Building equipment abnormity identification method and system based on machine learning

Abstract

The invention relates to the technical field of machine learning, and provides a building equipment abnormity identification method and system based on machine learning, which are used for realizing accurate detection and accurate early warning of building equipment abnormity. The method comprises the following steps: acquiring a continuous operation data set of target building equipment, wherein the continuous operation data set comprises multiple segments of equipment state recording units with timestamp marks; performing time-frequency domain feature extraction processing on the continuous operation data set to obtain a time-frequency domain feature set of the equipment state recording unit; calling a pre-constructed hybrid machine learning model to perform anomaly detection processing on the time-frequency domain feature set, and generating an anomaly recognition result of the equipment state recording unit; determining the anomaly type of the target building equipment and the distribution feature information of the anomaly type in the time dimension according to the anomaly recognition result; and generating a target early warning instruction containing a time positioning identifier based on the exception type and the time distribution feature information, and sending the equipment early warning instruction to a target equipment management terminal.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of machine learning, and particularly relates to a building equipment abnormality identification method and system based on machine learning. BACKGROUND

[0002] In the existing building equipment management technology, there are many deficiencies in the detection and early warning means of equipment abnormalities. Traditional technologies often only rely on single-dimensional data feature analysis, which cannot comprehensively grasp the equipment operation condition and is easy to miss complex and variable abnormal conditions. Although some technologies can obtain equipment operation data, they lack deep mining of data in the time dimension and are difficult to accurately locate the time and type of abnormality occurrence.

[0003] Among them, some abnormality detection methods based on simple models have low detection accuracy and reliability in the face of complex equipment operation modes and diversified abnormal conditions, and false detection and missed detection problems occur frequently. Moreover, most of the existing technologies fail to effectively integrate abnormal information and time distribution characteristics, resulting in the generated early warning instructions lacking precise time positioning, making it difficult for management personnel to timely and targetedly maintain the equipment, increasing the impact of equipment failure on the normal operation of the building, and reducing the service life and operation efficiency of the equipment. SUMMARY

[0004] The application provides a building equipment abnormality identification method and system based on machine learning, which realizes accurate detection and precise early warning of building equipment abnormalities, improves equipment management efficiency, and ensures stable operation of the equipment.

[0005] In a first aspect, the application embodiment provides a building equipment abnormality identification method based on machine learning, applied to a building equipment abnormality identification system, and the method comprises the following steps: obtaining a continuous operation data set of a target building equipment, wherein the continuous operation data set comprises a plurality of equipment state record units with time stamp markers; performing time-frequency domain feature extraction processing on the continuous operation data set to obtain a time-frequency domain feature set of the equipment state record unit; calling a pre-constructed hybrid machine learning model to perform abnormality detection processing on the time-frequency domain feature set to generate an abnormality identification result of the equipment state record unit; determining an abnormality type of the target building equipment and distribution characteristic information of the abnormality type in the time dimension according to the abnormality identification result; generating a target early warning instruction containing a time positioning identifier based on the abnormality type and the time distribution characteristic information, and sending the equipment early warning instruction to a target equipment management terminal.

[0006] In a second aspect, the embodiments of the present application provide a building equipment anomaly identification system, comprising a processor and a memory, wherein the memory stores a computer program which, when executed by the processor, causes the processor to perform the steps of the above method.

[0007] In a third aspect, the embodiments of the present application provide a computer-readable storage medium comprising a computer program, which, when executed on a building equipment anomaly identification system, is configured to cause the building equipment anomaly identification system to perform the steps of the above method.

[0008] The embodiments of the present application construct a complete and efficient building equipment anomaly identification and early warning system as a whole.

[0009] Firstly, by acquiring a continuous running data set with a time stamp, the state of the equipment at different times is recorded completely; the time-frequency domain feature extraction process can deeply mine data features from two dimensions of time domain and frequency domain, overcoming the limitations of single-dimensional analysis, more comprehensively depicting equipment running characteristics, and greatly improving the capture ability of abnormal features.

[0010] Secondly, the pre-constructed hybrid machine learning model combines the advantages of multiple algorithms to detect anomalies in the time-frequency domain feature set, significantly improving the accuracy and reliability of anomaly detection, and effectively reducing false positives and false negatives. According to the anomaly identification result, the abnormal type and spatiotemporal distribution feature information are determined, so that the manager can clearly understand the nature and development law of the anomaly, and provide a strong basis for targeted processing.

[0011] Further, based on the above information, a target early warning instruction containing a time positioning identifier is generated and sent to the terminal, realizing accurate early warning, so that the manager knows in advance the time and type of the anomaly, and thus arranges maintenance in a timely manner, greatly improving the timeliness and effectiveness of equipment maintenance, reducing the loss caused by equipment failure, ensuring the stable and reliable operation of building equipment, and comprehensively improving the intelligent level of building equipment management. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 A flowchart of a building equipment anomaly identification method based on machine learning provided by the embodiments of the present application.

[0013] Figure 2 A structural diagram of a building equipment anomaly identification system provided by the embodiments of the present application. DETAILED DESCRIPTION

[0014] To make the purposes, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions of the present application will be described below in connection with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments described in the present application document, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the scope of protection of the technical solutions of the present application.

[0015] Referring to Figure 1 It is a machine learning-based building equipment anomaly identification method provided in the embodiments of the present application, which can be applied to a building equipment anomaly identification system, and the specific process is as shown in steps 110-150.

[0016] Step 110: Obtain a continuous running data set of a target building equipment, wherein the continuous running data set contains multiple device state record units with time stamp markers.

[0017] In the embodiments of the present application, a refrigeration unit in an air conditioning system of a large commercial center is taken as a target building equipment, which runs 24 hours a day to provide a suitable temperature environment for the commercial center. In order to obtain a continuous running data set, data acquisition devices are deployed at key components and sensors of the refrigeration unit, which will collect various running parameters of the refrigeration unit in real time, such as the working current of the compressor, the temperature of the condenser, the pressure of the refrigerant, etc. Every certain time interval (for example, every minute), the acquisition device will organize the collected device state parameters and attach the current time stamp to form a device state record unit. With the passage of time, these device state record units with time stamp markers are continuously accumulated, thus constituting the continuous running data set of the target building equipment.

[0018] Step 120: Perform time-frequency domain feature extraction processing on the continuous running data set to obtain a time-frequency domain feature set of the device state record unit.

[0019] After obtaining the continuous running data set of the refrigeration unit, in order to deeply understand the running features of the equipment, it is necessary to perform time-frequency domain feature extraction processing. Through the above processing, key information in the equipment running data can be mined from two angles of time domain and frequency domain, so as to more comprehensively describe the running state of the equipment. The time domain features can reflect the change law of the device state parameters with time, and the frequency domain features can reveal the distribution of different frequency components in the signal. The combination of the two can more accurately depict the overall picture of the equipment running state, and provide strong support for accurate identification of anomalies.

[0020] In an optional embodiment, the time-frequency domain feature extraction processing on the continuous operation data set to obtain the time-frequency domain feature set of the device state record unit comprises: Step 121: performing time window division processing on the continuous operation data set to obtain a plurality of data segment units having a continuous time sequence relationship, each data segment unit corresponding to a fixed time length of device state record.

[0021] For the continuous operation data set of the refrigeration unit, time window division is performed according to a fixed time length (for example, 1 hour). Starting from the starting time of the data set, each 1 hour is taken as a time window, and the data is divided into a plurality of data segment units having a continuous time sequence relationship, so that each data segment unit contains the device state record of the refrigeration unit in the 1 hour. For example, the first data segment unit records the changes of the compressor operating current, condenser temperature, refrigerant pressure and other parameters of the refrigeration unit from 0 o'clock to 1 o'clock; the second data segment unit corresponds to the device operating state from 1 o'clock to 2 o'clock. Through the above time window division processing, the continuous operation data can be cut into small segments for processing and analysis, while the time sequence of the data is retained.

[0022] Step 122: performing time domain feature extraction processing on each data segment unit to extract the fluctuation amplitude feature, change rate feature and periodic fluctuation feature of the device state parameter in the data segment unit as a time domain feature subset.

[0023] For each data segment unit, the time domain features of each device state parameter of the refrigeration unit are extracted. Taking the compressor operating current as an example, the fluctuation amplitude feature is calculated, that is, the difference between the maximum value and the minimum value of the compressor operating current in the data segment unit is found; the change rate feature is calculated by calculating the ratio of the difference of the compressor operating current at adjacent time points to the time interval; the periodic fluctuation feature is obtained by performing autocorrelation analysis on the compressor operating current sequence to identify the repeating pattern period length in the sequence. Similarly, similar processing is performed on the condenser temperature, refrigerant pressure and other device state parameters. Combining these fluctuation amplitude features, change rate features and periodic fluctuation features extracted for different device state parameters, a time domain feature subset of each data segment unit is formed. The above time domain feature subset can comprehensively reflect the change characteristics of the device state parameters of the refrigeration unit in each data segment unit corresponding to the time period.

