A Machine Learning-Based Method and System for Anomaly Identification in Building Equipment

Abstract

This application relates to the field of machine learning technology, and provides a machine learning-based method and system for identifying anomalies in building equipment, enabling accurate detection and precise early warning of anomalies in building equipment. The method includes: acquiring a continuous operating data set of the target building equipment, the continuous operating data set containing multiple equipment status record units with timestamps; performing time-frequency domain feature extraction processing on the continuous operating data set to obtain a time-frequency domain feature set of the equipment status record units; calling a pre-built hybrid machine learning model to perform anomaly detection processing on the time-frequency domain feature set, generating anomaly identification results for the equipment status record units; determining the anomaly type of the target building equipment and the distribution characteristics of the anomaly type in the time dimension based on the anomaly identification results; generating a target early warning command containing a time location identifier based on the anomaly type and time distribution characteristics, and sending the equipment early warning command to the target equipment management terminal.

Description

Technical Field

[0001] This application belongs to the field of machine learning technology, specifically relating to a machine learning-based method and system for identifying anomalies in building equipment. Background Technology

[0002] Current building equipment management technologies have many shortcomings in terms of detection and early warning methods for equipment anomalies. Traditional technologies often rely solely on single-dimensional data feature analysis, failing to comprehensively grasp the equipment's operational status and easily overlooking complex and ever-changing anomalies. While some technologies can acquire equipment operational data, they lack in-depth analysis of the data over time, making it difficult to accurately pinpoint the time and type of anomalies.

[0003] Among these, some anomaly detection methods based on simple models suffer from low accuracy and reliability when faced with complex equipment operating modes and diverse anomaly situations, resulting in frequent false positives and false negatives. Moreover, most existing technologies fail to effectively integrate anomaly information with temporal distribution characteristics, leading to a lack of precise time positioning in the generated warning instructions. This makes it difficult for managers to perform timely and targeted equipment maintenance, increasing the impact of equipment failures on normal building operations and reducing equipment lifespan and operating efficiency. Summary of the Invention

[0004] This application provides a machine learning-based method and system for identifying building equipment anomalies, which can accurately detect and provide precise early warnings of building equipment anomalies, improve equipment management efficiency, and ensure stable equipment operation.

[0005] In a first aspect, embodiments of this application provide a machine learning-based method for identifying building equipment anomalies, applied to a building equipment anomaly identification system, the method comprising:

[0006] Acquire a continuous operating data set of equipment in the target building, wherein the continuous operating data set includes multiple equipment status record units with timestamps;

[0007] The continuous operating data set is subjected to time-frequency domain feature extraction processing to obtain the time-frequency domain feature set of the device status recording unit;

[0008] A pre-built hybrid machine learning model is invoked to perform anomaly detection processing on the time-frequency domain feature set, generating anomaly identification results for the device status recording unit;

[0009] Based on the anomaly identification results, determine the anomaly type of the target building equipment and the distribution characteristics of the anomaly type in the time dimension;

[0010] Based on the anomaly type and the time distribution feature information, a target early warning instruction containing a time location identifier is generated, and the device early warning instruction is sent to the target device management terminal.

[0011] Secondly, embodiments of this application provide a building equipment anomaly identification system, which includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the above-described method.

[0012] Thirdly, embodiments of this application provide a computer-readable storage medium including a computer program, which, when run on a building equipment anomaly identification system, causes the building equipment anomaly identification system to perform the steps of the above-described method.

[0013] The embodiments of this application construct a complete and efficient system for identifying and warning of abnormal building equipment.

[0014] First, by acquiring a set of continuously running data with timestamps, the state of the equipment at different times is fully recorded; the time-frequency domain feature extraction processing can deeply mine data features from both time and frequency domains, overcoming the limitations of single-dimensional analysis, more comprehensively depicting the operating characteristics of the equipment, and greatly improving the ability to capture abnormal features.

[0015] Secondly, the pre-built hybrid machine learning model combines the advantages of multiple algorithms to perform anomaly detection on time-frequency domain feature sets, significantly improving the accuracy and reliability of anomaly detection and effectively reducing false positives and false negatives. Based on the anomaly identification results, the anomaly type and spatiotemporal distribution characteristics are determined, enabling managers to clearly grasp the nature and development patterns of anomalies and providing a strong basis for targeted handling.

[0016] Furthermore, based on the above information, a target early warning instruction containing a time location identifier is generated and sent to the terminal, realizing accurate early warning. This allows managers to know in advance the time and type of anomalies, thereby arranging maintenance in a timely manner. This greatly improves the timeliness and effectiveness of equipment maintenance, reduces losses caused by equipment failures, ensures the stable and reliable operation of building equipment, and comprehensively enhances the level of intelligence in building equipment management. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a machine learning-based method for identifying anomalies in building equipment, as provided in an embodiment of this application.

[0018] Figure 2 This is a schematic diagram of the structure of a building equipment anomaly identification system provided in an embodiment of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this application. Obviously, the described embodiments are only some embodiments of the technical solutions of this application, and not all embodiments. Based on the embodiments recorded in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the technical solutions of this application.

[0020] See Figure 1 This is a machine learning-based method for identifying abnormal building equipment provided in the embodiments of this application. This method can be applied to a building equipment abnormality identification system. The specific process is as follows: steps 110-150.

[0021] Step 110: Obtain the continuous operation data set of the target building equipment, wherein the continuous operation data set includes multiple equipment status record units with timestamps.

[0022] This application embodiment uses a chiller unit in the air conditioning system of a large commercial center as the target building equipment. This chiller unit operates 24 hours a day, providing a suitable temperature environment for the commercial center. To obtain its continuous operational data set, data acquisition devices are deployed at key components and sensors of the chiller unit. These devices collect various operating parameters of the chiller unit in real time, such as the compressor's operating current, condenser temperature, and refrigerant pressure. At regular time intervals (e.g., every minute), the acquisition devices organize the collected equipment status parameters and attach a current timestamp, forming a unit of equipment status record. Over time, these timestamped equipment status record units accumulate, constituting the continuous operational data set of the target building equipment.

[0023] Step 120: Perform time-frequency domain feature extraction processing on the continuous operation data set to obtain the time-frequency domain feature set of the device status recording unit.

[0024] After acquiring the continuous operating data set of the refrigeration unit, time-frequency domain feature extraction processing is required to gain a deeper understanding of the equipment's operating characteristics. This processing allows for the extraction of key information from both the time and frequency domains, providing a more comprehensive description of the equipment's operating status. Time-domain features reflect the changing patterns of equipment state parameters over time, while frequency-domain features reveal the distribution of different frequency components in the signal. Combining these two approaches provides a more accurate overall picture of the equipment's operating status, offering strong support for accurate anomaly identification.

[0025] In an optional embodiment, the step of performing time-frequency domain feature extraction processing on the continuously running data set to obtain the time-frequency domain feature set of the device status recording unit includes:

[0026] Step 121: Perform time window division processing on the continuous running data set to obtain multiple data segment units with continuous time series relationship, and each data segment unit corresponds to a device status record of fixed duration.

[0027] For continuous operating data sets of refrigeration units, time windows are divided according to a fixed duration (e.g., 1 hour). Starting from the start time of the data set, each 1-hour period serves as a time window, dividing the data into multiple data segments with a continuous time sequence relationship. Thus, each data segment contains the equipment status records of the refrigeration unit within that 1-hour period. For example, the first data segment records the changes in parameters such as compressor operating current, condenser temperature, and refrigerant pressure from 0:00 to 1:00; the second data segment corresponds to the equipment operating status from 1:00 to 2:00. This time window division process allows continuous operating data to be segmented into smaller, easier-to-process and analyze segments while preserving the temporal order of the data.

[0028] Step 122: Perform time-domain feature extraction processing on each data segmentation unit, and extract the fluctuation amplitude features, change rate features, and periodic fluctuation features of the device state parameters in the data segmentation unit as a subset of time-domain features.

[0029] For each data segment unit, time-domain features are extracted from various equipment status parameters of the refrigeration unit. Taking the compressor operating current as an example, its fluctuation amplitude feature is calculated, i.e., the difference between the maximum and minimum values ​​of the compressor operating current in that data segment unit is identified; the rate of change feature is calculated by the ratio of the difference in compressor operating current at adjacent time points to the time interval; autocorrelation analysis is performed on the compressor operating current sequence to identify the periodicity of recurring patterns in the sequence, thus obtaining the periodic fluctuation feature. Similarly, similar processing is performed on other equipment status parameters such as condenser temperature and refrigerant pressure. Combining these fluctuation amplitude features, rate of change features, and periodic fluctuation features extracted for different equipment status parameters constitutes the time-domain feature subset for each data segment unit. This time-domain feature subset can comprehensively reflect the changing characteristics of the equipment status parameters of the refrigeration unit within the time period corresponding to each data segment unit.

[0030] In a preferred embodiment, the step of performing time-domain feature extraction processing on each data segmentation unit, and extracting the fluctuation amplitude features, change rate features, and periodic fluctuation features of the device state parameters in the data segmentation unit as a time-domain feature subset, includes:

[0031] Step 1221: Calculate the difference between the maximum and minimum values ​​of the device status parameters in the data segmentation unit to obtain the fluctuation amplitude characteristics of the data segmentation unit; calculate the ratio of the difference between the device status parameters at adjacent time points in the data segmentation unit to the time interval to obtain the rate of change characteristics of the data segmentation unit; perform autocorrelation analysis on the device status parameter sequence of the data segmentation unit to identify the periodicity of the recurring patterns in the sequence to obtain the periodic fluctuation characteristics of the data segmentation unit.

[0032] When calculating the fluctuation range characteristic of the compressor operating current of a refrigeration unit, all recorded compressor operating current values ​​within the data segment unit are iterated to find the maximum and minimum values, and then the difference between them is calculated. For example, in a 1-hour data segment unit, the maximum value of the compressor operating current is 50 amps and the minimum value is 30 amps, then the fluctuation range characteristic of the compressor operating current in this data segment unit is 50-30=20 amps. For the rate of change characteristic, the ratio of the difference in compressor operating current at adjacent time points to the time interval is calculated. Wherein, the compressor operating current recorded at two adjacent time points is 40 amps and 42 amps respectively, and the time interval is 1 minute (60 seconds), then the rate of change characteristic is (42-40) / 60=1 / 30 amps per second.

[0033] When extracting periodic fluctuation features, autocorrelation analysis is performed on the compressor operating current sequence within the data segment unit. The autocorrelation analysis algorithm is used to determine the period length of the recurring patterns in the sequence. For example, if the analysis reveals a similar change pattern in the compressor operating current every 10 minutes, then the periodic fluctuation feature of the compressor operating current in that data segment unit is 10 minutes. Other equipment status parameters, such as condenser temperature and refrigerant pressure, are calculated in the same way to obtain their corresponding fluctuation amplitude, rate of change, and periodic fluctuation features.

