anomaly. When the rule anomaly value of the target sensing device is greater than the corresponding first preset rule anomaly threshold (e.g., 0.8), that is, when there is an obvious rule anomaly, the target sensing device has a physical logic conflict anomaly. The rule anomaly value is calculated when it is impossible to determine whether there is a trend anomaly or a structural anomaly. Therefore, if there is a rule anomaly, there will not be both a trend anomaly and a structural anomaly at the same time.

[0212] (2) Secondly, determine whether there are structural or trend anomalies:

[0213] When the structural anomaly value of the target sensing device is greater than the corresponding first preset structural anomaly threshold (e.g., 0.8) and the regular anomaly value is less than the corresponding second preset regular anomaly threshold (e.g., 0.5), that is, there is obvious structural anomaly but no obvious regular anomaly, then the target sensing device has a structural anomaly.

[0214] When the trend anomaly value of the target sensing device is greater than the corresponding first trend anomaly threshold (e.g., 0.7), the extreme anomaly value is the second preset extreme anomaly value (0, no anomaly), the structural anomaly value is less than the corresponding second preset structural anomaly threshold (lower value), and the regular anomaly value is less than the corresponding third preset regular anomaly threshold (lower value), that is, when there is a significant trend anomaly and there are no extreme anomalies, structural anomalies, or regular anomalies, then the target sensing device has a trend anomaly.

[0215] Since rule outliers are calculated when it is impossible to determine whether trend anomalies or structural anomalies exist, when rule outliers exist, it is possible that either trend anomalies or structural anomalies exist, or neither trend anomalies nor structural anomalies exist.

[0216] (3) Then, determine whether there is an extreme anomaly. When the extreme anomaly value of the target sensing device is the first preset extreme anomaly value (1, an anomaly exists), then the target sensing device has an extreme anomaly.

[0217] (4) Finally, if none of the extreme outliers, trend outliers, structural outliers, and rule outliers of the target sensing device meet any of the above conditions, then the outlier of the target sensing device is an unattributed outlier. For unattributed outliers, the one or two outliers with the highest values ​​are listed as reference attribution prompts, which may be prompted as: "suspected trend outlier + structural disturbance" or "low-level logical conflict", etc. Unattributed outliers will also be recorded in the background manual review or training sample queue for further optimization of thresholds or judgment rules.

[0218] Among them, the first preset rule abnormal threshold > the second preset rule abnormal threshold > the third preset rule abnormal threshold, the first preset structure abnormal threshold > the second preset structure abnormal threshold, and the values ​​of the second preset structure abnormal threshold and the third preset rule abnormal threshold are relatively low, and can be set to 0.3 or lower.

[0219] The system can also set anomaly levels and corresponding response strategies based on the total anomaly value. For example: 0.00 ≤ total anomaly value < 0.30, normal, negligible, no response required; 0.30 ≤ total anomaly value < 0.60, minor anomaly, recording and observation recommended; 0.60 ≤ total anomaly value < 0.85, moderate anomaly, background alert recommended; total anomaly value ≥ 0.85: severe anomaly, immediate alarm recommended and manual verification. These range values ​​are set based on experience in debugging and engineering of anomaly distribution in the building system, and can be further adaptively optimized based on historical labeled data.

[0220] Furthermore, response strategies can be further refined based on the cause, level, and duration of sensor data anomalies, as shown in Table 3, which illustrates the correspondence between the cause, level, duration, and response strategy of sensor data anomalies.

[0221] Table 3. Correspondence between causes, levels, durations, and response strategies for sensor data anomalies.

[0222]

[0223] Furthermore, the total abnormal values ​​of different sensing devices can be weighted and summed to obtain the overall abnormal situation of the sensing devices in the building system. When the overall abnormal value exceeds the set overall abnormal threshold, it is considered that there is an overall abnormality in the sensing devices in the building system. Based on the above method, the specific sensing device that has the abnormality and the cause of the abnormal sensing device can be determined.