[0024] As a preferred embodiment, the time domain feature extraction processing on each data segment unit to extract the fluctuation amplitude feature, change rate feature and periodic fluctuation feature of the device state parameter in the data segment unit as a time domain feature subset comprises: Step 1221: Calculate the difference between the maximum and minimum values of the device state parameters in the data segment unit to obtain the fluctuation amplitude feature of the data segment unit; calculate the ratio of the difference between the device state parameters at adjacent time points to the time interval to obtain the change rate feature of the data segment unit; perform autocorrelation analysis on the sequence of device state parameters in the data segment unit to identify the periodically repeating pattern length in the sequence to obtain the periodic fluctuation feature of the data segment unit.

[0025] In calculating the fluctuation amplitude feature of the working current of the compressor of the refrigeration unit, all recorded working current values of the compressor in the data segment unit are traversed to find the maximum and minimum values, and then the difference between the two values is calculated. For example, in a certain 1-hour data segment unit, the maximum working current of the compressor is 50 amperes and the minimum working current is 30 amperes, and then the fluctuation amplitude feature of the working current of the compressor in the data segment unit is 50-30=20 amperes. For the change rate feature, the ratio of the difference between the working currents of the compressor at adjacent time points to the time interval is calculated. Among them, the working currents of the compressor recorded at adjacent two time points are 40 amperes and 42 amperes, and the time interval is 1 minute (60 seconds), and then the change rate feature is (42-40) / 60=1 / 30 amperes per second.

[0026] In performing the periodic fluctuation feature extraction, autocorrelation analysis is performed on the sequence of working currents of the compressor in the data segment unit. Through autocorrelation analysis algorithm, the periodically repeating pattern length in the sequence is found. For example, it is found through analysis that the working current of the compressor will appear a similar change pattern every 10 minutes, and then the periodic fluctuation feature of the working current of the compressor in the data segment unit is 10 minutes. The condenser temperature, refrigerant pressure and other device state parameters are calculated in the same way to obtain the corresponding fluctuation amplitude feature, change rate feature and periodic fluctuation feature.

[0027] Step 1222: Normalize the fluctuation amplitude feature, the change rate feature and the periodic fluctuation feature to eliminate the dimensional differences between different parameters; and combine the normalized fluctuation amplitude feature, the change rate feature and the periodic fluctuation feature into the time domain feature subset of the data segment unit.

[0028] Since the dimensions of the fluctuation amplitude feature, the change rate feature and the periodic fluctuation feature are different, in order to facilitate subsequent analysis and processing, normalization processing is needed. A normalization algorithm (such as the maximum and minimum normalization algorithm) is used to map these feature values to a unified interval (such as [0, 1]). For the fluctuation amplitude feature of the compressor working current, the original value is 20 amperes, the maximum value of the fluctuation amplitude feature in the data segmentation unit is 30 amperes, and the minimum value is 10 amperes. After the maximum and minimum normalization algorithm processing, the normalized fluctuation amplitude feature value is (20-10) / (30-10)=0.5. Similarly, the change rate feature and the periodic fluctuation feature are also normalized in a similar manner. The normalized fluctuation amplitude feature, change rate feature and periodic fluctuation feature are combined in a certain order to form the time domain feature subset of the data segmentation unit. For example, the normalized fluctuation amplitude feature, change rate feature and periodic fluctuation feature of the compressor working current are arranged in sequence to form a vector as part of the time domain feature subset. The corresponding features of other device state parameters are also combined in the same way to ultimately form the complete time domain feature subset of the data segmentation unit.

[0029] Step 123: performing frequency domain conversion processing on each data segmentation unit to convert the time series signal of the data segmentation unit into a frequency domain signal representation.

[0030] In order to further analyze the characteristics of the refrigeration unit operation data in the frequency domain, frequency domain conversion processing needs to be performed on each data segmentation unit. Through frequency domain conversion, the frequency information hidden in the time series signal can be revealed, so that the distribution of different frequency components in the device operation process can be found. Different device faults may produce abnormal signals in some frequency range, and through frequency domain analysis these signals can be captured to provide additional clues for accurate identification of abnormalities.

[0031] As a preferred embodiment, the frequency domain conversion processing on each data segmentation unit to convert the time series signal of the data segmentation unit into a frequency domain signal representation comprises: Step 1231: performing windowing processing on the time series signal of the data segmentation unit to suppress the spectrum leakage phenomenon of the signal edge, to obtain a windowed time series signal.

[0032] Before the time series signal of a certain data segment unit of the refrigeration unit (for example, the time series of the compressor working current) is converted into the frequency domain, a windowing process is performed. An appropriate window function (for example, a Hanning window) is selected, and the window function is multiplied with the original time series signal. The function of the window function is to truncate the signal, so that the part of the signal within the window can be more smoothly transitioned to zero, thereby suppressing the spectral leakage phenomenon of the signal edge. Among them, the original compressor working current time series signal is a sequence with a length of N, the Hanning window function is multiplied point by point with the sequence, and the windowed time series signal is obtained. Through the above windowing process, the accuracy of subsequent frequency domain conversion can be improved, and the influence of spectral leakage on the frequency domain analysis result can be reduced.

[0033] Step 1232: based on the fast Fourier transform algorithm, the windowed time series signal is processed to convert the time domain signal into the frequency domain signal; the real part and the imaginary part information of the frequency domain signal are extracted to generate a frequency distribution list containing the frequency value and the corresponding amplitude.

[0034] Optionally, the fast Fourier transform algorithm can efficiently convert the time domain signal into the frequency domain signal, and reveal the distribution of different frequency components in the signal. After the fast Fourier transform algorithm processing, the frequency domain signal is obtained. For the frequency domain signal, the real part and the imaginary part information are extracted. The real part and the imaginary part information can completely describe the characteristics of the frequency domain signal. According to the real part and the imaginary part information, the amplitude corresponding to each frequency is calculated, and the amplitude represents the intensity of the frequency component in the signal. The frequency value and the corresponding amplitude are arranged into a list to form a frequency distribution list containing the frequency value and the corresponding amplitude. Through the frequency distribution list, the energy distribution of the compressor working current signal at different frequencies can be directly understood.

[0035] Step 1233: the frequency distribution list is subjected to a smoothing filter process to eliminate high-frequency noise interference, and a smoothed frequency domain signal representation is obtained; the smoothed frequency domain signal representation is associated with the timestamp information of the data segment unit and stored to form a frequency domain signal representation with time-frequency correspondence.

[0036] To eliminate high-frequency noise interference in the frequency distribution list, a smoothing filter is used. For example, a Gaussian filter algorithm is used to perform weighted average processing on each data point in the frequency distribution list, so that the change between adjacent data points is more smooth, thereby effectively suppressing high-frequency noise. After Gaussian filter algorithm processing, a smooth frequency domain signal representation is obtained. For example, there are some data points with large fluctuations in a certain high frequency band in the original frequency distribution list, which are smoothed out after Gaussian filter algorithm processing, and the data is more stable. Then, the smooth frequency domain signal representation is stored in association 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 to form a data structure with time-frequency correspondence. In subsequent analysis, the corresponding frequency domain signal can be quickly located according to the timestamp information, facilitating analysis and comparison of frequency domain characteristics of equipment operation in different time periods.

[0037] Step 124: Perform energy distribution analysis on the frequency domain signal representation to extract energy proportion features, main frequency component features, and frequency component stability features in different frequency intervals as a frequency domain feature subset.

[0038] The processed frequency domain signal representation with time-frequency correspondence is analyzed for energy distribution. First, the entire frequency range is divided into multiple frequency intervals, such as 0-10Hz, 10-20Hz, 20-30Hz, etc. The energy proportion feature in each frequency interval is calculated by summing the squares of all frequency values corresponding to the amplitude in that frequency interval, then comparing the energy of each frequency interval with the total energy to obtain the energy proportion of that frequency interval. For example, in the 0-10Hz frequency interval, the sum of the squares of all frequency values corresponding to the amplitude is E1, and the total energy is E, so the energy proportion of that frequency interval is E1 / E. Then, find the frequency interval with the largest energy proportion, and the frequency component in that interval is the main frequency component feature. For the frequency component stability feature, analyze the change of the amplitude of each frequency component over time in the frequency distribution list. If the amplitude of a certain frequency component fluctuates little over a period of time, it means that the frequency component has high stability. These energy proportion features, main frequency component features, and frequency component stability features in different frequency intervals are combined to form a frequency domain feature subset, which can reflect the energy distribution, main frequency component, and stability of the frequency component of the refrigeration unit operation signal in different frequency intervals, providing important frequency domain information for anomaly detection.

[0039] Step 125: Perform association matching processing on the time domain feature subset and the features of the corresponding data segmentation unit in the frequency domain feature subset to generate a time-frequency domain feature set for each device state record unit.