[0034] Step 1222: Normalize the fluctuation amplitude feature, the rate of change feature, and the periodic fluctuation feature to eliminate the dimensional differences between different parameters; combine the normalized fluctuation amplitude feature, rate of change feature, and periodic fluctuation feature into a time-domain feature subset of the data segmentation unit.

[0035] Because the fluctuation amplitude, rate of change, and periodic fluctuation characteristics have different dimensions, normalization is required for subsequent analysis and processing. A normalization algorithm (e.g., max-min normalization) is used to map these characteristic values ​​to a unified interval (e.g., [0, 1]). For the fluctuation amplitude characteristic of the compressor operating current, its original value is 20 amperes. The maximum value of the fluctuation amplitude characteristic within this data segment unit is 30 amperes, and the minimum value is 10 amperes. After processing with the max-min normalization algorithm, the normalized fluctuation amplitude characteristic value is (20-10) / (30-10) = 0.5. Similarly, a similar normalization process is performed on the rate of change and periodic fluctuation characteristics. Combining the normalized fluctuation amplitude, rate of change, and periodic fluctuation characteristics in a certain order forms a time-domain feature subset for this data segment unit. For example, arranging the normalized compressor operating current fluctuation amplitude, rate of change, and periodic fluctuation characteristics sequentially forms a vector, which is part of the time-domain feature subset. The corresponding features of other device status parameters are also combined in the same way to ultimately form a complete temporal feature subset of the data segmentation unit.

[0036] Step 123: Perform frequency domain conversion processing on each data segment unit to convert the time series signal of the data segment unit into a frequency domain signal representation.

[0037] To further analyze the frequency domain characteristics of the refrigeration unit's operating data, frequency domain transformation is required for each data segment. This transformation reveals the hidden frequency information within the time-series signal, allowing us to observe the distribution of different frequency components during equipment operation. Different equipment faults may generate abnormal signals within certain frequency ranges; frequency domain analysis can capture these signals, providing additional clues for accurate anomaly identification.

[0038] As a preferred embodiment, the step of performing frequency domain transformation processing on each data segment unit, converting the time-series signal of the data segment unit into a frequency domain signal representation, includes:

[0039] Step 1231: Window the time series signal of the data segmentation unit to suppress the spectral leakage at the signal edges and obtain the windowed time series signal.

[0040] Before performing frequency domain transformation on the time-series signal of a specific data segment of a refrigeration unit (e.g., the time-series of compressor operating current), windowing is first applied. A suitable window function (e.g., the Hanning window) is selected and multiplied with the original time-series signal. The purpose of the window function is to truncate the signal, allowing the portion within the window to transition to zero more smoothly, thereby suppressing spectral leakage at signal edges. In this case, the original compressor operating current time-series signal is a sequence of length N. Multiplying the Hanning window function point-by-point with this sequence yields the windowed time-series signal. This windowing process improves the accuracy of the subsequent frequency domain transformation and reduces the impact of spectral leakage on the frequency domain analysis results.

[0041] Step 1232: Process the windowed time series signal based on the Fast Fourier Transform algorithm to convert the time domain signal into a frequency domain signal; extract the real and imaginary parts of the frequency domain signal to generate a frequency distribution list containing frequency values ​​and corresponding amplitudes.

[0042] Optionally, the Fast Fourier Transform (FFT) algorithm can efficiently convert a time-domain signal into a frequency-domain signal, revealing the distribution of different frequency components in the signal. After processing by the FFT algorithm, a frequency-domain signal is obtained. For this frequency-domain signal, its real and imaginary parts are extracted. The real and imaginary parts can completely describe the characteristics of the frequency-domain signal. Based on the real and imaginary parts, the amplitude corresponding to each frequency is calculated, and the amplitude represents the intensity of that frequency component in the signal. The frequency values ​​and their corresponding amplitudes are compiled into a list, forming a frequency distribution list containing the frequency values ​​and their corresponding amplitudes. Through this frequency distribution list, the energy distribution of the compressor's operating current signal at different frequencies can be intuitively understood.

[0043] Step 1233: Perform smoothing filtering on the frequency distribution list to eliminate high-frequency noise interference and obtain a smooth frequency domain signal representation; associate and store the smooth frequency domain signal representation with the timestamp information of the data segmentation unit to form a frequency domain signal representation with time-frequency correspondence.

[0044] To eliminate high-frequency noise interference in the frequency distribution list, a smoothing filter is applied. For example, a Gaussian filter algorithm is used to perform a weighted average on each data point in the frequency distribution list, making the changes between adjacent data points smoother and effectively suppressing high-frequency noise. After Gaussian filtering, a smoothed frequency domain signal representation is obtained. For instance, if there are some data points with large fluctuations in a certain high-frequency band in the original frequency distribution list, these fluctuations are smoothed out after Gaussian filtering, making the data more stable. Next, the smoothed frequency domain signal representation is associated with the timestamp information of the data segmentation unit. Each frequency value and amplitude information in the frequency domain signal representation is bound to the timestamp corresponding to the data segmentation unit, forming a data structure with a time-frequency correspondence. In subsequent analysis, the corresponding frequency domain signal can be quickly located based on the timestamp information, facilitating the analysis and comparison of the frequency domain characteristics of the device operation in different time periods.

[0045] Step 124: Perform energy distribution analysis on the frequency domain signal representation, and extract the energy proportion characteristics, main frequency component characteristics and frequency component stability characteristics of different frequency intervals as frequency domain feature subsets.

[0046] Energy distribution analysis is performed on the frequency domain signal representation with time-frequency correspondence obtained after processing. First, the entire frequency range is divided into multiple frequency intervals, such as 0-10Hz, 10-20Hz, and 20-30Hz. The energy proportion characteristic within each frequency interval is calculated by summing the squares of the amplitudes corresponding to all frequency values ​​within that interval, and then comparing the sum of the energy in each interval with the total energy to obtain the energy proportion of that interval. For example, in the 0-10Hz frequency interval, the sum of the squares of the amplitudes corresponding to all frequency values ​​is E1, and the total energy is E; therefore, the energy proportion of this frequency interval is E1 / E. Next, the frequency interval with the largest energy proportion is identified; the frequency components within this interval are the dominant frequency components. The stability characteristics of the frequency components are determined by analyzing the changes in the amplitude of each frequency component in the frequency distribution list over time. If the amplitude of a certain frequency component fluctuates little over a period of time, it indicates that the frequency component has high stability. By combining the energy proportion characteristics, main frequency component characteristics, and frequency component stability characteristics of these different frequency ranges, a frequency domain feature subset is formed. These frequency domain features can reflect the energy distribution, main frequency components, and stability of the frequency components of the refrigeration unit's operating signal in different frequency ranges, providing important frequency domain information for anomaly detection.

[0047] Step 125: Perform association matching processing between the time-domain feature subset and the features of the corresponding data segmentation units in the frequency-domain feature subset to generate a time-frequency domain feature set for each device status recording unit.

[0048] Optionally, the time-domain feature subset and frequency-domain feature subset of each data segment unit are correlated and matched. Since both the time-domain and frequency-domain feature subsets are extracted for the same data segment unit, they are inherently related. The fluctuation amplitude, rate of change, and periodic fluctuation features in the time-domain feature subset are combined with the energy proportion features, dominant frequency component features, and frequency component stability features of different frequency intervals in the frequency-domain feature subset according to certain rules. For example, the fluctuation amplitude features can be correlated with the energy proportion features, the rate of change features with the dominant frequency component features, and the periodic fluctuation features with the frequency component stability features. Through the above correlation and matching process, a time-frequency domain feature set for each equipment status recording unit is generated. This time-frequency domain feature set integrates information from both the time and frequency domains, and can more comprehensively and accurately describe the operating status characteristics of the refrigeration unit within the time period corresponding to each data segment unit.

[0049] Step 130: Call the pre-built hybrid machine learning model to perform anomaly detection processing on the time-frequency domain feature set, and generate the anomaly identification result of the device status recording unit.

[0050] Optionally, the time-frequency domain feature set of the generated refrigeration unit equipment status recording unit is input into a pre-built hybrid machine learning model. This model combines the advantages of multiple machine learning algorithms, enabling more accurate analysis and processing of the time-frequency domain feature set, thereby identifying anomalies. The hybrid machine learning model has mastered the characteristic patterns of normal equipment operation through learning and training on a large amount of historical data. When a new time-frequency domain feature set is input, the model compares it with the learned normal patterns to determine if any anomalies exist. In this way, potential problems during the operation of the refrigeration unit can be detected in a timely manner.

[0051] In a preferred design approach, the step of calling a pre-built hybrid machine learning model to perform anomaly detection processing on the time-frequency domain feature set and generating anomaly identification results for the device status recording unit includes:

[0052] Step 131: Input the time-frequency domain feature set into the LSTM neural network of the hybrid machine learning model to model the time-series dependency of the time-frequency domain feature set and generate a sequence feature representation with time-series correlation information.

[0053] First, the time-frequency domain feature set of the refrigeration unit is input into the LSTM neural network of the hybrid machine learning model. LSTM neural networks excel at processing time-series data and can capture the temporal dependencies between features at different times in the time-frequency domain feature set. Before input, the time-frequency domain feature set undergoes necessary preprocessing to ensure that the data format and range meet the requirements of the LSTM neural network. For example, each feature in the feature set is normalized to ensure that its value range is within the appropriate interval. Then, the time-frequency domain feature set is input into the input layer of the LSTM neural network in timestamp order. The LSTM neural network processes the input time-frequency domain feature set through its internal memory units, forget gates, and output gates. The memory units retain the feature information of the previous equipment state record units, the forget gate filters out noisy historical feature information, and the output gate fuses the filtered historical information with the current feature input vector to generate a sequence feature representation with temporal correlation information. This sequence feature representation not only contains the time-frequency domain feature information of the current time but also incorporates relevant information from previous time points, which can better reflect the changing trend of the refrigeration unit's operating status over time.

[0054] In one implementation, the step of inputting the time-frequency domain feature set into the LSTM neural network of the hybrid machine learning model to model the time-series dependency of the time-frequency domain feature set and generate a sequence feature representation with temporal correlation information includes:

[0055] Step 1311: Perform feature dimension alignment processing on the time-frequency domain feature set to unify the number of feature dimensions of different device state recording units.