[0224] Unlike traditional single-algorithm judgment, this invention proposes a multi-layered fusion method for attributing anomalies in building system sensor data. This method employs a multi-layered fusion mechanism, prioritizing statistics, then structure, and finally physical logic, possessing dynamic collaborative characteristics of layer-by-layer enhancement, cross-verification, and result feedback. Based on historical data, it obtains extreme anomalies, trend anomalies, and structural anomalies in the current sensor data, and executes an engineering rule set to obtain rule-based anomalies for targeted verification. After obtaining the anomaly values ​​corresponding to each type of anomaly, these anomalies are uniformly scored, and anomaly level, type, and response suggestions are output accordingly. This multi-layered fusion method for attributing anomalies in building system sensor data possesses fault tolerance, interpretability, and practicality, and is suitable for anomaly early warning needs in complex building environments.

[0225] Example 2:

[0226] A multi-layered fusion system for attributing data anomalies in building system sensors, such as Figure 2As shown, it includes a current data module 100, a historical data module 200, an extreme anomaly module 300, a data removal module 400, a trend anomaly module 500, a structural anomaly module 600, a rule anomaly module 700, and an anomaly attribution module 800.

[0227] The current data module 100 acquires current sensing data from different sensing devices in the building system;

[0228] The historical data module 200 acquires historical data from different sensing devices in the building system based on time series, obtains historical sensing data sequences, and preprocesses the historical sensing data sequences to obtain preprocessed sensing data sequences.

[0229] The extreme anomaly module 300 determines extreme anomaly values ​​based on whether the current sensing data exceeds a reasonable range; wherein, the reasonable range is obtained by dividing the preprocessed sensing data sequence based on quartiles;

[0230] The data removal module 400 removes data that exceeds the reasonable range from the preprocessed sensor data sequence to obtain a non-extreme sensor data sequence.

[0231] The trend anomaly module 500 calculates the deviation of the current sensing data from the data in the non-extreme sensing data sequence to obtain the trend anomaly value.

[0232] The structural anomaly module 600 infers the current sensing data using a pre-trained structural anomaly model to obtain the predicted value of the current data; and compares the predicted value of the current data with the current sensing data to obtain the structural anomaly value.

[0233] The rule anomaly module 700, based on the current sensing data and combined with the operating rule model, compares the total value of rule anomaly weights that meet the anomaly conditions with the total value of rule anomaly weights that apply the anomaly conditions to obtain the rule anomaly value; wherein, the operating rule model is constructed based on the anomaly conditions of the sensing device and the rule anomaly weights that meet the anomaly conditions, and the anomaly conditions are set according to the operating rules of the sensing device in the building system.

[0234] The anomaly attribution module 800 compares extreme anomalies, trend anomalies, structural anomalies, and regular anomalies with a preset anomaly threshold to obtain the cause of the sensor device data anomaly.

[0235] Various changes and modifications made without departing from the spirit and scope of this invention, and all equivalent technical solutions, also fall within the scope of this invention.

[0236] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0237] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0238] This invention is described with reference to flowchart illustrations and / or block diagrams of the method, terminal device (system), and computer program product according to the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0239] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0240] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0241] It should be noted that:

[0242] The phrase "an embodiment" or "an embodiment" used in this specification means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the invention. Therefore, the phrase "an embodiment" or "an embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment.

[0243] Furthermore, it should be noted that the shapes and names of the parts and components described in the specific embodiments described in this specification may differ. All equivalent or simple variations made to the structure, features, and principles described in this patent concept are included within the protection scope of this patent. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to replace them, as long as they do not depart from the structure of this invention or exceed the scope defined in these claims, they should all fall within the protection scope of this invention.