[0040] Optionally, the time domain feature subset and the frequency domain feature subset of each data segment unit are associated and matched. Since the time domain feature subset and the frequency domain feature subset are extracted for the same data segment unit, there is an inherent connection between them. The fluctuation amplitude feature, the change rate feature and the periodic fluctuation feature in the time domain feature subset are combined with the energy proportion feature, the main frequency component feature and the frequency component stability feature in the frequency domain feature subset according to certain rules. For example, the fluctuation amplitude feature can be associated with the energy proportion feature, the change rate feature can be associated with the main frequency component feature, and the periodic fluctuation feature can be associated with the frequency component stability feature. Through the above associated matching processing, a time-frequency domain feature set of each device state record unit is generated, which comprehensively combines the time domain and frequency domain information and can more comprehensively and accurately describe the running state features of the refrigeration unit in each data segment unit corresponding to the time period.

[0041] Step 130: calling a pre-built hybrid machine learning model to perform anomaly detection processing on the time-frequency domain feature set to generate an anomaly recognition result of the device state record unit.

[0042] Optionally, the generated time-frequency domain feature set of the refrigeration unit device state record unit is input into a pre-built hybrid machine learning model. The model combines the advantages of multiple machine learning algorithms and can more accurately analyze and process the time-frequency domain feature set to identify abnormal conditions therein. The hybrid machine learning model has mastered the feature patterns under normal device running state 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 whether there is an anomaly. In this way, potential problems in the refrigeration unit running process can be discovered in a timely manner.

[0043] Under a preferred design idea, the calling of the pre-built hybrid machine learning model to perform anomaly detection processing on the time-frequency domain feature set to generate the anomaly recognition result of the device state record unit includes: Step 131: inputting the time-frequency domain feature set into an 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 sequence association information.

[0044] First, the time-frequency domain feature set of the refrigeration unit is input into the LSTM neural network of the hybrid machine learning model. The LSTM neural network is good at processing time series data and can capture the time series dependency between features at different times in the time-frequency domain feature set. Before input, the time-frequency domain feature set is preprocessed as necessary to ensure that the format and range of the data meet the requirements of the LSTM neural network. For example, normalize each feature in the feature set to have a value range 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 cells, forget gates, and output gates. The memory cells retain the feature information of the previous device state recording unit, the forget gates filter out noisy historical feature information, and the output gates fuse the filtered historical information with the current feature input vector to generate a sequence feature representation with time sequence correlation information. This sequence feature representation not only contains the time-frequency domain feature information at the current time, but also incorporates relevant information from previous times, better reflecting the trend of the refrigeration unit's operating state over time.

[0045] In one implementation, the input of the time-frequency domain feature set into the LSTM neural network of the hybrid machine learning model models the time series dependency of the time-frequency domain feature set and generates a sequence feature representation with time sequence correlation information, including: 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.

[0046] Since the number of feature dimensions of different device state recording units in the time-frequency domain feature set of the refrigeration unit may differ, before inputting it into the LSTM neural network, it needs to be processed for feature dimension alignment. Check the time-frequency domain feature set of each device state recording unit to determine the maximum number of feature dimensions. Then, for device state recording units with insufficient feature dimensions, fill or expand them to make their feature dimension number consistent with the maximum value. For example, if a device state recording unit has only 5 feature dimensions in its time-frequency domain feature set, while other units have a maximum of 8 feature dimensions, you can add some default values (such as 0) to the unit's feature set or expand it reasonably based on existing features to expand its feature dimension to 8. Through the above feature dimension alignment processing, the time-frequency domain feature sets of different device state recording units can have a unified format when input into the LSTM neural network, facilitating network processing and learning.

[0047] Step 1312: input the dimension-aligned time-frequency domain feature set in chronological order into the input layer of the LSTM neural network to generate an initial feature input vector.

[0048] The time-frequency domain feature set of the refrigeration unit after the feature dimension alignment processing is sequentially input into the input layer of the LSTM neural network according to the chronological order of the timestamps. The time-frequency domain feature set of each device state recording unit is input into the neurons of the input layer as a whole. The neurons of the input layer will preliminarily process and convert the input features into an initial feature input vector suitable for internal processing of the LSTM neural network. For example, the input layer may perform weighted summation and other operations on the input features to combine multiple features into an initial feature input vector in the form of a vector.

[0049] Step 1313: perform historical information memory processing on the initial feature input vector through the memory unit of the LSTM neural network to retain the feature information of the previous device state recording unit and obtain historical information.

[0050] It can be understood that the memory unit of the LSTM neural network will comprehensively process the initial feature input vector and the historical information retained at the previous time. For the operation data of the refrigeration unit, the memory unit will remember the time-frequency domain feature information of the previous device state recording unit. For example, when processing the initial feature input vector related to the operating current of the compressor at the current time, the memory unit will review the parameter at the previous time and other time-frequency domain feature information related thereto. In the above manner, the memory unit accumulates a series of historical information about the operating state of the refrigeration unit, which not only contains the change of the current parameter, but also covers the performance of other features related thereto at different times, which enables the network to capture the evolution law of the device 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 dependency in the operation data of the refrigeration unit.

[0051] Step 1314: perform screening processing on the historical information through the forget gate of the LSTM neural network to filter out noise historical feature information and obtain screened historical information.

[0052] Optionally, the forgetting gate evaluates and filters the historical information stored in the memory cell in the LSTM neural network, to determine which historical feature information is noise and which is valuable information for the current analysis. For the operation data of the refrigeration unit, the forgetting gate analyzes the correlation between each feature in the historical information and the current device state. For example, if the condenser temperature abnormally fluctuates at a certain historical time, but subsequent data shows that this is only a random noise event, the forgetting gate will reduce attention to this historical information, or even filter it out. Through the above filtering process, the forgetting gate can remove noise historical feature information that may interfere with network judgment, so that the historical information retained in the memory cell is more refined and valuable. The filtered historical information obtained in this way can more accurately reflect the long-term trends and important features of the refrigeration unit operation state, and help improve the accuracy of the network analysis of the current device state.

[0053] Step 1315: The output gate of the LSTM neural network is used to fuse the filtered historical information and the current feature input vector, to generate a sequence feature representation containing time series correlation information.

[0054] In detail, the output gate is responsible for fusing the filtered historical information and the feature input vector at the current time. In the example of the refrigeration unit, the output gate combines the historical information about the compressor operating current, condenser temperature, refrigerant pressure, etc. filtered by the forgetting gate, with the initial feature input vector of these parameters corresponding to the current time. The output gate performs weighted summation and other operations on the historical information and the current feature according to certain weight distribution rules, so that the generated sequence feature representation contains both the information of the current device state and the memory of the past related state. For example, the output gate can dynamically adjust the weights of the historical information and the current feature according to the stability of the current device operation, etc. If the current device operation state is relatively stable, it may rely more on historical information to generate 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 sequence feature representation containing time series correlation information generated finally can comprehensively reflect the relationship and changes between the operation states of the refrigeration unit at different times.

[0055] Step 132: The sequence feature representation is input into the Isolation Forest algorithm of the hybrid machine learning model to perform outlier detection processing on the sequence feature representation, to identify abnormal feature points in the sequence feature representation that deviate from the normal mode.

[0056] In this step, the sequence feature representation containing time series correlation information generated by the LSTM neural network is input into the Isolation Forest algorithm of the hybrid machine learning model. The Isolation Forest algorithm is an algorithm specifically used for outlier detection, which can quickly and effectively identify data points deviating from the normal mode in high-dimensional data space. For the sequence feature representation of the chiller unit, the Isolation Forest algorithm will build multiple decision trees, and each decision tree will randomly divide the data points in the sequence feature representation. In the division process, normal data points are usually divided into lower levels of the tree, while outliers are quickly divided into leaf nodes of the tree. In this way, the Isolation Forest algorithm can calculate the isolation score of each data point, and the higher the isolation score, the more likely the data point is an outlier. For example, for a certain data point about the compressor working current in the sequence feature representation, if the data point is quickly divided into the leaf node in the decision tree built by the Isolation Forest algorithm, and its isolation score is significantly higher than that of other data points, then the data point is identified as an abnormal feature point that may deviate from the normal mode. Through the above outlier detection process, those points that are significantly different from the normal operation mode can be found from the sequence feature representation, providing key clues for subsequent judgment of whether the device is abnormal.

[0057] Step 133: Contextual correlation analysis processing is performed on the abnormal feature points, combining the feature information of the adjacent device state record units in the time-frequency domain feature set to determine whether the abnormal feature points are persistent abnormalities or incidental abnormalities.