[0056] Because the number of feature dimensions may differ among different device status recording units in the time-frequency domain feature set of a refrigeration unit, feature dimension alignment is required before inputting them into an LSTM neural network. The time-frequency domain feature set of each device status recording unit is checked to determine the maximum number of feature dimensions. Then, for device status recording units with insufficient feature dimensions, padding or expansion is used to match the maximum number of feature dimensions. For example, if a device status recording unit has only 5 feature dimensions in its time-frequency domain feature set, while other units have a maximum of 8, its feature dimensions can be expanded to 8 by adding default values ​​(e.g., 0) to the feature set or by reasonably expanding existing features. This feature dimension alignment ensures that the time-frequency domain feature sets of different device status recording units have a uniform format when input into the LSTM neural network, facilitating network processing and learning.

[0057] Step 1312: Input the dimension-aligned time-frequency domain feature set into the input layer of the LSTM neural network in timestamp order to generate the initial feature input vector.

[0058] The time-frequency domain feature set of the refrigeration unit, after feature dimension alignment, is sequentially input into the input layer of the LSTM neural network according to the timestamp order. The time-frequency domain feature set of each device status recording unit is input as a whole into the neurons of the input layer. The neurons of the input layer perform preliminary processing and transformation on these input features, converting them into an initial feature input vector suitable for the internal processing of the LSTM neural network. For example, the input layer may perform operations such as weighted summation on the input features, combining multiple features into a vector form of the initial feature input vector.

[0059] Step 1313: The initial feature input vector is processed by the memory unit of the LSTM neural network to retain the feature information of the previous device state recording unit, and the historical information is obtained.

[0060] It's understandable that the memory unit of an LSTM neural network comprehensively processes the initial feature input vector and historical information retained from previous time steps. For the refrigeration unit's operating data, the memory unit remembers the time-frequency domain feature information from previous equipment state recording units. For example, when processing the initial feature input vector related to the compressor's operating current at the current moment, the memory unit recalls that parameter and other related time-frequency domain feature information from the previous moment. Through this method, the memory unit accumulates a series of historical information about the refrigeration unit's operating state. This information not only includes the changes in the current parameters but also covers the performance of other related features at different times. This allows the network to capture the evolution of the equipment's operating state over time. This historical information memory processing mechanism enables the LSTM neural network to fully utilize past experience to understand the current situation when facing complex time-series data, thereby better processing and analyzing the time-series dependencies in the refrigeration unit's operating data.

[0061] Step 1314: The historical information is filtered through the forget gate of the LSTM neural network to remove noisy historical feature information and obtain the filtered historical information.

[0062] Optionally, the forget gate in the LSTM neural network evaluates and filters the historical information stored in the memory unit, determining which historical features are noise and which are valuable for current analysis. For the refrigeration unit's operating data, the forget gate analyzes the correlation between various features in the historical information and the current equipment status. For example, if the condenser temperature fluctuates abnormally at a certain historical moment, but subsequent data indicates that this is just a random noise event, the forget gate will reduce its focus on this historical information or even filter it out. Through the above filtering process, the forget gate can remove noisy historical features that may interfere with the network's judgment, making the historical information retained in the memory unit more refined and valuable. The filtered historical information obtained can more accurately reflect the long-term trend and important characteristics of the refrigeration unit's operating status, helping to improve the accuracy of the network's analysis of the current equipment status.

[0063] Step 1315: The filtered historical information and the current feature input vector are fused through the output gate of the LSTM neural network to generate a sequence feature representation containing temporal correlation information.

[0064] In detail, the output gate is responsible for fusing the filtered historical information with the current feature input vector. In the example of the refrigeration unit, the output gate combines the historical information, such as compressor operating current, condenser temperature, and refrigerant pressure, filtered by the forget gate, with the initial feature input vectors of these parameters at the current moment. The output gate performs weighted summation and other operations on the historical information and current features according to certain weighting rules, ensuring that the generated sequence feature representation includes both information about the current equipment state and memories of past related states. For example, the output gate may dynamically adjust the weights of historical information and current features based on factors such as the stability of the current equipment operation. If the current equipment operating state is relatively stable, it may rely more on historical information to generate the sequence feature representation; if some abnormal changes occur, the weight of the current feature input vector will be increased. Through the above fusion process, the final generated sequence feature representation, containing temporal correlation information, can comprehensively reflect the connections and changes between the operating states of the refrigeration unit at different times.

[0065] Step 132: Input the sequence feature representation into the isolated forest algorithm of the hybrid machine learning model to perform outlier detection processing on the sequence feature representation and identify abnormal feature points in the sequence feature representation that deviate from the normal pattern.

[0066] In this step, the sequence feature representation containing temporal correlation information generated by the LSTM neural network is input into the Isolation Forest algorithm, a hybrid machine learning model. The Isolation Forest algorithm is specifically designed for outlier detection, capable of quickly and effectively identifying data points that deviate from normal patterns in high-dimensional data spaces. For the sequence feature representation of a refrigeration unit, the Isolation Forest algorithm constructs multiple decision trees, each randomly partitioning the data points in the sequence feature representation. During partitioning, normal data points are typically partitioned at lower levels of the tree, while outliers are partitioned to leaf nodes more quickly. Through this method, the Isolation Forest algorithm calculates the isolation score for each data point; a higher isolation score indicates a greater likelihood of the data point being an outlier. For example, for a data point in the sequence feature representation related to the compressor's operating current, if this data point is quickly partitioned to a leaf node in the decision tree constructed by the Isolation Forest algorithm, and its isolation score is significantly higher than other data points, then this data point is identified as an anomalous feature point that may deviate from the normal pattern. Through the outlier detection process described above, we can identify those points that differ significantly from the normal operating mode from the sequence feature representation, providing crucial clues for subsequent judgment on whether the equipment is abnormal.

[0067] Step 133: Perform contextual analysis on the abnormal feature points, and combine the feature information of adjacent device status record units in the time-frequency domain feature set to determine whether the abnormal feature points are persistent or sporadic.

[0068] For identified anomalous feature points, further analysis is needed to determine their nature, classifying them as either persistent or intermittent anomalies. This requires contextual analysis, combining the feature information of adjacent equipment status recording units in the time-frequency domain feature set. Taking an anomalous feature point in a refrigeration unit as an example, this anomalous feature point is a large fluctuation in the compressor's operating current. First, examine the feature information of the compressor's operating current and other relevant equipment status parameters (such as condenser temperature, refrigerant pressure, etc.) in the adjacent equipment status recording units before and after this anomalous feature point. If similar anomalous fluctuations in the compressor's operating current are consistently observed in multiple adjacent equipment status recording units, and other related parameters also show corresponding anomalous changes, then this anomalous feature point can be determined as a persistent anomaly. This may indicate a systemic fault or problem with the refrigeration unit's compressor. Conversely, if only this anomalous feature point shows a significant anomaly, while the relevant parameters in adjacent equipment status recording units remain normal, then this anomalous feature point is likely an intermittent anomaly, possibly caused by brief external interference or momentary equipment instability. Through the above contextual analysis, the nature of the anomalous feature point can be determined more accurately.

[0069] Step 134: If so, the degree of abnormality of the abnormal feature points is quantitatively evaluated to generate an abnormal feature descriptor containing the abnormality confidence.

[0070] When an anomaly is determined to be persistent, its degree of anomaly needs to be quantitatively assessed to more accurately describe and analyze the equipment's abnormal condition. An appropriate quantitative assessment algorithm (such as one based on statistical analysis and machine learning) is used, combining historical data and current time-frequency domain feature information, to evaluate the degree of anomaly. For anomalies in refrigeration units, such as abnormal fluctuations in compressor operating current, the assessment algorithm considers factors such as the degree of deviation from the normal operating range, the duration of the fluctuation, and changes in related parameters. For example, if the abnormal fluctuation value of the compressor operating current far exceeds the normal operating range and lasts for a long time, while the related condenser temperature and refrigerant pressure also show significant abnormal changes, the assessment algorithm will give a high quantified value for the degree of anomaly. Based on this quantified value, an anomaly feature descriptor containing anomaly confidence is generated. The anomaly confidence indicates the credibility of the judgment that the anomaly feature is indeed abnormal. For example, an anomaly feature descriptor is obtained through evaluation, which contains an anomaly severity quantification value of 80 (ranging from 0 to 100) and an anomaly confidence level of 90%. This indicates that there is a 90% certainty that the anomaly feature point represents a real anomaly in the refrigeration unit and that the anomaly severity is high. The above anomaly feature descriptor can clearly convey the severity and credibility information of the anomaly.

[0071] Step 135: Associate and map the abnormal feature descriptor with the timestamp information of the device status recording unit to generate the abnormal identification result of the device status recording unit.

[0072] It is understandable that the generated anomaly feature descriptors, which include anomaly confidence levels, are associated and mapped with the timestamp information of the equipment status recording units. For each equipment status recording unit of the refrigeration unit, if an anomaly feature point exists and a corresponding anomaly feature descriptor is generated, then the descriptor is bound to the timestamp corresponding to that equipment status recording unit. For example, if the timestamp corresponding to a certain equipment status recording unit is 10:00 AM, and its anomaly feature descriptor shows an anomaly severity quantification value of 75 and an anomaly confidence level of 85%, then these two pieces of information are associated. Through the above association mapping process, anomaly identification results for the equipment status recording units are generated. These anomaly identification results not only contain information about whether the equipment is abnormal, but also clarify the time of the anomaly occurrence, as well as the degree and confidence level of the anomaly. For example, equipment managers can quickly locate the time of the anomaly occurrence based on the anomaly identification results, and, combined with the anomaly severity and confidence level information, decide what measures to take to handle the abnormal situation.

[0073] Step 140: Determine the anomaly type of the target building equipment and the distribution characteristics of the anomaly type in the time dimension based on the anomaly identification results.

[0074] Optionally, based on the anomaly identification results of the generated refrigeration unit equipment status record unit, further analysis can be performed to determine the anomaly type of the target building equipment and the distribution characteristics of that anomaly type over time. This step can provide more targeted guidance for equipment maintenance and management. Through detailed analysis of the anomaly identification results, different anomalies can be classified and their respective anomaly types determined. Simultaneously, by statistically analyzing anomalies occurring at different time points, the distribution characteristics of anomaly types over time can be understood, thereby enabling proactive preventative and response measures.

[0075] As one implementation, determining the anomaly type of the target building equipment and the distribution characteristics of the anomaly type over time based on the anomaly identification result includes:

[0076] Step 141: Parse the abnormal feature descriptors in the abnormality identification results and extract the time-frequency domain feature patterns corresponding to the abnormal feature descriptors.