Claims

1. A method for attributing data anomalies in a multi-layered fusion building system sensor device, characterized in that, Includes the following steps: Acquire current sensing data from different sensing devices in the building system; Historical data from different sensing devices in a building system are obtained based on time series data to obtain historical sensing data sequences. These historical sensing data sequences are then preprocessed to obtain preprocessed sensing data sequences. Extreme outliers are determined based on whether the current sensor data exceeds a reasonable range; wherein, the reasonable range is obtained by dividing the preprocessed sensor data sequence based on quartiles; Data exceeding the reasonable range in the preprocessed sensor data sequence is removed to obtain a non-extreme sensor data sequence. Calculate the deviation of the current sensor data from the data in the non-extreme sensor data sequence to obtain the trend outlier; The current sensor data is inferred by a pre-trained structural anomaly model to obtain the predicted value of the current data; the predicted value of the current data is compared with the current sensor data to obtain the structural anomaly value. Based on the current sensor data and combined with the operation rule model, the total value of the rule anomaly weight that meets the anomaly conditions is compared with the total value of the rule anomaly weight that applies the anomaly conditions to obtain the rule anomaly value; wherein, the operation rule model is constructed based on the sensor device anomaly conditions and the rule anomaly weight that meets the anomaly conditions, and the anomaly conditions are set according to the operation rules of the sensor device in the building system. By comparing extreme outliers, trend outliers, structural outliers, and regular outliers with preset outlier thresholds, the causes of data anomalies in the sensing device can be obtained. The step of comparing extreme outliers, trend outliers, structural outliers, and regular outliers with a preset outlier threshold to determine the cause of the sensor data anomaly includes the following steps: The extreme outliers, trend outliers, structural outliers, and regular outliers of the target sensing device are weighted and summed to obtain the total outlier value of the target sensing device. When the total abnormal value of the target sensing device exceeds the preset total abnormal value threshold, the target sensing device in the building system is abnormal, and the reason for the abnormality is as follows: When the rule anomaly value of the target sensing device is greater than the corresponding first preset rule anomaly threshold, the target sensing device has a physical logic conflict anomaly. When the structural anomaly value of the target sensing device is greater than the corresponding first preset structural anomaly threshold and the regular anomaly value is less than the corresponding second preset regular anomaly threshold, the target sensing device has a structural anomaly. When the trend anomaly value of the target sensing device is greater than the corresponding first trend anomaly threshold, there is no extreme anomaly, the structural anomaly value is less than the corresponding second preset structural anomaly threshold, and the regular anomaly value is less than the corresponding third preset regular anomaly threshold, then the target sensing device has a trend anomaly. When the target sensing device exhibits an extreme anomaly, the cause of the anomaly is an extreme anomaly. Among them, the first preset rule abnormal threshold > the second preset rule abnormal threshold > the third preset rule abnormal threshold, and the first preset structure abnormal threshold > the second preset structure abnormal threshold.

2. The method for attributing data anomalies in multi-level fusion building system sensor devices according to claim 1, characterized in that, The determination of extreme outliers based on whether the current sensor data exceeds a reasonable range is as follows: If the current sensing data of the target sensing device exceeds the reasonable range of the corresponding sensing device data, then the target sensing device has an extreme anomaly, and the extreme anomaly value is the first preset extreme anomaly value; if the current sensing data of the target sensing device is within the reasonable range of the corresponding sensing device data, then the target sensing device does not have an extreme anomaly, and the extreme anomaly value is the second preset extreme anomaly value. The reasonable range of the data from the sensing device is defined as follows: in, Indicates the first... The first quartile of the data from the sensor device Indicates the first... The third quartile of the data from the sensor device Indicates the first... The interquartile range of data from each sensor device , They represent the first The weights of the first and third quartiles of the data from each sensor device.