[0058] For the identified abnormal feature points, further analysis of their properties is needed to determine whether they are persistent abnormalities or incidental abnormalities, which requires contextual correlation analysis combined with the feature information of the adjacent device state record units in the time-frequency domain feature set. Taking a certain abnormal feature point of the chiller unit as an example, the abnormal feature point is a large fluctuation value of the compressor working current. First, the feature information of the compressor working current and other related device state parameters (such as condenser temperature, refrigerant pressure, etc.) in the adjacent device state record units before and after the abnormal feature point is checked. If the compressor working current continuously appears similar abnormal fluctuations in multiple adjacent device state record units, and other related parameters also appear corresponding abnormal changes, then it can be determined that the abnormal feature point is a persistent abnormality, which may mean that the compressor of the chiller unit has some systematic failure or problem. On the contrary, if only the abnormal feature point appears obvious abnormality, and the related parameters in the adjacent device state record units remain normal, then the abnormal feature point is likely to be an incidental abnormality, which may be caused by temporary external interference or transient instability of the device. Through the above contextual correlation analysis processing, the properties of the abnormal feature points can be more accurately determined.

[0059] Step 134: If yes, a quantitative evaluation process is performed on the abnormal degree of the abnormal feature point to generate an abnormal feature descriptor containing abnormal confidence.

[0060] When the abnormal feature point is determined to be a persistent abnormality, a quantitative evaluation process is needed to evaluate its abnormal degree in order to more accurately describe and analyze the abnormal situation of the equipment. An appropriate quantitative evaluation algorithm (such as an evaluation algorithm based on statistical analysis and machine learning) is used to evaluate the abnormal degree of the abnormal feature point in combination with historical data and current time-frequency domain feature information. For the abnormal feature point of the refrigeration unit, such as the abnormal fluctuation of the compressor operating current, the evaluation algorithm considers factors such as the deviation from the normal operating range, the duration, and the changes in related parameters. For example, if the abnormal fluctuation value of the compressor operating current is far beyond the normal operating range and lasts for a long time, and the condenser temperature and refrigerant pressure also show a large abnormal change, the evaluation algorithm will give a higher quantitative value of the abnormal degree. According to the quantitative value, an abnormal feature descriptor containing abnormal confidence is generated. Abnormal confidence represents the confidence of the judgment that the abnormal feature point is indeed abnormal. For example, through evaluation, an abnormal feature descriptor is obtained, which contains an abnormal degree quantitative value of 80 (the value range is 0-100) and an abnormal confidence of 90%, indicating that there is a 90% confidence that the abnormal feature point represents the real abnormal situation of the refrigeration unit, and the abnormal degree is high. The above abnormal feature descriptor can clearly convey the severity and confidence information of the abnormality.

[0061] Step 135: Perform association mapping processing of the abnormal feature descriptor and the timestamp information of the device state recording unit to generate an abnormal identification result of the device state recording unit.

[0062] It can be understood that the generated abnormal feature descriptor containing abnormal confidence is associated with the timestamp information of the device state recording unit. For each device state recording unit of the refrigeration unit, if there is an abnormal feature point and a corresponding abnormal feature descriptor is generated, the descriptor is bound to the timestamp corresponding to the device state recording unit. For example, a device state recording unit corresponds to a timestamp of 10 am, and its abnormal feature descriptor shows an abnormal degree quantitative value of 75 and an abnormal confidence of 85%, then the two information are associated. Through the above association mapping processing, an abnormal identification result of the device state recording unit is generated, which not only contains information about whether the device has an abnormality, but also clearly shows the time of the abnormality and the degree and confidence of the abnormality. For example, the device management personnel can quickly locate the time point of the abnormality according to the abnormal identification result, and decide what measures to take to handle the abnormal situation in combination with the degree and confidence information of the abnormality.

[0063] Step 140: determining the abnormal type of the target building equipment and the distribution characteristic information of the abnormal type in the time dimension according to the abnormal identification result.

[0064] Optionally, according to the abnormal identification result of the generated refrigeration unit equipment state record unit, the abnormal type of the target building equipment and the distribution characteristic information of the abnormal type in the time dimension are further analyzed and determined. This step can provide more targeted guidance for the maintenance and management of the equipment. Through detailed analysis of the abnormal identification result, different abnormal conditions can be classified to determine the abnormal type to which they belong. At the same time, by counting and analyzing abnormal conditions at different time points, the distribution characteristics of the abnormal type in the time dimension can be mastered, so that preventive and response measures can be taken in advance.

[0065] As an implementation manner, the determining of the abnormal type of the target building equipment and the distribution characteristic information of the abnormal type in the time dimension according to the abnormal identification result comprises: Step 141: analyzing the abnormal feature descriptor in the abnormal identification result, and extracting the time-frequency domain feature mode corresponding to the abnormal feature descriptor.

[0066] The abnormal feature descriptor in the abnormal identification result of the refrigeration unit is analyzed in detail. Taking an abnormal feature descriptor as an example, it contains an abnormal degree quantitative value, an abnormal confidence and some feature information related to the abnormality. By analyzing these information, the time-frequency domain feature mode corresponding to the abnormal feature descriptor is extracted. For example, the abnormal feature descriptor mentions the abnormal fluctuation of the compressor working current, and combined with the related time-frequency domain feature information, it is found that the abnormal fluctuation shows rapid rise and fall of the current value in the time domain, and abnormal increase of energy in a certain specific frequency range in the frequency domain. Combining these time domain and frequency domain features forms the time-frequency domain feature mode corresponding to the abnormal feature descriptor. This time-frequency domain feature mode can accurately depict the performance of the abnormal condition in the time-frequency domain.

[0067] Step 142: matching the time-frequency domain feature mode with a preset abnormal type feature library to determine the abnormal type corresponding to the abnormal feature descriptor.

[0068] The extracted time-frequency domain feature pattern is compared and matched with a preset abnormal type feature library. The abnormal type feature library stores time-frequency domain feature patterns of various known abnormal types. For a refrigeration unit, the abnormal type feature library can include feature patterns of different abnormal types such as compressor failure, condenser failure, refrigerant leakage, etc. The similarity between the current extracted time-frequency domain feature pattern and each abnormal type feature pattern in the library is calculated. For example, a similarity matching algorithm (such as a cosine similarity algorithm) is used to calculate the similarity score between the current time-frequency domain feature pattern and a certain abnormal type feature pattern in the library. If the similarity score of a certain abnormal type feature pattern exceeds a preset threshold (e.g. 0.8), it is determined that the abnormal type corresponding to the abnormal feature descriptor is the matched abnormal type. For example, it is found through matching that the current time-frequency domain feature pattern has a high similarity to the feature pattern of compressor failure, exceeding the threshold, so it can be determined that the abnormal type corresponding to the abnormal feature descriptor is compressor failure. Through the above matching process, the abnormal situation can be quickly and accurately classified into known abnormal types.

[0069] Step 143: Extract the timestamp information of the abnormal feature descriptor in the abnormal recognition result, and count the occurrence frequency of each abnormal type in different time intervals.

[0070] The timestamp information corresponding to each abnormal feature descriptor is extracted from the abnormal recognition result of the refrigeration unit. Then, the time is divided according to a certain time interval (e.g. by day, by hour, etc.). Taking division of time interval by day as an example, the occurrence frequency of each abnormal type in a day is counted. For each day, all abnormal recognition results are traversed to find all abnormal feature descriptors that occur in that day, and are classified and counted according to their corresponding abnormal types. For example, in a certain day, it is found that 3 abnormal feature descriptors correspond to the abnormal type of compressor failure, and 2 abnormal feature descriptors correspond to the abnormal type of condenser failure. Through the above statistical method, the occurrence frequency of different abnormal types in different time intervals can be clearly understood.

[0071] Step 144: Time series analysis is performed on the occurrence frequency of the abnormal type to identify the active period and the non-active period of the abnormal type in the time dimension.

[0072] Optionally, time series analysis algorithm (such as sliding average algorithm, autoregressive moving average algorithm, etc.) is used to analyze and process the time series of the occurrence frequency of different abnormal types obtained by statistics. Taking the compressor fault abnormal type as an example, the average occurrence frequency in each time window is calculated by the sliding average algorithm. The time window is 3 days, and the 3-day sliding average of the compressor fault abnormal occurrence frequency is calculated every day. Then, the average occurrence frequency difference of adjacent time windows is compared. If the average occurrence frequency of a certain time window is significantly higher than that of the previous time window and exceeds a certain threshold (for example, 1.5 times of the average value), the starting point of the active period is identified; if the average occurrence frequency is significantly lower than that of the previous time window and lower than a certain threshold (for example, 0.5 times of the average value), the end point of the active period is identified. In the non-active period, the time interval with an occurrence frequency lower than the preset frequency (for example, less than 1 time per day) is counted as the non-active period. The boundaries of the preliminary determined active period and non-active period are smoothed, for example, linear interpolation method is used to eliminate the period division error caused by accidental fluctuations, and the smoothed active period and non-active period are obtained. The smoothed active period and non-active period are associated with the compressor fault abnormal type and stored to form the active and non-active period information of the abnormal type in the time dimension.