[0077] A detailed analysis of the anomaly feature descriptors in the anomaly identification results of the refrigeration unit is performed. Taking a certain anomaly feature descriptor as an example, it includes a quantified value of the anomaly severity, anomaly confidence level, and some feature information related to the anomaly. By analyzing this information, the time-frequency domain feature pattern corresponding to the anomaly feature descriptor is extracted. For example, the anomaly feature descriptor mentions abnormal fluctuations in the compressor operating current. Combining the relevant time-frequency domain feature information, it is found that this abnormal fluctuation manifests as a rapid rise and fall in the current value in the time domain, and as an abnormal increase in energy within a specific frequency range in the frequency domain. Combining these time-domain and frequency-domain features forms the time-frequency domain feature pattern corresponding to the anomaly feature descriptor. This time-frequency domain feature pattern can accurately characterize the manifestation of the anomaly in the time-frequency domain.

[0078] Step 142: Match the time-frequency domain feature pattern with a preset anomaly type feature library to determine the anomaly type corresponding to the anomaly feature descriptor.

[0079] The extracted time-frequency domain feature patterns are compared and matched with a pre-defined anomaly type feature library. This library stores time-frequency domain feature patterns for various known anomaly types. For refrigeration units, the library may contain feature patterns for different anomaly types such as compressor failure, condenser failure, and refrigerant leakage. The similarity between the currently extracted time-frequency domain feature pattern and each anomaly type feature pattern in the library is calculated. For example, a similarity matching algorithm (such as cosine similarity algorithm) is used to calculate the similarity score between the current time-frequency domain feature pattern and a certain anomaly type feature pattern in the library. If the similarity score of an anomaly type feature pattern exceeds a pre-defined threshold (e.g., 0.8), the anomaly type corresponding to that anomaly feature descriptor is determined to be the matched anomaly type. For example, if the matching reveals that the current time-frequency domain feature pattern has a high similarity to the compressor failure feature pattern, exceeding the threshold, then the anomaly type corresponding to that anomaly feature descriptor can be determined to be compressor failure. Through this matching process, anomalies can be quickly and accurately classified into known anomaly types.

[0080] Step 143: Extract the timestamp information of the anomaly feature descriptors in the anomaly identification results, and count the frequency of occurrence of each anomaly type in different time intervals.

[0081] The timestamp information corresponding to each anomaly feature descriptor is extracted from the anomaly identification results of the refrigeration unit. Then, the time is divided according to a certain time interval (e.g., by day, by hour, etc.). Taking the time interval divided by day as an example, the frequency of occurrence of various anomaly types within each day is statistically analyzed. For each day, all anomaly identification results are iterated to find all anomaly feature descriptors that appear within that day, and they are classified and statistically analyzed according to their corresponding anomaly types. For example, on a certain day, it is found that 3 anomaly feature descriptors correspond to the anomaly type of compressor failure, and 2 anomaly feature descriptors correspond to the anomaly type of condenser failure. Through the above statistical method, the frequency of occurrence of different anomaly types in different time intervals can be clearly understood.

[0082] Step 144: Perform time series analysis on the occurrence frequency of the anomaly type to identify the active and inactive periods of the anomaly type in the time dimension.

[0083] Optionally, time series analysis algorithms (such as moving average algorithms, autoregressive moving average algorithms, etc.) are used to analyze and process the time series of occurrence frequencies of different anomaly types obtained statistically. Taking compressor failure anomaly type as an example, the average occurrence frequency within each time window is calculated using the moving average algorithm. The time window is 3 days, and the 3-day moving average of the daily compressor failure anomaly occurrence frequency is calculated. Then, the difference in average occurrence frequency between adjacent time windows is compared. If the average occurrence frequency of a certain time window is significantly higher than the previous time window and exceeds a certain threshold (e.g., 1.5 times the average), then that time point is identified as the starting point of an active period; if the average occurrence frequency is significantly lower than the previous time window and falls below a certain threshold (e.g., 0.5 times the average), then that time point is identified as the ending point of an active period. In inactive periods, time intervals with occurrence frequencies lower than a preset frequency (e.g., less than once per day) are defined as inactive periods. These initially determined active and inactive periods are then smoothed using methods such as linear interpolation to eliminate time period division errors caused by random fluctuations, resulting in smoothed active and inactive periods. The smoothed active and inactive periods are associated with the compressor fault anomaly type and stored to form the active and inactive period information of that anomaly type in the time dimension.

[0084] As one implementation, the step of performing time series analysis on the frequency of occurrence of the anomaly type to identify active and inactive periods of the anomaly type in the time dimension includes:

[0085] Step 1440: Perform sliding window statistical processing on the occurrence frequency time series of the anomaly type, calculate the average occurrence frequency within each time window; compare the difference in average occurrence frequency between adjacent time windows, identify the time point of sudden increase in occurrence frequency as the starting point of the active period; identify the time point of sudden decrease in occurrence frequency as the ending point of the active period; within the inactive period, statistically define the time interval with occurrence frequency lower than a preset frequency as the inactive period; perform boundary smoothing processing on the active and inactive periods to eliminate the time period division error caused by accidental fluctuations, and obtain smoothed active and inactive periods; associate and store the smoothed active and inactive periods with the anomaly type to form the time distribution feature information of the anomaly type.

[0086] For a time series of occurrence frequencies of a certain type of anomaly in a refrigeration unit (such as condenser failure), a sliding window size (e.g., 5 hours) is set. Starting from the beginning of the time series, the average occurrence frequency within each 5-hour time window is calculated sequentially. For example, if condenser failure occurs twice in the first 5-hour time window, the average occurrence frequency is 2 / 5 = 0.4 times / hour. Then, the average occurrence frequency of the current time window is compared with the average occurrence frequency of the adjacent previous time window. If the average occurrence frequency of the current time window is 0.6 times / hour, and the average occurrence frequency of the previous time window is 0.2 times / hour, and this increase exceeds a preset threshold (e.g., 0.3 times / hour), then the starting point of the current time window is identified as the start point of an active period. Conversely, if the average occurrence frequency of the current time window is 0.1 times / hour, and the average occurrence frequency of the previous time window is 0.4 times / hour, and the decrease exceeds a preset threshold (e.g., 0.2 times / hour), then the ending point of the current time window is identified as the end point of an active period. During inactive periods, a frequency threshold is preset (e.g., 0.1 times / hour), and time intervals where the frequency of occurrence is below this threshold are statistically analyzed. For example, if the frequency of condenser failures consistently falls below 0.1 times / hour during a certain time period, then that time period is considered inactive. Since actual data may exhibit occasional fluctuations, the boundaries between the initially determined active and inactive periods are smoothed. Linear interpolation can be used to adjust the boundaries between active and inactive periods, making them more accurate and reasonable. Finally, the smoothed active and inactive periods are associated with the condenser failure anomaly type and stored to form the temporal distribution characteristic information for that anomaly type. This temporal distribution characteristic information clearly shows the activity and inactivity of condenser failures in different time periods, facilitating advance maintenance and monitoring planning.

[0087] Step 145: Construct the time distribution curve of the anomaly type based on the active and inactive time periods, and generate time distribution feature information containing time intervals and corresponding frequencies.

[0088] Understandably, based on the previously determined active and inactive time periods for various anomaly types in the refrigeration unit, a time distribution curve is constructed for each anomaly type. The horizontal axis represents time, and the vertical axis represents the frequency of occurrence of the anomaly type. For each anomaly type, active and inactive time periods are marked on the time axis, and curves are plotted based on the frequency of occurrence within different time intervals. For example, for the compressor failure anomaly type, its active time period (e.g., 10:00 AM - 2:00 PM) and inactive time period (e.g., 8:00 PM - 6:00 AM) are clearly marked on the time axis. Within the active time period, based on the statistically obtained frequency data, the corresponding frequency value is found on the vertical axis, and these points are connected to form part of the curve; within the inactive time period, due to the lower frequency of occurrence, the curve will be closer to the horizontal axis.

[0089] The above method generates time distribution feature information containing time intervals and corresponding frequencies. The time distribution curve can intuitively show the frequency changes of each anomaly type at different times. Equipment managers can quickly understand the time distribution pattern of anomaly types by observing the curve, providing a basis for formulating more effective maintenance plans and monitoring strategies.

[0090] Step 150: Generate a target early warning instruction containing a time location identifier based on the anomaly type and the time distribution feature information, and send the device early warning instruction to the target device management terminal.

[0091] Based on the identified anomaly type of the refrigeration unit and its distribution characteristics over time, a target early warning instruction containing a time location identifier is generated to promptly notify the equipment management terminal to take appropriate measures. This target early warning instruction must not only include information about the anomaly type but also specify the timeframe in which the anomaly may occur, thus helping equipment managers to handle the situation effectively. By sending the early warning instruction to the target equipment management terminal, rapid response and effective management of equipment anomalies can be achieved.

[0092] Optionally, generating a target early warning instruction containing a time location identifier based on the anomaly type and the time distribution feature information includes:

[0093] Step 151: Parse the preset early warning rule base corresponding to the anomaly type, and extract the early warning priority identifier and maintenance strategy code associated with the anomaly type.

[0094] For identified anomalies in the refrigeration unit, such as compressor failure, the system automatically parses a pre-defined warning rule base. This pre-built database stores detailed rule information corresponding to various anomaly types. Within this rule base, for the compressor failure anomaly, the system identifies the associated warning priority identifier and maintenance strategy code. The warning priority identifier indicates the urgency of the anomaly; for example, a compressor failure might be set as high priority because it could severely impact the normal operation of the refrigeration unit, thereby affecting the overall ambient temperature of the commercial center. The maintenance strategy code is a set of codes indicating the maintenance measures to be taken for this anomaly type. For example, the maintenance strategy code might contain a series of operational steps and requirements to guide maintenance personnel in performing effective repair and maintenance work. Through this method, crucial information related to the anomaly type is accurately extracted from the pre-defined warning rule base.

[0095] Step 152: Extract the active and inactive time periods from the time distribution feature information to determine the current active time interval of the anomaly type.

[0096] From the generated temporal distribution characteristics of refrigeration unit anomaly types, the active and inactive time periods for each anomaly type (such as compressor failure) are extracted. Then, combined with the current time, the current active time interval for this anomaly type is determined. For example, if the current time is 9:00 AM, and the active time for compressor failure is 10:00 AM to 2:00 PM, then the current active time interval is determined to be 10:00 AM to 2:00 PM. By clearly defining the current active time interval, equipment managers can clearly understand during which time periods the anomaly requires focused attention, allowing for advance preparation and scheduling of appropriate maintenance work.

[0097] Step 153: Calculate the probability of the anomaly type recurring in a specified subsequent time period based on the current active time interval.