3. The method for attributing data anomalies in multi-level fusion building system sensor devices according to claim 1, characterized in that, The process of calculating the deviation of the current sensing data from the data in the non-extreme sensing data sequence to obtain trend outliers includes the following steps: Obtain the most recent data of the first preset duration in the non-extreme sensing data sequence to obtain the first duration sensing data sequence, and then obtain the mean and standard deviation of the first duration sensing data sequence; Calculate the skewness of the non-extreme sensing data sequence to obtain the first sequence skewness; When the skewness of the first sequence is less than the first preset skewness threshold, the deviation from the baseline value is the first preset deviation threshold. When the skewness of the first sequence is greater than or equal to the first preset skewness threshold, the non-extreme sensing data sequence and the current sensing data are transformed to obtain the transformed data sequence and the transformed current sensing data; the skewness of the transformed data sequence is calculated to obtain the second sequence skewness; the deviation from the reference value is obtained by comparing the second sequence skewness with the first preset skewness threshold and the second preset skewness threshold. The deviation value of the transformation trend is obtained based on the mean and standard deviation of the current sensing data and the first time-duration sensing data sequence. Based on the deviation from the baseline value and the deviation from the transformed trend value, trend anomalies are obtained; The deviation value of the transformation trend is represented as follows: The trend outliers are represented as follows: in, Indicates the first Deviation value of the changing trend of data from individual sensors Indicates the first Transformation of data from individual sensors to the current sensor data. Indicates the current timestamp. Indicates the first preset duration. Indicates the first The average value of the first time-long sensor data sequence from each sensor device. Indicates the first The standard deviation of the first duration sensor data sequence from each sensor device. Indicates the first Outliers in the trend of individual sensors This indicates that the smaller of the multiple values ​​will be returned. Indicates the first The deviation of the data from the reference value of each sensor device.

4. The method for attributing data anomalies in multi-level fusion building system sensor devices according to claim 3, characterized in that, The process of transforming the non-extreme sensing data sequence and the current sensing data to obtain the transformed data sequence and the transformed current sensing data includes the following steps: When all data in the non-extreme sensing data sequence are positive, the non-extreme sensing data sequence is subjected to maximum likelihood estimation to obtain the first transformation parameter; based on the first transformation parameter, the non-extreme sensing data sequence and the current sensing data are subjected to the first transformation to obtain the first transformed data sequence and the first transformed current sensing data. When there are negative numbers in the non-extreme sensing data sequence, the maximum likelihood estimation process is performed on the non-extreme sensing data sequence to obtain the second transformation parameter; based on the second transformation parameter, the second transformation process is performed on the non-extreme sensing data sequence and the current sensing data to obtain the second transformed data sequence and the second transformed current sensing data. The first transformation process is represented as follows: The second transformation process is represented as follows: in, Represents the first transformed data sequence or the second transformed data sequence. The first data sequence of the sensor device One data point, Indicating the first non-extreme sensing data sequence The first data sequence of the sensor device One data point, Indicates the first The first transformation parameter of the data from the sensor device Indicates the first The second transformation parameter of the data from the sensor device.

5. The method for attributing data anomalies in multi-level fusion building system sensor devices according to claim 3, characterized in that, The step of obtaining the deviation from the baseline value by comparing the second sequence skewness with the first preset skewness threshold and the second preset skewness threshold includes the following steps: When the skewness of the second sequence is less than the first preset skewness threshold, the deviation reference value is the first preset deviation threshold; When the skewness of the second sequence is greater than or equal to the first preset skewness threshold and less than the second preset skewness threshold, the deviation benchmark value is the second preset deviation threshold. When the second sequence skewness is greater than or equal to the second preset skewness threshold, the deviation benchmark value is the third preset deviation threshold; wherein, the first preset deviation threshold > the second preset deviation threshold > the third preset deviation threshold.