[0073] As an implementation manner, the time series analysis processing on the occurrence frequency of the abnormal type, identifying the active period and non-active period of the abnormal type in the time dimension, comprises: Step 1440: sliding window statistical processing is performed on the time series of the occurrence frequency of the abnormal type, the average occurrence frequency in each time window is calculated, the average occurrence frequency difference of adjacent time windows is compared, the time point of sudden increase of the occurrence frequency is identified as the starting point of the active period, the time point of sudden decrease of the occurrence frequency is identified as the end point of the active period, the time interval with an occurrence frequency lower than the preset frequency is counted as the non-active period in the non-active period, the boundaries of the active period and non-active period are smoothed to eliminate the period division error caused by accidental fluctuations, and the smoothed active period and non-active period are obtained. The smoothed active period and non-active period are associated with the abnormal type and stored to form the time distribution characteristic information of the abnormal type.

[0074] For the time series of the occurrence frequency of a certain type of abnormality of the refrigeration unit (such as condenser failure), a sliding window size (for example, 5 hours) is set. Starting from the starting point of the time series, the average occurrence frequency in each 5-hour time window is calculated in turn. For example, the condenser failure occurred 2 times in the first 5-hour time window, and 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, the average occurrence frequency of the previous time window is 0.2 times / hour, and the increase exceeds the preset threshold (for example, 0.3 times / hour), the starting time point of the current time window is identified as the starting point of the active period. Conversely, if the average occurrence frequency of the current time window is 0.1 times / hour, the average occurrence frequency of the previous time window is 0.4 times / hour, and the decrease exceeds the preset threshold (for example, 0.2 times / hour), the end time point of the current time window is identified as the end point of the active period. In the non-active period, a frequency threshold (for example, 0.1 times / hour) is preset, and the time interval with an occurrence frequency lower than the threshold is counted. For example, in a certain time period, the occurrence frequency of the condenser failure is always lower than 0.1 times / hour, and the time period is a non-active period. Since there may be accidental fluctuations in the actual data, the boundaries of the initially determined active period and non-active period are smoothed. The linear interpolation method can be used to adjust the boundaries of the active period and non-active period to make them more accurate and reasonable. Finally, the smoothed active period and non-active period are associated with the condenser failure abnormality type and stored to form the time distribution characteristic information of the abnormality type. The above time distribution characteristic information can clearly show the active and non-active situation of the condenser failure in different time periods, facilitating advance maintenance and monitoring arrangements.

[0075] Step 145: constructing a time distribution curve of the abnormality type based on the active period and the non-active period, and generating time distribution characteristic information containing time intervals and corresponding frequencies.

[0076] It can be understood that, according to the active period and non-active period information of various abnormal types of the refrigeration unit determined in the foregoing, a time distribution curve of each abnormal type is constructed. The abscissa represents time, and the ordinate represents the occurrence frequency of the abnormal type. For each abnormal type, the active period and the non-active period are marked on the time axis, and a curve is drawn according to the occurrence frequency in different time intervals. For example, for the compressor fault abnormal type, its active period (such as 10 am-2 pm) and non-active period (such as 8 pm-6 am) are clearly marked on the time axis. In the active period, the corresponding frequency value is found on the ordinate according to the statistical occurrence frequency data, and these points are connected to form part of the curve; in the non-active period, the curve approaches the abscissa because of the low occurrence frequency.

[0077] In the above manner, time distribution characteristic information containing time intervals and corresponding frequencies is generated. The time distribution curve can intuitively show the occurrence frequency variation of each abnormal type at different times, and the device management personnel can quickly understand the time distribution law of the abnormal type by observing the curve, thereby providing a basis for formulating more effective maintenance plans and monitoring strategies.

[0078] Step 150: generating a target early warning instruction containing a time positioning identifier based on the abnormal type and the time distribution characteristic information, and sending the device early warning instruction to a target device management terminal.

[0079] According to the determined abnormal type of the refrigeration unit and the distribution characteristic information thereof in the time dimension, a target early warning instruction containing a time positioning identifier is generated to timely notify the device management terminal to take corresponding measures. The target early warning instruction not only contains the information of the abnormal type, but also clearly indicates the time range in which the abnormality may occur, thereby helping the device management personnel to perform targeted processing. By sending the early warning instruction to the target device management terminal, rapid response and effective management of the device abnormality can be achieved.

[0080] Optionally, the generating of the target early warning instruction containing the time positioning identifier based on the abnormal type and the time distribution characteristic information comprises: Step 151: analyzing a preset early warning rule library corresponding to the abnormal type, and extracting an early warning priority identifier and a maintenance strategy code associated with the abnormal type.

[0081] For a determined abnormal type of the refrigeration unit, such as a compressor failure abnormal type, the system automatically parses a preset warning rule library. The preset warning rule library is a pre-constructed database that stores detailed rule information corresponding to various abnormal types. In this rule library, for the compressor failure abnormal type, the associated warning priority identifier and maintenance strategy code are found. The warning priority identifier is used to indicate the urgency of the abnormal situation. For example, a compressor failure may be set as a high priority because it can seriously affect the normal operation of the refrigeration unit, and in turn affect the overall environmental temperature of the commercial center. The maintenance strategy code is a set of codes that indicate the maintenance measures to be taken for the abnormal type. For example, the maintenance strategy code may include a series of operation steps and requirements to guide the maintenance personnel to carry out effective maintenance and repair work. In this way, important information related to the abnormal type is accurately extracted from the preset warning rule library.

[0082] Step 152: Extract the active period and inactive period in the time distribution feature information to determine the current active time interval of the abnormal type.

[0083] From the time distribution feature information of the generated abnormal type of the refrigeration unit, the active period and inactive period information of the abnormal type (such as compressor failure) are extracted. Then, combined with the current time point, the current active time interval of the abnormal type is determined. For example, the current time is 9 am, and the active period of the compressor failure is 10 am-2 pm, so the current active time interval is determined to be 10 am-2 pm. By determining the current active time interval, the equipment manager can clearly know that the abnormal situation needs to be focused on in what time period, so as to prepare in advance and arrange the corresponding maintenance work.

[0084] Step 153: Calculate the re-occurrence probability of the abnormal type in a specified subsequent period according to the current active time interval.

[0085] Based on the current determined active time interval, the probability of the abnormal type (such as compressor failure) of the refrigeration unit occurring again in a specified subsequent period is calculated using a corresponding probability calculation model. The probability calculation model can be established based on statistical analysis of historical data, and it considers various factors such as the frequency of the abnormal type occurring in the past similar time period, the current operating state of the equipment, environmental factors, etc. For example, by statistically analyzing the occurrence of compressor failures after a plurality of similar active time periods in the past, it is found that if a failure occurs in a certain active time period, the probability of it occurring again in the next 24 hours is 30%. In combination with the actual operating parameters of the current refrigeration unit, maintenance records, and current environmental temperature, etc., the base probability is adjusted. If the operating temperature of the current compressor is high and recent maintenance work has been delayed, the probability of failure occurring again in the specified subsequent period (such as the next 24 hours) may increase to 40% according to the model calculation. Through the above method, the likelihood of the abnormal type occurring again in the future can be accurately estimated, providing a basis for formulating reasonable warning and maintenance strategies.

[0086] Step 154: The occurrence probability is associated and weighted with the warning priority identifier of the abnormal type to generate a dynamic warning level parameter.

[0087] The calculated probability of the abnormal type (such as compressor failure) of the refrigeration unit occurring again in a specified subsequent period is associated and weighted with the warning priority identifier corresponding to the abnormal type. The warning priority identifier represents the urgency of the abnormal situation, and the occurrence probability reflects the likelihood of the abnormality occurring in the future. Through a preset weighting algorithm, these two factors are combined 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 occurrence probability is 40% (i.e. 0.4). The weighting algorithm may provide that the dynamic warning level parameter = warning priority weight x occurrence probability + other adjustment factors (other adjustment factors in the present application are 0), then the dynamic warning level parameter = 0.6 x 0.4 = 0.24 (the dynamic warning level parameter value range is 0-1), which can comprehensively reflect the urgency of the abnormal situation and the likelihood of future occurrence, providing an important basis for generating more accurate and effective warning instructions subsequently. Through the above association and weighting processing, the warning level can be dynamically adjusted according to the actual situation, which is more in line with the actual operation of the equipment.

[0088] Step 155: The dynamic warning level parameter, maintenance strategy code, and current active time interval are information fused to generate a target warning instruction containing a time positioning identifier.

[0089] Optionally, the three key information, i.e., the generated dynamic early warning level parameter, the maintenance strategy code, and the current active time interval, are processed by information fusion. Information fusion is the process of integrating different types of information that are related to each other to form a complete, accurate, and clear target early warning instruction. For example, the dynamic early warning level parameter (such as 0.24), the maintenance strategy code (such as a set of detailed maintenance operation step codes), and the current active time interval (10 am-2 pm) are combined. During the combination process, the generated target early warning instruction is arranged according to certain format and rules, so that it is clear and explicit. For example, the target early warning instruction may be presented in the form of text: “Refrigeration unit compressor fault warning! Dynamic early warning level: 0.24 (high), current active time interval: 10 am-2 pm, please perform the corresponding maintenance operation according to the maintenance strategy code [corresponding code content].” Through the above information fusion process, the generated target early warning instruction contains not only the severity and likelihood of the abnormal situation, but also the time range of the abnormal situation and the corresponding maintenance measures, providing comprehensive and specific guidance for the equipment management terminal.