[0098] Based on the currently determined active time interval, a corresponding probability calculation model is used to calculate the probability of the recurrence of this type of anomaly (such as compressor failure) in the refrigeration unit within a specified subsequent period. This probability calculation model can be established based on historical data statistical analysis, taking into account various factors such as the frequency of occurrence of this anomaly type in similar past time periods, the current operating status of the equipment, and environmental factors. For example, by statistically analyzing the occurrence of compressor failures after several similar active time periods in the past, it is found that if a failure occurs once in a certain active time period, the probability of it recurring in the next 24 hours is 30%. This basic probability is adjusted by combining the actual operating parameters of the refrigeration unit, maintenance records, and current ambient temperature. If the current compressor operating temperature is high and recent maintenance work has been delayed, the probability of the failure recurring in the specified subsequent period (such as the next 24 hours) may increase to 40% according to the model calculation. Through this method, the likelihood of the anomaly type recurring in the future can be estimated relatively accurately, providing a basis for formulating reasonable early warning and maintenance strategies.

[0099] Step 154: Associate and weight the probability of recurrence with the warning priority identifier of the anomaly type to generate dynamic warning level parameters.

[0100] The calculated probability of recurrence of a refrigeration unit anomaly type (such as compressor failure) within a specified subsequent time period is weighted and correlated with the corresponding warning priority identifier. The warning priority identifier represents the urgency of the anomaly, while the recurrence probability reflects the likelihood of the anomaly occurring in the future. A preset weighting algorithm combines these two factors to generate a dynamic warning level parameter. For example, for compressor failure, the warning priority identifier is high, the weight of high priority is 0.6, and the recurrence probability is 40% (i.e., 0.4). The weighting algorithm may specify that the dynamic warning level parameter = warning priority weight × recurrence probability + other adjustment factors (other adjustment factors are 0 in this embodiment), then the dynamic warning level parameter = 0.6 × 0.4 = 0.24 (the dynamic warning level parameter ranges from 0 to 1). This dynamic warning level parameter comprehensively reflects the urgency of the anomaly and the likelihood of its future occurrence, providing an important basis for generating more accurate and effective warning instructions. Through the above-mentioned weighted correlation processing, the warning level can be dynamically adjusted according to the actual situation, better reflecting the actual operating conditions of the equipment.

[0101] Step 155: Perform information fusion processing on the dynamic early warning level parameters, maintenance strategy codes, and current active time intervals to generate a target early warning instruction containing time location identifiers.

[0102] Optionally, the generated dynamic warning level parameter, maintenance strategy code, and current active time interval are fused together. Information fusion integrates these different but related pieces of information to form a complete, accurate, and clearly guiding target warning instruction. For example, the dynamic warning level parameter (e.g., 0.24), maintenance strategy code (e.g., a set of detailed maintenance operation step codes), and current active time interval (10:00 AM - 2:00 PM) can be combined. During this combination process, the information is arranged according to a specific format and rules, making the generated target warning instruction clear and concise. For example, the target warning instruction might be presented in text form: "Refrigeration unit compressor fault warning! Dynamic warning level: 0.24 (higher), current active time interval is 10:00 AM - 2:00 PM, please perform the corresponding maintenance operation according to the maintenance strategy code [corresponding code content]." Through the above information fusion, the generated target warning instruction includes information on the severity and probability of the abnormal situation, as well as clearly defining the time range of the abnormality and the corresponding maintenance measures, providing comprehensive and specific guidance for the equipment management terminal.

[0103] Preferably, the step of fusing the dynamic early warning level parameters, maintenance strategy codes, and current active time intervals to generate a target early warning instruction containing a time location identifier includes:

[0104] Step 1551: Assign a corresponding warning symbol to the dynamic warning level parameter, wherein the warning symbol is positively correlated with the warning level parameter.

[0105] Assign corresponding warning symbols to the dynamic warning level parameters (e.g., 0.24) generated by the refrigeration unit. Warning symbols are an intuitive representation used to quickly convey the severity of the warning. Generally, warning symbols are positively correlated with the warning level parameter; that is, the higher the warning level parameter, the more attention the corresponding warning symbol receives. For example, for warning level parameters between 0 and 0.3, the assigned warning symbol might be a "yellow exclamation mark"; between 0.3 and 0.6, it might be an "orange triangle"; and between 0.6 and 1, it might be a "red alert" icon. For a dynamic warning level parameter of 0.24, a "yellow exclamation mark" is assigned as the warning symbol. Therefore, when equipment managers receive a warning instruction, they can quickly understand the approximate severity of the warning by checking the warning symbol, without needing to examine the specific level parameter in detail, thus improving the efficiency of information acquisition and processing.

[0106] Step 1552: Encode the maintenance strategy into maintenance operation description text that includes maintenance steps.

[0107] Specifically, the step of encoding the maintenance strategy into maintenance operation description text containing maintenance step instructions includes:

[0108] Step 15521: Obtain the multi-segment coding structure of the maintenance strategy code, which consists of a maintenance action type identifier segment, an action object identifier segment, and an operation condition identifier segment arranged in sequence.

[0109] For the maintenance strategy coding of refrigeration unit compressor failures, the first step is to obtain its multi-segment coding structure. For example, the maintenance strategy code is "123-45-678", where "123" is the maintenance action type identifier segment, representing the category of maintenance action required for this anomaly, such as "repairing the compressor motor"; "45" is the action object identifier segment, specifying the specific equipment component targeted by the maintenance action, possibly corresponding to "compressor motor windings"; and "678" is the operating condition identifier segment, specifying the conditions that must be met when performing the maintenance operation, such as "the ambient temperature must be between 20-30 degrees Celsius, and the equipment must be in a stopped state". Through the above multi-segment coding structure, the key elements of the maintenance operation can be accurately defined, providing a clear framework for the subsequent accurate conversion into maintenance operation description text.

[0110] Step 15522: Call the pre-built maintenance action knowledge base, which stores basic step description units corresponding one-to-one with the maintenance action type identifier segment, equipment component positioning description units corresponding one-to-one with the action object identifier segment, and execution constraint description units corresponding one-to-one with the operation condition identifier segment.

[0111] Optionally, the system calls a pre-built maintenance action knowledge base, a rich database storing detailed information on various maintenance action types, action objects, and operating conditions. For the maintenance action type identifier segment "123" in the maintenance strategy code, the corresponding basic step description unit is found in the knowledge base. The basic step description unit contains standard action verbs and general operation direction information; for example, the basic step description unit for "repairing the compressor motor" might be "First, disconnect the power, then use specialized tools to disassemble the motor housing...". For the action object identifier segment "45", the corresponding equipment component positioning description unit is found. This unit contains the component name and spatial location information, such as "The compressor motor winding is located inside the compressor on the left side and is connected to other components through a set interface." For the operating condition identifier segment "678", the corresponding execution constraint description unit is obtained, which includes environmental parameter range requirements and safety precautions, such as "The ambient temperature must be maintained between 20-30 degrees Celsius during operation; be sure to wear insulating protective equipment to avoid the risk of electric shock." By calling the maintenance action knowledge base, detailed and accurate maintenance operation information can be obtained to generate a complete maintenance operation description text.

[0112] Step 15523: Perform segmented parsing processing on the multi-segmented coding structure to extract the encoded values ​​of the maintenance action type identifier segment, action object identifier segment, and operation condition identifier segment.

[0113] The maintenance strategy code "123-45-678" is segmented and parsed to extract the code values ​​"123" for the maintenance action type identifier, "45" for the action object identifier, and "678" for the operation condition identifier. This segmented parsing is to accurately retrieve relevant information from the maintenance action knowledge base and combine this information logically to generate a maintenance operation description text that meets actual needs. By clearly defining the values ​​of each code segment, the accuracy and effectiveness of subsequent information matching and text generation processes can be ensured.

[0114] Step 15524: Based on the encoded value of the maintenance action type identifier segment, the corresponding basic step description unit is obtained by matching from the maintenance action knowledge base. The basic step description unit includes standard action verbs and general operation direction information.

[0115] Based on the extracted maintenance action type identifier segment code value "123", a precise match is performed in the maintenance action knowledge base to obtain the corresponding basic step description unit. As mentioned earlier, this basic step description unit might be "First, disconnect the power, then use professional tools to disassemble the motor housing and check the motor windings for signs of damage..." This basic step description unit provides the basic flow and action direction for performing maintenance operations and is the core part of constructing a complete maintenance operation description text. It uses standard action verbs, such as "disconnect," "disassemble," and "check," clearly indicating the operations that maintenance personnel need to perform. At the same time, the general operation direction information also provides maintenance personnel with a general idea and sequence of operations.

[0116] Step 15525: Based on the encoded value of the action object identifier segment, obtain the corresponding equipment component positioning description unit from the maintenance action knowledge base. The equipment component positioning description unit includes the component name and spatial location information.

[0117] Based on the code value "45" of the action object identifier segment, the corresponding equipment component location description unit is located in the maintenance action knowledge base. This unit details the component name and spatial location information, such as "the compressor motor winding is located inside the compressor on the left side and is connected to other components through a set interface." This information is crucial for maintenance personnel, helping them quickly and accurately locate the specific equipment component requiring maintenance and avoiding misoperation or failure to find the target component during operation. By combining the equipment component location description unit with the basic step description unit, the maintenance operation description text becomes more specific and executable.

[0118] Step 15526: Based on the encoded value of the operation condition identifier segment, obtain the corresponding execution constraint description unit from the maintenance action knowledge base. The execution constraint description unit includes environmental parameter range requirements and safety precautions information.

[0119] Based on the coded value "678" of the operating condition identifier segment, the corresponding execution constraint description unit is retrieved from the maintenance action knowledge base. This unit contains environmental parameter range requirements and safety precautions, such as "The ambient temperature must be maintained between 20-30 degrees Celsius during operation. Wear appropriate insulating protective equipment to avoid the risk of electric shock." This information is crucial for ensuring the safe and effective execution of maintenance operations. Including the execution constraint description unit in the maintenance operation description text reminds maintenance personnel to pay attention to environmental conditions and safety issues during operation, preventing equipment damage or personal injury due to unmet conditions or neglect of safety precautions.

[0120] Step 15527: Perform semantic fusion processing on the standard action verbs of the basic step description unit and the component names and spatial location information of the device component positioning description unit to generate an action description clause containing the operation object.

[0121] The standard action verbs in the basic step description unit are fused with the component names and spatial location information in the equipment component location description unit. For example, if the basic step description unit includes the action "cut off the power," and the equipment component location description unit indicates that the power control switch is located on the right side inside the compressor control cabinet, then after semantic fusion processing, the generated action description clause containing the operation object might be "Locate the power control switch on the right side inside the compressor control cabinet and cut off the power." This fusion makes the action description more specific and clear, allowing maintenance personnel to clearly understand which specific component requires which operation, thus improving the accuracy and operability of the maintenance operation description text.

[0122] Step 15528: Perform condition association processing between the action description clause and the environmental parameter range requirements and safety precautions information of the execution constraint description unit to generate a step constraint description clause containing the operation prerequisites.