6. The method for attributing data anomalies in multi-level fusion building system sensor devices according to claim 1, characterized in that, The process of reasoning about the current sensing data using a pre-trained structural anomaly model to obtain a predicted value for the current data, and comparing the predicted value with the current sensing data to obtain structural anomaly values, includes the following steps: By inferring the current sensing data of the target sensing device through a pre-trained structural anomaly model of the target sensing device, the predicted value of the current data of the target sensing device is obtained. The current reconstruction error of the target sensor is obtained by squaring the difference between the current sensing data and the current data prediction value of the target sensor. Based on the preset reconstruction error threshold and the current reconstruction error of the target sensing device, the structural anomaly value of the target sensing device is obtained; The structural anomaly values ​​of other sensing devices were calculated; The structural anomalies of the target sensing device are represented as follows: in, Indicates the first The structural anomaly value of each sensing device is the structural anomaly value of the target sensing device. Indicates the first Current timestamp of each sensor device The reconstruction error is the current reconstruction error of the target sensing device. , Indicates the first The current sensing data of the target sensing device is the data from the current sensing device. Indicates the first Current data prediction values ​​of each sensor device Indicates the first The preset reconstruction error threshold for each sensor device This indicates that the smaller of the multiple values ​​will be returned.

7. The method for attributing data anomalies in multi-level fusion building system sensor devices according to claim 1, characterized in that, The process of comparing the total anomaly weight of rules that meet the anomaly conditions with the total anomaly weight of rules that apply the anomaly conditions, based on the current sensor data and in conjunction with the operating rule model, to obtain the rule anomaly value includes the following steps: If the current sensing data of the target sensing device meets the abnormal conditions, the corresponding rule abnormal weight is added to the target meets the rule abnormal weight, and then the total value of the target meets the rule abnormal weight is obtained. If the current sensing data of the target sensing device is applicable to abnormal conditions, the corresponding rule abnormal weight is added to the target applicable rule abnormal weight, thereby obtaining the total value of the target applicable rule abnormal weight. The rule anomaly value of the target sensing device is obtained by the ratio of the total weight of the target satisfying rule anomalies to the total weight of the target applying rule anomalies. Calculate regular outliers from other sensors; The rule-based outliers of the target sensing device are represented as follows: in, Indicates the first The regular abnormal values ​​of the sensor device Indicates the first The target of each sensing device satisfies the rule of the total weight of anomalies. Indicates the first The total weight of the abnormality applicable rules for the target of each sensing device.

8. A multi-level fusion data anomaly attribution system for building system sensors, characterized in that, The method described in any one of claims 1 to 7 is capable of being implemented, including a current data module, a historical data module, an extreme anomaly module, a data removal module, a trend anomaly module, a structural anomaly module, a rule anomaly module, and an anomaly attribution module; The current data module acquires current sensing data from different sensing devices in the building system; The historical data module acquires historical data from different sensing devices in the building system based on time series, obtains historical sensing data sequences, and preprocesses the historical sensing data sequences to obtain preprocessed sensing data sequences. The extreme anomaly module determines extreme anomaly values ​​based on whether the current sensing data exceeds a reasonable range; wherein, the reasonable range is obtained by dividing the preprocessed sensing data sequence based on quartiles; The data removal module removes data that exceeds the reasonable range from the preprocessed sensor data sequence to obtain a non-extreme sensor data sequence. The trend anomaly module calculates the deviation of the current sensing data from the data in the non-extreme sensing data sequence to obtain the trend anomaly value. The structural anomaly module infers the current sensing data using a pre-trained structural anomaly model to obtain the predicted value of the current data; it then compares the predicted value of the current data with the current sensing data to obtain the structural anomaly value. The rule anomaly module, based on the current sensor data and combined with the operating rule model, compares the total value of rule anomaly weights that meet the anomaly conditions with the total value of rule anomaly weights that apply the anomaly conditions to obtain the rule anomaly value; wherein, the operating rule model is constructed based on the sensor device anomaly conditions and the rule anomaly weights that meet the anomaly conditions, and the anomaly conditions are set according to the operating rules of the sensor devices in the building system. The anomaly attribution module compares extreme anomalies, trend anomalies, structural anomalies, and regular anomalies with preset anomaly thresholds to obtain the cause of the anomaly in the sensor data.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

10. A multi-layered fusion building system sensor data anomaly attribution device, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 7.

Images