[0090] Preferably, the information fusion processing of the dynamic early warning level parameter, the maintenance strategy code, and the current active time interval generates a target early warning instruction containing a time positioning identifier, which comprises: Step 1551: Assign a corresponding early warning identifier to the dynamic early warning level parameter, and the early warning identifier is positively correlated with the early warning level parameter.

[0091] The dynamic early warning level parameter (such as 0.24) generated for the refrigeration unit is assigned a corresponding early warning identifier. The early warning identifier is a direct representation method for quickly conveying the severity of the early warning. Generally, the early warning identifier is positively correlated with the early warning level parameter, i.e., the higher the early warning level parameter, the more attention the corresponding early warning identifier can attract. For example, for the early warning level parameter between 0-0.3, the assigned early warning identifier may be a “yellow exclamation mark”; for 0.3-0.6, it may be an “orange triangle”; for 0.6-1, it may be a “red alert” icon. For the dynamic early warning level parameter 0.24, a “yellow exclamation mark” is assigned as the early warning identifier. Thus, when the equipment management personnel receive the early warning instruction, they can quickly understand the approximate severity of the early warning by checking the early warning identifier, without the need to check the specific level parameter in detail, improving the efficiency of information acquisition and processing.

[0092] Step 1552: Convert the maintenance strategy code into a maintenance operation description text containing maintenance step descriptions.

[0093] In detail, the converting the maintenance strategy code into the maintenance operation description text containing maintenance step descriptions comprises: Step 15521: obtaining a multi-section coding structure of the maintenance strategy code, the multi-section coding structure being composed of a maintenance action type identification section, an action object identification section and an operation condition identification section arranged in sequence.

[0094] For the maintenance strategy code of the compressor fault of the refrigeration unit, the multi-section coding structure thereof is obtained first. For example, the maintenance strategy code is "123-45-678", wherein "123" is the maintenance action type identification section, representing the category of the maintenance action needed to be taken for the abnormality, for example, possibly indicating "repairing the compressor motor"; "45" is the action object identification section, explicitly indicating the specific equipment component to which the maintenance action is directed, possibly corresponding to "the winding of the compressor motor"; and "678" is the operation condition identification section, specifying the conditions needed to be met when the maintenance operation is performed, such as "the ambient temperature needs to be between 20-30 degrees Celsius, and the equipment needs to be in the shutdown state". Through the above multi-section coding structure, each key element of the maintenance operation can be accurately defined, providing a clear framework for the subsequent accurate conversion into the maintenance operation description text.

[0095] Step 15522: calling a pre-constructed maintenance action knowledge base, the maintenance action knowledge base storing a basic step description unit corresponding to the maintenance action type identification section one by one, an equipment component positioning description unit corresponding to the action object identification section one by one, and an execution constraint description unit corresponding to the operation condition identification section one by one.

[0096] Optionally, the system calls a pre-built maintenance action knowledge base, which is a rich database storing detailed information related to various maintenance action types, action objects, and operating conditions. For the maintenance action type identification segment "123" in the maintenance strategy code, find the corresponding basic step description unit 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" may be "first, cut off the power, then use professional tools to disassemble the motor housing...". For the action object identification segment "45", find the corresponding device component positioning description unit, which contains component name and spatial location information, such as "the compressor motor winding is located inside the compressor on the left side, connected to other components through the setting interface". For the operating condition identification segment "678", match the corresponding execution constraint description unit, which contains environmental parameter range requirements and safety precautions information, such as "the operating environment temperature should be kept between 20-30 degrees Celsius, and it is necessary to wear good insulation protective equipment to avoid electric shock danger". By calling the maintenance action knowledge base, detailed and accurate maintenance operation related information can be obtained to generate complete maintenance operation description text.

[0097] Step 15523: Segment parsing processing is performed on the multi-segment code structure to extract the code values of the maintenance action type identification segment, the action object identification segment, and the operating condition identification segment.

[0098] The maintenance strategy code "123-45-678" is segmented and parsed to extract the code value "123" of the maintenance action type identification segment, the code value "45" of the action object identification segment, and the code value "678" of the operating condition identification segment. This segmentation and parsing is to accurately obtain the corresponding information from the maintenance action knowledge base and combine these information according to the correct logic to generate the maintenance operation description text that meets the actual needs. By specifying the values of each code segment, the accuracy and effectiveness of the subsequent information matching and text generation process can be ensured.

[0099] Step 15524: Based on the code value of the maintenance action type identification segment, match and obtain the corresponding basic step description unit from the maintenance action knowledge base, which contains standard action verbs and general operation direction information.

[0100] According to the extracted maintenance action type identification segment code value "123", an accurate match is made in the maintenance action knowledge base to obtain the corresponding basic step description unit. As described earlier, the basic step description unit can be "first, cut off the power supply, then use professional tools to disassemble the motor shell, check whether the motor winding has damage signs …", which provides the basic flow and action direction of the maintenance operation and is the core part of building a complete maintenance operation description text. It uses standard action verbs such as "cut off", "disassemble", "check", etc., clearly indicating the operations that the maintenance personnel need to perform, and the general operation direction information also provides the general idea and sequence of the operation for the maintenance personnel.

[0101] Step 15525: Based on the code value of the action object identification segment, the corresponding device component positioning description unit is matched and obtained from the maintenance action knowledge base, which contains component name and spatial position information.

[0102] According to the code value "45" of the action object identification segment, the corresponding device component positioning description unit is found in the maintenance action knowledge base, which details the component name and spatial position information, such as "the compressor motor winding is located inside the compressor on the left side, connected to other components through a set interface". This information is very important for maintenance personnel, as it can help them quickly and accurately find the specific device component that needs maintenance, avoiding the risk of misoperation or failure to find the target component during the operation. By combining the device component positioning description unit with the basic step description unit, the maintenance operation description text can be made more specific and executable.

[0103] Step 15526: Based on the code value of the operation condition identification segment, the corresponding execution constraint description unit is matched and obtained from the maintenance action knowledge base, which contains environmental parameter range requirements and safety precautions information.

[0104] According to the code value "678" of the operation condition identification segment, the corresponding execution constraint description unit is matched and obtained from the maintenance action knowledge base, which contains environmental parameter range requirements and safety precautions information, such as "the operating environment temperature should be kept between 20-30 degrees Celsius, and it is necessary to wear good insulation protective equipment to avoid electric shock danger". These information is an important guarantee to ensure the safety and effectiveness of the maintenance operation. Including the content of the execution constraint description unit in the maintenance operation description text can remind the maintenance personnel to pay attention to the environmental conditions and safety issues during the operation, preventing equipment damage or personnel injury due to not meeting the conditions or neglecting safety matters.

[0105] Step 15527: The standard action verb of the basic step description unit is semantically fused with the component name and spatial location information of the equipment component positioning description unit to generate an action description clause containing an operation object.

[0106] The standard action verb in the basic step description unit is fused with the component name and spatial location information of the equipment component positioning description unit. For example, the action of "cut off power" in the basic step description unit and the indication that the power control switch is located on the right side of the compressor control cabinet in the equipment component positioning description unit. After semantic fusion processing, the generated action description clause containing the operation object may be "find the power control switch on the right side of the compressor control cabinet and cut off the power". Through the above fusion, the action description is more specific and clear, and the maintenance personnel can clearly know which specific component to perform the operation, which improves the accuracy and operability of the maintenance operation description text.

[0107] Step 15528: The action description clause is conditionally associated with the environmental parameter range requirement and safety precautions information of the execution constraint description unit to generate a step constraint description clause containing operation premise conditions.

[0108] The generated action description clause is conditionally associated with the information of the execution constraint description unit. Taking the generated action description clause "find the power control switch on the right side of the compressor control cabinet and cut off the power" and the execution constraint description unit "the ambient temperature during operation should be kept between 20-30 degrees Celsius, and the insulating protective equipment should be worn to avoid electric shock" as an example, the generated step constraint description clause containing operation premise conditions is "find the power control switch on the right side of the compressor control cabinet and cut off the power under the premise that the ambient temperature is kept between 20-30 degrees Celsius and the insulating protective equipment is worn". The above step constraint description clause clearly indicates the premise conditions of the operation, so that the maintenance personnel can fully consider various limiting factors when performing the operation, ensuring the safety and correctness of the operation.

[0109] Step 15529: The action description clause and the step constraint description clause are logically concatenated according to the arrangement order of the multi-section coding structure to generate a maintenance step description sequence with sequential association; the maintenance step description sequence is optimized for natural language fluency to adjust the syntax structure and conjunction usage of the sentence to generate a maintenance operation description text containing complete operation logic.