[0123] The generated action description clauses are conditionally associated with the information in the execution constraint description units. Taking the previously generated action description clause "Locate the power control switch on the right side inside the compressor control cabinet and disconnect the power" and the execution constraint description unit "The ambient temperature must be maintained between 20-30 degrees Celsius during operation; be sure to wear appropriate insulating protective equipment to avoid the risk of electric shock" as an example, the generated step constraint description clause containing the operational prerequisites is: "Under the premise that the ambient temperature is maintained between 20-30 degrees Celsius and appropriate insulating protective equipment is worn, locate the power control switch on the right side inside the compressor control cabinet and disconnect the power." This step constraint description clause clarifies the operational prerequisites, enabling maintenance personnel to fully consider various limiting factors when performing operations, ensuring the safety and correctness of the operation.

[0124] Step 15529: According to the arrangement order of the multi-segment encoding structure, logically concatenate the action description clause and the step constraint description clause to generate a maintenance step description sequence with sequential correlation; perform natural language fluency optimization processing on the maintenance step description sequence, adjust the grammatical structure of the statements and the use of conjunctions, and generate a maintenance operation description text containing complete operation logic.

[0125] Following the multi-segment structure of the maintenance strategy coding, the generated action description clauses and step constraint description clauses are logically linked. For example, other related action description clauses and step constraint description clauses, such as "After confirming damage to the motor windings, replace the motor windings using professional tools, and maintain a clean operating environment during the replacement process," are connected in a logical order to form a sequential maintenance step description sequence. Then, this maintenance step description sequence undergoes natural language fluency optimization. The grammatical structure of the statements is checked for correctness, and the use of conjunctions is examined for appropriateness. Adjustments are made to any unclear or fluent passages. For example, conjunctions such as "first...then...and..." are used appropriately to make the entire maintenance operation description text more fluent and understandable. Finally, a maintenance operation description text containing complete operational logic is generated. This text provides maintenance personnel with detailed, accurate, and easy-to-understand operational guidance, helping them successfully complete the maintenance work of the refrigeration unit.

[0126] Step 1553: Convert the current active time interval into a time range descriptor, which includes a start time point and an end time point.

[0127] This converts the current active time interval (e.g., 10:00 AM - 2:00 PM) of the refrigeration unit's abnormal type (such as compressor failure) into a time range descriptor. A time range descriptor is a more standardized and easier-to-understand way of representing time, explicitly including a start and end time. For example, converting 10:00 AM to "2024-01-01 10:00:00" (current date is January 1, 2024) and 2:00 PM to "2024-01-01 14:00:00" results in a time range descriptor of "2024-01-01 10:00:00 to 2024-01-01 14:00:00". This type of time range descriptor can more accurately and clearly convey the time information of abnormal activity when generating subsequent warning commands, avoiding misunderstandings or incorrect operations caused by ambiguous time representations.

[0128] Step 1554: Perform structured combination processing on the warning symbol, maintenance operation description text and time range descriptor to generate the main body of the equipment warning instruction containing the time positioning identifier.

[0129] The assigned warning symbols (such as a "yellow exclamation mark"), the generated maintenance operation description text (detailed maintenance steps for compressor failure), and the time range descriptor ("2024-01-01 10:00:00 to 2024-01-01 14:00:00") are structured and combined. Following a specific format and rules, this information is integrated to form a complete equipment warning instruction body. For example, the equipment warning instruction body might be presented in the following format: "[Yellow exclamation mark] Refrigeration unit compressor failure warning! Time range: 2024-01-01 10:00:00 to 2024-01-01 14:00:00. Maintenance operation: Under the premise that the ambient temperature is maintained between 20-30 degrees Celsius and insulated protective equipment is worn, locate the power control switch on the right side inside the compressor control cabinet and disconnect the power... (followed by detailed maintenance steps)." Through the above structured combination processing, the generated equipment early warning command can clearly and comprehensively convey abnormal information, time location, and maintenance operation requirements, providing strong support for equipment management.

[0130] Step 1555: Add device identification information to the device warning command body. The device identification information is used to uniquely identify the target building device.

[0131] Add equipment identification information to the generated equipment warning command body. Each refrigeration unit, as the target building equipment, has unique identification information, such as equipment number "L001" and equipment name "Refrigeration Unit 1, Building A, Commercial Center". Adding this equipment identification information to the equipment warning command body makes the warning command more specific and targeted. The added warning command might look like this: "Equipment Number: L001, Equipment Name: Refrigeration Unit 1, Building A, Commercial Center [Yellow Exclamation Mark] Refrigeration Unit Compressor Fault Warning! Time Range: 2024-01-01 10:00:00 to 2024-01-01 14:00:00. Maintenance Operation: Under the premise that the ambient temperature is maintained between 20-30 degrees Celsius and insulated protective equipment is worn, locate the power control switch on the right side inside the compressor control cabinet and disconnect the power... (detailed maintenance steps follow)". By adding equipment identification information, the equipment management terminal can accurately identify the specific equipment targeted by the warning command, avoiding misoperation or incorrect handling, improving the accuracy and effectiveness of the warning command, and helping equipment managers quickly locate and handle problematic equipment.

[0132] Step 1556: Perform verification and encoding processing on the main body of the device warning instruction and the device identification information to generate a target warning instruction with error verification function.

[0133] To ensure the accuracy and integrity of equipment warning commands during transmission and reception, the body of the warning command, which includes equipment identification information, undergoes verification encoding. This can be achieved based on a preset verification encoding algorithm (such as Cyclic Redundancy Check (CRC) algorithm). This algorithm generates a checksum based on the content of the warning command body and the equipment identification information. Taking a previously generated warning command as an example, all information, including the equipment number, equipment name, warning identifier, time range descriptor, and maintenance operation description text, is used as input. A checksum is then calculated using the CRC algorithm. For example, the calculated checksum might be "12AB". Then, the checksum is added to the end of the warning instruction to form a target warning instruction with error verification function: "Equipment No.: L001, Equipment Name: Refrigeration Unit 1, Block A, Commercial Center [Yellow Exclamation Mark] Refrigeration Unit Compressor Failure Warning! Time Range: 2024-01-01 10:00:00 to 2024-01-01 14:00:00. Maintenance Operation: Under the premise that the ambient temperature is maintained between 20-30 degrees Celsius and insulated protective equipment is worn, locate the power control switch on the right side inside the compressor control cabinet and disconnect the power supply... (Detailed maintenance steps follow) Checksum: 12AB". When the target equipment management terminal receives this warning instruction, it will use the same checksum algorithm to verify the received information. If the calculated checksum matches the received checksum, it means that no error occurred during the transmission of the warning instruction; if they do not match, it means that there may be data loss or error, and the equipment management terminal can request a resend of the warning instruction, thereby ensuring the accuracy and reliability of the warning instruction.

[0134] Step 156: Standardize the format of the device warning command to obtain a standardized warning command; the standardized warning command matches the communication protocol requirements of the target device management terminal.

[0135] The generated target warning commands with error checking functionality undergo format standardization to ensure they match the communication protocol requirements of the target device management terminal. Different device management terminals may have different communication protocol requirements, including specifications for data format, character encoding, message length, etc. For example, the target device management terminal requires warning commands to be transmitted in XML format with UTF-8 character encoding. Furthermore, the target warning commands are rearranged according to the XML format, organizing each information element (device number, device name, warning symbol, time range, maintenance operation description, checksum, etc.) according to the prescribed tags and structure. Simultaneously, the character encoding of the entire command is converted to UTF-8.

[0136] After the above-mentioned format standardization process, a standardized early warning command is obtained. This standardized early warning command meets the communication protocol requirements of the target device management terminal, ensuring that it can be correctly identified and parsed during transmission. When the device management terminal receives the standardized early warning command, it can accurately extract various information from it according to its internal parsing mechanism, thereby taking appropriate measures in a timely manner to handle the abnormal situation of the refrigeration unit.

[0137] As a non-limiting embodiment, after sending the device warning command to the target device management terminal, the method further includes:

[0138] Collect early warning processing feedback information returned by the target device management terminal. The early warning processing feedback information includes the execution status identifier of the early warning command and the subsequent operation data record of the corresponding device status.

[0139] Semantic parsing is performed on the early warning processing feedback information to extract the processing result label corresponding to the execution status identifier and the sequence of changes in device status parameters in subsequent operation data records;

[0140] The labels of the processing results are associated and mapped with the corresponding time-frequency domain feature sets and anomaly identification results to construct an extended set of anomaly samples that includes the dimensions of the processing results;

[0141] The pre-trained model performance evaluation module is invoked to perform model performance analysis on the extended set of abnormal samples, and to identify the false detection rate and false negative rate characteristics of the hybrid machine learning model under different anomaly types.

[0142] Based on the false positive rate and false negative rate features, model parameter adjustment suggestions are generated. The adjustment suggestions include the memory unit update strategy of LSTM neural network and the outlier threshold correction strategy of isolated forest algorithm. The model parameter adjustment suggestions are input into the parameter optimization interface of hybrid machine learning model to complete the dynamic optimization and update process of model.

[0143] After sending standardized early warning commands to the target equipment management terminal, the terminal processes the commands and returns feedback information. Based on the commands, the terminal takes appropriate maintenance actions and records the execution status, such as "processed," "processing," or "processing failed." Simultaneously, the terminal continues to collect subsequent operational data on the refrigeration unit's equipment status, such as changes in parameters like compressor operating current, condenser temperature, and refrigerant pressure after the maintenance operations. This feedback information is then sent back to the system.

[0144] After receiving the early warning feedback information, the system first performs semantic parsing. For execution status identifiers, they are converted into corresponding processing result labels; for example, "processed" is converted to "successfully processed," and "processing failed" is converted to "processing unsuccessfully processed." For subsequent operational data records, the system extracts the sequence of changes in equipment status parameters, such as recording the minute-by-minute change in compressor operating current after maintenance. Next, the processing result labels are associated with the previously generated time-frequency domain feature sets and anomaly identification results. For example, if the processing result label is "successfully processed," and the corresponding anomaly identification result is a compressor malfunction, then the successful processing label is associated with the time-frequency domain feature set corresponding to the anomaly, indicating that the equipment status changed accordingly after the anomaly was processed.

[0145] Through the above connections, an extended set of anomaly samples is constructed, incorporating the dimension of processing results. This extended set contains more information about the anomaly processing outcomes, providing richer data for model performance analysis. Then, the pre-trained model performance evaluation module is invoked to perform model performance analysis on the extended set of anomaly samples. This module analyzes the false positive and false negative rates of the hybrid machine learning model under different anomaly types. For example, for compressor fault anomalies, it checks whether the model misdetects normal operation as an anomaly (false positive rate) and whether it misses actual anomalies (false negative rate). Based on the analyzed false positive and false negative rate characteristics, model parameter adjustment suggestions are generated. For example, if a high false positive rate is found for compressor fault anomalies, it might be suggested to adjust the memory unit update strategy of the LSTM neural network to better learn normal operation modes; if a high false negative rate is found, it might be suggested to correct the outlier threshold of the isolated forest algorithm to more accurately identify anomalies. Finally, these model parameter adjustment suggestions are input into the parameter optimization interface of the hybrid machine learning model. The model will adjust its parameters according to these suggestions, completing dynamic optimization and update processing.