[0110] According to the arrangement order of the multi-section structure coded according to the maintenance strategy, the generated action description clauses and step constraint description clauses are logically concatenated. For example, there are other related action description clauses and step constraint description clauses, such as “after confirming that the motor winding is damaged, use professional tools to replace the motor winding, and keep the operating environment clean during the replacement process”, and the like. These clauses are connected in a reasonable order to form a maintenance step description sequence with sequential correlation. Then, the maintenance step description sequence is subjected to natural language fluency optimization processing. It is checked whether the grammatical structure of the sentence is correct, whether the conjunctions are used appropriately, and some places where the expression is not clear or fluent are adjusted. For example, conjunctions such as “first … then … and …” are reasonably used to make the entire maintenance operation description text read more smoothly and easily understood. Finally, a maintenance operation description text containing complete operation logic is generated, which can provide detailed, accurate and easy-to-understand operation guidance for maintenance personnel, helping them to successfully complete the maintenance work of the refrigeration unit.

[0111] Step 1553: converting the current active time interval into a time range descriptor, the time range descriptor including a starting time point and an ending time point.

[0112] The current active time interval (such as 10 am-2 pm) of the refrigeration unit abnormal type (such as compressor failure) is converted into a time range descriptor. The time range descriptor is a more standardized and easy-to-understand time representation, which explicitly includes a starting time point and an ending time point. For example, 10 am is converted to “2024-01-01 10:00:00” (the current date is January 1, 2024), and 2 pm is converted to “2024-01-01 14:00:00”. The generated time range descriptor is “2024-01-01 10:00:00 to 2024-01-01 14:00:00”. This time range descriptor can more accurately and clearly convey the time information of the active abnormality in the subsequent generation of warning instructions, avoiding misunderstandings or errors due to unclear time representation.

[0113] Step 1554: structurally combining the warning identifier, the maintenance operation description text, and the time range descriptor to generate a device warning instruction body containing a time positioning identifier.

[0114] The assigned early warning identifier symbol (such as a "yellow exclamation mark"), the generated maintenance operation description text (detailed maintenance step instructions 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. According to certain formats and rules, these information are integrated to form a complete device early warning instruction body. For example, the device early warning instruction body may be 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: Find the power control switch on the right side of the compressor control cabinet under the premise of maintaining the ambient temperature between 20-30 degrees Celsius and wearing appropriate insulation protective equipment, turn off the power…… (follow-up detailed maintenance steps)". Through the above structured combination processing, the generated device early warning instruction body can clearly and comprehensively convey abnormal information, time positioning and maintenance operation requirements, providing strong support for device management.

[0115] Step 1555: Add device identification information to the device early warning instruction body, which is used to uniquely identify the target building device.

[0116] Add device identification information to the generated device early warning instruction body. The refrigeration unit as the target building device has unique identification information, such as device number "L001" and device name "Commercial Center A Refrigeration Unit No. 1". Adding these device identification information to the device early warning instruction body makes the early warning instruction more specific and targeted. The added early warning instruction may be as follows: "Device number: L001, device name: Commercial Center A Refrigeration Unit No. 1 [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: Find the power control switch on the right side of the compressor control cabinet under the premise of maintaining the ambient temperature between 20-30 degrees Celsius and wearing appropriate insulation protective equipment, turn off the power…… (follow-up detailed maintenance steps)". By adding device identification information, the device management terminal can accurately know the specific device that the early warning instruction is directed to, avoiding misoperation or error handling, improving the accuracy and effectiveness of the early warning instruction, and helping device management personnel quickly locate and handle problem devices.

[0117] Step 1556: Perform error checking coding processing on the device early warning instruction body and device identification information to generate a target early warning instruction with error checking function.

[0118] To ensure the accuracy and integrity of the device warning instruction during transmission and reception, the device warning instruction body containing device identification information is subjected to check encoding processing, which can be implemented based on a preset check encoding algorithm (such as a cyclic redundancy check algorithm, CRC algorithm). The algorithm generates a check code based on the contents of the device warning instruction body and the device identification information. Taking the previously generated warning instruction as an example, all information such as device number, device name, warning identifier, time range descriptor, and maintenance operation description text is taken as input, and a check code is calculated by the CRC algorithm. For example, the calculated check code is "12AB". Then, the check code is added to the end of the warning instruction to form the target warning instruction with error checking function: "Device number: L001, device name: Commercial Center A refrigeration unit No. 1 [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: Find the power control switch on the right side in the compressor control cabinet under the premise of keeping the environmental temperature between 20-30 degrees Celsius and wearing good insulation protective equipment, turn off the power … (follow-up detailed maintenance steps) Check code: 12AB". When the target device management terminal receives the warning instruction, it will use the same check encoding algorithm to check the received information. If the calculated check code is consistent with the received check code, it means that the warning instruction has not been corrupted during transmission; if it is not consistent, it means that there may be data loss or errors, and the device management terminal can request to resend the warning instruction, thereby ensuring the accuracy and reliability of the warning instruction.

[0119] Step 156: Format standardization processing is performed on the device warning instruction to obtain a standardized warning instruction; the standardized warning instruction matches the communication protocol requirements of the target device management terminal.

[0120] The generated target warning instruction with error checking function is subjected to format standardization processing to ensure that it can match the communication protocol requirements of the target device management terminal. Different device management terminals may have different communication protocol requirements, including data format, character encoding, message length, etc. For example, the target device management terminal requires that the warning instruction be transmitted in XML format and the character encoding be UTF-8. Further, the target warning instruction is rearranged according to the XML format, and each information element (device number, device name, warning identifier, time range, maintenance operation description, check code, etc.) is organized according to the specified tags and structure. At the same time, the character encoding of the entire instruction is converted to UTF-8.

[0121] After the above format standardization processing, the standardized early warning instruction is obtained, which can meet the communication protocol requirements of the target device management terminal and ensure that it can be correctly recognized and parsed during transmission. When the device management terminal receives the standardized early warning instruction, it can accurately extract the information therein according to its internal parsing mechanism, so as to timely take corresponding measures to handle the abnormal situation of the refrigeration unit.

[0122] As a non-limiting example, after sending the device early warning instruction to the target device management terminal, it further includes: Collecting early warning processing feedback information returned by the target device management terminal, which contains execution state identification of the early warning instruction and subsequent running data records of the corresponding device state; Performing semantic analysis processing on the early warning processing feedback information to extract the processing result label corresponding to the execution state identification and the device state parameter change sequence in the subsequent running data records; Correlating and mapping the processing result label with the corresponding time-frequency domain feature set and abnormal identification result to construct an abnormal sample expansion set containing the processing result dimension; Calling a pre-trained model effect evaluation module to perform model performance analysis processing on the abnormal sample expansion set to identify the false detection rate and the missed detection rate characteristics of the mixed machine learning model under different abnormal types; Generating model parameter adjustment suggestions based on the false detection rate and the missed detection rate characteristics, which include memory cell update strategies of the LSTM neural network and outlier threshold correction strategies of the isolation forest algorithm; inputting the model parameter adjustment suggestions into the parameter optimization interface of the mixed machine learning model to complete the dynamic optimization and update processing of the model.

[0123] After sending the standardized early warning instruction to the target device management terminal, the device management terminal will process the early warning instruction and return early warning processing feedback information. The device management terminal will take corresponding maintenance operations according to the early warning instruction and record the execution state identification, such as "processed", "processing", "processing failed", etc. At the same time, the target device management terminal will also continue to collect the subsequent running data records of the corresponding device state of the refrigeration unit, such as the change of parameters such as compressor working current, condenser temperature, refrigerant pressure, etc. after maintenance operation, and these feedback information is sent back to the system.

[0124] After the system receives the early warning processing feedback information, semantic analysis processing is first performed. For the execution state identifier, it is converted into the corresponding processing result label, for example, "processed" is converted into "successfully processed", and "processing failed" is converted into "processing unsuccessful". For the subsequent running data record, the change sequence of the equipment state parameter therein is extracted, for example, the change value of the compressor working current every minute after the maintenance operation is recorded. Then, the processing result label is associated with the corresponding time-frequency domain feature set and the abnormality identification result generated before, and the association mapping processing is performed. For example, if the processing result label is "successfully processed", and the corresponding abnormality identification result is a compressor fault abnormality, the label of the successful processing is associated with the time-frequency domain feature set corresponding to the abnormality, indicating that the abnormality has caused a corresponding change in the equipment state after processing.