[0146] Through the above methods, hybrid machine learning models can continuously adapt to the actual operating conditions of equipment, improving the accuracy and reliability of anomaly detection.

[0147] As a non-limiting embodiment, after sending the device warning command to the target device management terminal, the method further includes:

[0148] Continuously acquire the continuous supplementary operation data set of the target building equipment after the warning command is sent. The supplementary operation data set contains equipment status record units with the same timestamp marking rules as the historical continuous operation data set.

[0149] Time-frequency domain feature extraction processing is performed on the supplementary runtime data set to obtain a supplementary time-frequency domain feature set;

[0150] The supplementary time-frequency domain feature set is input into the hybrid machine learning model for anomaly detection processing to generate supplementary anomaly identification results.

[0151] Extract novel anomaly feature patterns from the supplementary anomaly identification results that do not match the preset anomaly type feature library; perform feature clustering processing on the novel anomaly feature patterns, and divide the novel anomaly type clusters based on feature similarity measurement rules;

[0152] A unique anomaly type identifier is assigned to each new anomaly type cluster, and the time-frequency domain feature pattern corresponding to the new anomaly type cluster is extracted as the new anomaly feature descriptor;

[0153] Add the new anomaly type identifier and the new anomaly feature descriptor to the preset anomaly type feature library to complete the dynamic expansion and update of the anomaly type feature library.

[0154] After sending an equipment warning command, the system continuously acquires a supplementary set of operational data for the refrigeration unit following the warning command. This supplementary operational data set follows the same timestamp rules as the historical continuous operational data set, such as recording equipment status parameters every minute. The acquisition equipment continues to monitor various parameters of the refrigeration unit in real time, such as compressor operating current, condenser temperature, and refrigerant pressure, and records them according to the timestamp rules, forming new equipment status record units. These units constitute the supplementary operational data set. Next, time-frequency domain feature extraction processing is performed on the supplementary operational data set, in the same way as the previous processing of historical data. First, a time window is divided, then time-domain feature subsets and frequency-domain feature subsets are extracted separately, and finally, the two are correlated and matched to obtain a supplementary time-frequency domain feature set. The supplementary time-frequency domain feature set is input into a hybrid machine learning model for anomaly detection processing. The model analyzes these new feature sets based on learned normal and abnormal patterns to generate supplementary anomaly identification results.

[0155] In supplementing the anomaly identification results, a thorough check is performed to identify any novel anomaly feature patterns that do not match the pre-defined anomaly type feature library. For example, a new compressor operating current fluctuation pattern might be discovered that does not match any pattern in the pre-defined anomaly type feature library. For such novel anomaly feature patterns, feature clustering is performed. A feature similarity metric (e.g., Euclidean distance metric) is used to group novel anomaly feature patterns with similar characteristics together, classifying them into different novel anomaly type clusters. For example, by calculating the Euclidean distance between different novel anomaly feature patterns, patterns with closer distances are grouped into one cluster. A unique anomaly type identifier is assigned to each novel anomaly type cluster, such as "Anomaly Type X1", "Anomaly Type X2", etc. Simultaneously, the time-frequency domain feature pattern corresponding to each novel anomaly type cluster is extracted as a novel anomaly feature descriptor. Finally, the novel anomaly type identifier and novel anomaly feature descriptor are added to the pre-defined anomaly type feature library. This dynamically expands and updates the anomaly type feature library, enabling it to contain more anomaly type information and improving the system's ability to identify and handle novel anomalies.

[0156] As time goes on, new operational data and new anomalies are constantly being discovered and addressed, and the anomaly type feature library is continuously updated and improved, enabling the system to better adapt to the ever-changing operating conditions of the refrigeration unit.

[0157] It is understandable that in time-domain feature extraction (such as fluctuation amplitude, rate of change, and periodic fluctuation features), the statistical module in Scikit-learn can be used to calculate the difference between the maximum and minimum values, the ratio of the difference between adjacent time points to the time interval, and perform autocorrelation analysis. Its built-in functions (such as sklearn.preprocessing for normalization) can be used to automatically eliminate dimensional differences and generate a consistent feature subset. For frequency-domain transformation, windowing (such as the Hanning window) and smoothing filtering can be implemented based on the FFT algorithm of NumPy and SciPy (such as numpy.fft). The Gaussian filtering function of the scipy.signal module can be used to automatically perform weighted averaging to suppress noise, thereby obtaining a smooth frequency signal.

[0158] In this embodiment, during the anomaly detection stage of the hybrid machine learning model, the memory units, forget gates, and output gates of the LSTM neural network can be processed using standard implementations of TensorFlow or Keras frameworks (such as tf.keras.layers.LSTM). Based on publicly available gating mechanisms, time series dependencies are automatically modeled. Meanwhile, the outlier detection of the Isolation Forest algorithm is directly implemented using the IsolationForest class of Scikit-learn, whose built-in decision tree partitioning strategy efficiently identifies anomalous feature points, thereby simplifying manual parameter settings.

[0159] In addition, probability calculations (such as the probability of an anomaly recurring) can be performed by analyzing historical active period data through existing time series prediction techniques (such as the ARIMA model), automatically outputting quantitative values, and dynamically adjusting the warning level in conjunction with the rule base; while maintenance strategy encoding conversion uses existing NLP tools (such as rule-based or pre-trained models) to parse multi-segment structures and generate natural language text.

[0160] Therefore, the processes of time-frequency domain feature extraction, model detection, and early warning command generation can be further optimized, and communication reliability can be ensured through error checking coding protocols (such as CRC algorithms), achieving high-precision anomaly identification. This ensures that the equipment management terminal receives standardized early warning commands and optimizes its response, thereby improving the efficiency of building equipment monitoring.

[0161] The embodiments of this application construct a complete and efficient system for identifying and warning of abnormal building equipment.

[0162] First, by acquiring a set of continuously running data with timestamps, the state of the equipment at different times is fully recorded; the time-frequency domain feature extraction processing can deeply mine data features from both time and frequency domains, overcoming the limitations of single-dimensional analysis, more comprehensively depicting the operating characteristics of the equipment, and greatly improving the ability to capture abnormal features.

[0163] Secondly, the pre-built hybrid machine learning model combines the advantages of multiple algorithms to perform anomaly detection on time-frequency domain feature sets, significantly improving the accuracy and reliability of anomaly detection and effectively reducing false positives and false negatives. Based on the anomaly identification results, the anomaly type and spatiotemporal distribution characteristics are determined, enabling managers to clearly grasp the nature and development patterns of anomalies and providing a strong basis for targeted handling.

[0164] Furthermore, based on the above information, a target early warning instruction containing a time location identifier is generated and sent to the terminal, realizing accurate early warning. This allows managers to know in advance the time and type of anomalies, thereby arranging maintenance in a timely manner. This greatly improves the timeliness and effectiveness of equipment maintenance, reduces losses caused by equipment failures, ensures the stable and reliable operation of building equipment, and comprehensively enhances the level of intelligence in building equipment management.

[0165] Based on the same inventive concept, embodiments of this application also provide a building equipment anomaly identification system. See also... Figure 2 As shown, it is a schematic diagram of a possible building equipment anomaly identification system provided in an embodiment of this application. Figure 2The building equipment anomaly identification system 200 includes a processor 210 and a memory 220. The memory 220 stores computer programs executable by the processor 210. By executing the instructions stored in the memory 220, the processor 210 can perform the steps of the aforementioned machine learning-based building equipment anomaly identification method.

[0166] Based on the same inventive concept, embodiments of this application provide a computer-readable storage medium including a computer program. When the computer program is run on a building equipment anomaly identification system, the computer program is used to cause the building equipment anomaly identification system to perform the steps of the aforementioned machine learning-based building equipment anomaly identification method. In some possible implementations, various aspects of the machine learning-based building equipment anomaly identification method provided in this application can also be implemented in the form of a program product, including a computer program. When the program product is run on a building equipment anomaly identification system, the computer program is used to cause the building equipment anomaly identification system to perform the steps in the aforementioned machine learning-based building equipment anomaly identification method. For example, the building equipment anomaly identification system can perform, for example, ... Figure 1 The steps are shown in the figure.

[0167] In the technical solutions involved in the above embodiments of the present invention, whether performing comparison calculations of multi-dimensional features or constructing composite parameters, if there are problems caused by significant differences in the number of dimensions, units of measurement, and semantic meanings of different features, those skilled in the art, based on their professional knowledge and past practical experience, can fully understand that these differences need to be properly handled so that the calculation results are accurate and comparable, and to avoid situations such as logical confusion and unclear mathematical meaning.

[0168] In detail, when faced with features of different numbers of dimensions, those skilled in the art can employ various strategies, including but not limited to feature selection, feature extraction, and kernel function processing, in order to accurately calculate the similarity, matching degree, or feature distance between different features.

[0169] In order to achieve comparability alignment of feature spaces when comparing multidimensional features, those skilled in the art can use a variety of existing and common technical means, including but not limited to standardization preprocessing, mapping transformation, and spatial projection.

[0170] In the process of constructing composite parameters (such as loss function values), different parameter terms often have different dimensions. Those skilled in the art can use normalization processing or an adaptive weight allocation mechanism based on distribution characteristics.

[0171] The aforementioned general techniques for solving feature matching and loss balance problems are all common knowledge in this field. These techniques have been fully verified and widely used in numerous practical applications, and those skilled in the art can skillfully and flexibly apply these methods to handle similar problems involving differences in dimensions.

[0172] The formulas and calculation processes involved in the embodiments of this application, whether used for multidimensional feature comparison or composite loss function construction, strictly adhere to the principle of dimensional correspondence. Each variable in the formula has a clear and explicit physical meaning, and its operational logic fully conforms to basic mathematical and physical logic. The calculation results are necessarily the reasonable results expected by this application. Those skilled in the art are capable of effectively solving various problems arising from the number of dimensions, dimensional differences, etc., in the multidimensional feature comparison calculation and composite loss function construction in the embodiments, based on specific data conditions and business needs, by comprehensively utilizing the above-mentioned general technical means, thus ensuring the accuracy, reliability, and implementability of the technical solution of this invention.

Claims

1. A method for identifying anomalies in building equipment based on machine learning, characterized in that, The method includes: Acquire a continuous operating data set of equipment in the target building, wherein the continuous operating data set includes multiple equipment status record units with timestamps; The continuous operating data set is subjected to time-frequency domain feature extraction processing to obtain the time-frequency domain feature set of the device status recording unit; A pre-built hybrid machine learning model is invoked to perform anomaly detection processing on the time-frequency domain feature set, generating anomaly identification results for the device status recording unit; Determining the anomaly type of the target building equipment and the distribution characteristics of the anomaly type in the time dimension based on the anomaly identification results includes: parsing the anomaly feature descriptors in the anomaly identification results and extracting the time-frequency domain feature patterns corresponding to the anomaly feature descriptors; matching the time-frequency domain feature patterns with a preset anomaly type feature library to determine the anomaly type corresponding to the anomaly feature descriptors; extracting the timestamp information of the anomaly feature descriptors in the anomaly identification results and counting the occurrence frequency of each anomaly type in different time intervals; performing time series analysis on the occurrence frequency of the anomaly types to identify the active and inactive time periods of the anomaly types in the time dimension; constructing the time distribution curve of the anomaly type based on the active and inactive time periods to generate time distribution characteristic information containing time intervals and corresponding frequencies; The step of performing time series analysis on the occurrence frequency of the anomaly type to identify active and inactive periods of the anomaly type in the time dimension includes: performing sliding window statistical processing on the time series of the occurrence frequency of the anomaly type to calculate the average occurrence frequency within each time window; comparing the difference in average occurrence frequency between adjacent time windows to identify the time point of sudden increase in occurrence frequency as the starting point of the active period; identifying the time point of sudden decrease in occurrence frequency as the ending point of the active period; within the inactive period, statistically identifying the time interval with occurrence frequency lower than a preset frequency as the inactive period; performing boundary smoothing processing on the active and inactive periods to eliminate the time period division error caused by accidental fluctuations, obtaining smoothed active and inactive periods; associating and storing the smoothed active and inactive periods with the anomaly type to form the time distribution feature information of the anomaly type; Based on the anomaly type and the time distribution feature information, a target early warning instruction containing a time location identifier is generated, and the target early warning instruction is sent to the target device management terminal; The step of generating a target early warning instruction containing a time location identifier based on the anomaly type and the time distribution feature information includes: parsing a preset early warning rule base corresponding to the anomaly type, extracting an early warning priority identifier and maintenance strategy code associated with the anomaly type; extracting active and inactive time periods from the time distribution feature information, and determining the current active time interval of the anomaly type; calculating the probability of the anomaly type recurring in a specified subsequent time period based on the current active time interval; performing a weighted association processing between the recurrence probability and the early warning priority identifier of the anomaly type to generate a dynamic early warning level parameter; performing information fusion processing on the dynamic early warning level parameter, the maintenance strategy code, and the current active time interval to generate a target early warning instruction containing a time location identifier; performing format standardization processing on the target early warning instruction to obtain a standardized early warning instruction; and matching the communication protocol requirements of the target device management terminal.

2. The machine learning-based building equipment anomaly identification method according to claim 1, characterized in that, The step of performing time-frequency domain feature extraction processing on the continuously running data set to obtain the time-frequency domain feature set of the device status recording unit includes: The continuous operating data set is divided into time windows to obtain multiple data segment units with continuous time series relationship. Each data segment unit corresponds to a device status record of a fixed duration. For each data segmentation unit, time-domain feature extraction processing is performed to extract the fluctuation amplitude features, change rate features, and periodic fluctuation features of the device state parameters in the data segmentation unit as a time-domain feature subset; Each data segmentation unit undergoes frequency domain transformation to convert the time-series signal of the data segmentation unit into a frequency domain signal representation. The frequency domain signal representation is subjected to energy distribution analysis and processing to extract the energy proportion characteristics, main frequency component characteristics and frequency component stability characteristics of different frequency intervals as frequency domain feature subsets; The time-domain feature subset is associated and matched with the features of the corresponding data segmentation units in the frequency-domain feature subset to generate a time-frequency domain feature set for each device status recording unit.

3. The machine learning-based building equipment anomaly identification method according to claim 2, characterized in that, The step of performing time-domain feature extraction processing on each data segment unit, extracting the fluctuation amplitude features, change rate features, and periodic fluctuation features of the device state parameters in the data segment unit as a time-domain feature subset, includes: The difference between the maximum and minimum values ​​of the device status parameters in the data segmentation unit is calculated to obtain the fluctuation amplitude characteristics of the data segmentation unit; The ratio of the difference in device state parameters at adjacent time points in the data segmentation unit to the time interval is calculated to obtain the rate of change characteristic of the data segmentation unit. Autocorrelation analysis is performed on the device status parameter sequence of the data segmentation unit to identify the periodicity of recurring patterns in the sequence, thereby obtaining the periodic fluctuation characteristics of the data segmentation unit. The fluctuation amplitude feature, the rate of change feature, and the periodic fluctuation feature are normalized to eliminate the dimensional differences between different parameters. The normalized fluctuation amplitude characteristics, rate of change characteristics, and periodic fluctuation characteristics are combined into a time-domain feature subset of the data segmentation unit; The step of performing frequency domain transformation on each data segment unit, converting the time-series signal of the data segment unit into a frequency domain signal representation, includes: The time series signal of the data segmentation unit is windowed to suppress spectral leakage at the signal edges, resulting in a windowed time series signal. The windowed time series signal is processed using the Fast Fourier Transform algorithm to convert the time domain signal into a frequency domain signal. Extract the real and imaginary parts of the frequency domain signal to generate a frequency distribution list containing frequency values ​​and corresponding amplitudes; The frequency distribution list is smoothed by filtering to eliminate high-frequency noise interference and obtain a smooth frequency domain signal representation. The smooth frequency domain signal representation is associated with and stored with the timestamp information of the data segmentation unit to form a frequency domain signal representation with a time-frequency correspondence.

4. The machine learning-based building equipment anomaly identification method according to claim 1, characterized in that, The step of calling a pre-built hybrid machine learning model to perform anomaly detection processing on the time-frequency domain feature set and generating anomaly identification results for the device status recording unit includes: The time-frequency domain feature set is input into the LSTM neural network of the hybrid machine learning model to model the time-series dependency of the time-frequency domain feature set and generate a sequence feature representation with time-series correlation information. The sequence feature representation is input into the isolated forest algorithm of the hybrid machine learning model to perform outlier detection processing on the sequence feature representation and identify abnormal feature points in the sequence feature representation that deviate from the normal pattern. The abnormal feature points are subjected to contextual correlation analysis, and combined with the feature information of adjacent device status recording units in the time-frequency domain feature set, it is determined whether the abnormal feature points are persistent or sporadic. If so, the degree of abnormality of the abnormal feature points is quantitatively evaluated to generate an abnormal feature descriptor containing the abnormality confidence level. The abnormal feature descriptor is associated with the timestamp information of the device status recording unit to generate the abnormal identification result of the device status recording unit.

5. The machine learning-based building equipment anomaly identification method according to claim 4, characterized in that, The step of inputting the time-frequency domain feature set into the LSTM neural network of the hybrid machine learning model to model the time-series dependency of the time-frequency domain feature set and generate a sequence feature representation with time-series correlation information includes: The time-frequency domain feature set is subjected to feature dimension alignment processing to unify the number of feature dimensions of different device status recording units; The dimension-aligned time-frequency domain feature set is input into the input layer of the LSTM neural network in timestamp order to generate the initial feature input vector; The initial feature input vector is processed by the memory unit of the LSTM neural network to retain the feature information of the previous device state recording unit, thus obtaining historical information. The historical information is filtered by the forget gate of the LSTM neural network to remove noisy historical features and obtain the filtered historical information. The output gate of the LSTM neural network is used to fuse the filtered historical information with the current feature input vector to generate a sequence feature representation containing temporal correlation information.

6. The machine learning-based building equipment anomaly identification method according to claim 1, characterized in that, The step of fusing the dynamic early warning level parameters, maintenance strategy codes, and the current active time interval to generate a target early warning instruction containing a time location identifier includes: The dynamic warning level parameter is assigned a corresponding warning symbol, and the warning symbol is positively correlated with the warning level parameter. The maintenance strategy is encoded into maintenance operation description text that includes maintenance steps. Convert the current active time interval into a time range descriptor, which includes a start time point and an end time point; The warning symbol, maintenance operation description text, and time range descriptor are combined in a structured manner to generate a target warning instruction body containing a time location identifier. Add device identification information to the target early warning command body, the device identification information being used to uniquely identify the target building equipment; The target warning instruction subject and device identification information are verified and encoded to generate a target warning instruction with error verification function.

7. The machine learning-based building equipment anomaly identification method according to claim 6, characterized in that, The step of encoding the maintenance strategy into maintenance operation description text containing maintenance step instructions includes: Obtain the multi-segment coding structure of the maintenance strategy code, which consists of a maintenance action type identifier segment, an action object identifier segment, and an operation condition identifier segment arranged in sequence. The pre-built maintenance action knowledge base is invoked. The maintenance action knowledge base stores basic step description units that correspond one-to-one with the maintenance action type identifier segment, equipment component positioning description units that correspond one-to-one with the action object identifier segment, and execution constraint description units that correspond one-to-one with the operation condition identifier segment. The multi-segment coding structure is segmented and parsed to extract the encoded values ​​of the maintenance action type identifier segment, action object identifier segment, and operation condition identifier segment; Based on the encoded value of the maintenance action type identifier segment, the corresponding basic step description unit is obtained by matching from the maintenance action knowledge base. The basic step description unit includes standard action verbs and general operation direction information. Based on the encoded value of the action object identifier segment, the corresponding equipment component positioning description unit is obtained from the maintenance action knowledge base. The equipment component positioning description unit includes the component name and spatial location information. Based on the encoded value of the operation condition identifier segment, the corresponding execution constraint description unit is obtained from the maintenance action knowledge base. The execution constraint description unit includes environmental parameter range requirements and safety precautions information. The standard action verbs of the basic step description unit are semantically fused with the component names and spatial location information of the device component positioning description unit to generate action description clauses containing the operation object. The action description clause is conditionally associated with the environmental parameter range requirements and safety precautions information of the execution constraint description unit to generate a step constraint description clause containing the operation prerequisites. According to the arrangement order of the multi-segment coding structure, the action description clause and the step constraint description clause are logically concatenated to generate a maintenance step description sequence with sequential correlation. The maintenance step description sequence is optimized for natural language fluency by adjusting the grammatical structure and conjunctions of the statements to generate a maintenance operation description text containing complete operation logic.

8. A building equipment anomaly identification system, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, causes the processor to perform the steps of any one of the methods described in claims 1 to 7.

Images