[0125] Through the above association, an abnormal sample expansion set containing a processing result dimension is constructed, which contains more information about the abnormality processing result and can provide more abundant data for model performance analysis. Then, the pre-trained model effect evaluation module is called to perform model performance analysis processing on the abnormal sample expansion set. The module will analyze the false detection rate and the missed detection rate characteristics of the mixed machine learning model under different abnormality types. For example, for the compressor fault abnormality type, it is checked whether the model misdetects the normal running state as an abnormality (false detection rate) and whether it misses the actual existing abnormality (missed detection rate). Based on the false detection rate and the missed detection rate characteristics obtained by the analysis, model parameter adjustment suggestions are generated. For example, if it is found that for the compressor fault abnormality, the false detection rate is high, it may be suggested to adjust the memory cell update strategy of the LSTM neural network so that it can better learn the normal running mode; if the missed detection rate is high, it may be suggested to correct the outlier threshold of the isolation forest algorithm to more accurately identify abnormal points. Finally, these model parameter adjustment suggestions are input into the parameter optimization interface of the mixed machine learning model, and the model will adjust its parameters according to these suggestions to complete the dynamic optimization and update processing.

[0126] In the above manner, the mixed machine learning model can continuously adapt to the actual situation of equipment operation, improve the accuracy and reliability of abnormality detection.

[0127] As a non-limiting embodiment, after the device early warning instruction is sent to the target device management terminal, the method further includes: Continuously acquiring a continuous supplementary running data set of the target building equipment after the early warning instruction is sent, the supplementary running data set containing device state record units with the same timestamp marking rule as the historical continuous running data set; Performing time-frequency domain feature extraction processing on the supplementary running data set to obtain a supplementary time-frequency domain feature set; The supplementary time-frequency domain feature set is input into the hybrid machine learning model for abnormality detection processing to generate a supplementary abnormality recognition result. New abnormality feature patterns that are not matched with the preset abnormality type feature library in the supplementary abnormality recognition result are extracted; the new abnormality feature patterns are subjected to feature clustering processing, and new abnormality type clusters are divided based on a feature similarity measurement rule; A unique abnormality type identifier is assigned to each new abnormality type cluster, and a time-frequency domain feature pattern corresponding to the new abnormality type cluster is extracted as a new abnormality feature descriptor; The new abnormality type identifier and the new abnormality feature descriptor are added to the preset abnormality type feature library to complete dynamic expansion and update processing of the abnormality type feature library.

[0128] After the sending device sends the early warning instruction, the system continuously acquires a continuous supplementary running data set of the refrigeration unit after the early warning instruction is sent. The supplementary running data set follows the same time stamp marking rule as the historical continuous running data set, for example, recording the device state parameters once every minute. The acquisition device will continue to monitor the various parameters of the refrigeration unit in real time, such as the compressor working current, the condenser temperature, the refrigerant pressure, etc., and record them according to the time stamp marking rule to form new device state record units. These units constitute the supplementary running data set. Then, the supplementary running data set is subjected to time-frequency domain feature extraction processing, which is the same as the processing method for the historical data. First, the time window is divided, then the time domain feature subset and the frequency domain feature subset are extracted respectively, and finally the two are associated and matched to obtain a supplementary time-frequency domain feature set. The supplementary time-frequency domain feature set is input into the hybrid machine learning model for abnormality detection processing. The model will analyze these new feature sets according to the learned normal patterns and abnormal patterns to generate a supplementary abnormality recognition result.

[0129] In the supplementary anomaly recognition result, it is carefully checked whether there is a new type of abnormal feature mode that does not match the preset abnormal type feature library. For example, a new compressor operating current fluctuation mode may be found, which does not match any mode in the preset abnormal type feature library. For this new type of abnormal feature mode, feature clustering processing is performed. A feature similarity measurement rule (such as the Euclidean distance measurement rule) is used to cluster new type of abnormal feature modes with similar features together and divide them into different new type of abnormal type clusters. For example, by calculating the Euclidean distance between different new type of abnormal feature modes, modes with closer distances are classified into a cluster. Each new type of abnormal type cluster is assigned a unique abnormal type identifier, such as "abnormal type X1", "abnormal type X2", etc. At the same time, the time-frequency domain feature mode corresponding to each new type of abnormal type cluster is extracted as a new type of abnormal feature descriptor. Finally, the new type of abnormal type identifier and the new type of abnormal feature descriptor are added to the preset abnormal type feature library, so that the abnormal type feature library is dynamically expanded and updated, and can contain more abnormal type information, improving the system's ability to recognize and handle new abnormal situations.

[0130] Over time, new supplementary operation data and new types of abnormalities are discovered and processed, and the abnormal type feature library is continuously updated and improved, so that the system can better adapt to the changing operating conditions of the refrigeration unit.

[0131] It can be understood that in the time domain feature extraction (such as fluctuation amplitude, change rate and periodic fluctuation features), the difference between the maximum value and the minimum value, the ratio of the difference between adjacent time points to the time interval, and the autocorrelation analysis can be realized by the statistical module in Scikit-learn, and the built-in function (such as sklearn.preprocessing processing normalization) is used to automatically eliminate the dimensional difference, and a consistent feature subset is generated. For frequency domain conversion processing, windowing (such as Hanning window) and smoothing filtering can be implemented based on the FFT algorithm of NumPy and SciPy (such as numpy.fft), and the Gaussian filtering function of scipy.signal module is used to automatically perform weighted average to suppress noise, so as to obtain a smooth frequency signal.

[0132] In the anomaly detection link of the mixed machine learning model in the embodiment of the application, the memory unit, the forgetting gate and the output gate of the LSTM neural network can be processed by means of the standard implementation (such as tf.keras.layers.LSTM) of TensorFlow or Keras framework, and the time series dependence is automatically modeled based on the public gating mechanism algorithm, and the outlier detection of the isolation forest algorithm is directly implemented by using the IsolationForest class of Scikit-learn, and the built-in decision tree division strategy is used to efficiently identify abnormal feature points, thereby simplifying manual parameter setting.

[0133] In addition, the probability calculation (such as the probability of abnormal recurrence) can analyze the historical active period data by using the existing time series prediction technology (such as the ARIMA model), automatically output the quantitative value, and dynamically adjust the warning level in combination with the rule base; and the maintenance strategy coding conversion is used to analyze the multi-section structure by using the existing NLP tool (such as a rule-based or pre-trained model) to generate natural language text.

[0134] Therefore, the time-frequency domain feature extraction, model detection, and early warning instruction generation processes can be further optimized, the communication reliability can be guaranteed by using the error checking coding protocol (such as the CRC algorithm), the high-precision abnormality identification can be realized, and thus the device management terminal can receive the standardized early warning instruction and optimize the response, and the building equipment monitoring efficiency can be improved.

[0135] The embodiments of the present application construct a complete and efficient building equipment abnormality identification and early warning system as a whole.

[0136] Firstly, the continuous running data set with timestamp markers is acquired to record the state of the equipment at different times; the time-frequency domain feature extraction processing can deeply mine the data features from the time domain and the frequency domain, overcome the limitations of single-dimensional analysis, more comprehensively depict the equipment running characteristics, and greatly improve the capture ability of abnormal features.

[0137] Secondly, the pre-constructed hybrid machine learning model combines the advantages of multiple algorithms to perform abnormal detection on the time-frequency domain feature set, significantly improves the accuracy and reliability of abnormal detection, and effectively reduces the false detection and missed detection. According to the abnormal identification result, the abnormal type and spatiotemporal distribution feature information are determined, so that the manager can clearly understand the nature and development law of the abnormality, and provide a strong basis for targeted processing.

[0138] Further, the target early warning instruction containing the time positioning identifier is generated based on the above information and sent to the terminal, precise early warning is realized, the manager can know the time and type of the abnormality in advance, so as to arrange the maintenance in time, greatly improve the timeliness and effectiveness of the equipment maintenance, reduce the loss caused by the equipment failure, guarantee the stable and reliable operation of the building equipment, and comprehensively improve the intelligent level of the building equipment management.

[0139] Based on the same inventive concept, the embodiments of the present application also provide a building equipment abnormality identification system. Referring to Figure 2 As shown in the figure, it is a structure schematic diagram of a possible building equipment abnormality identification system provided in the embodiments of the present application, Figure 2In some embodiments, the building equipment anomaly identification system 200 comprises a processor 210 and a memory 220. The memory 220 stores a computer program executable by the processor 210. The processor 210 can execute the steps of the machine learning based building equipment anomaly identification method described above by executing the instructions stored in the memory 220.

[0140] Based on the same inventive concept, the present application provides a computer readable storage medium comprising a computer program, which, when executed on a building equipment anomaly identification system, causes the building equipment anomaly identification system to perform the steps of the machine learning based building equipment anomaly identification method described above. In some possible implementations, various aspects of the machine learning based building equipment anomaly identification method provided by the present application can also be implemented in the form of a program product, which comprises a computer program, which, when executed on a building equipment anomaly identification system, causes the building equipment anomaly identification system to perform the steps of the machine learning based building equipment anomaly identification method described above, for example, the building equipment anomaly identification system can perform the steps shown in FIG. 13. Figure 1

[0141] In the technical solutions related to the above embodiments of the present application, whether it is a multi-dimensional feature comparison calculation or a composite parameter construction, if there are problems caused by significant differences in the number of dimensions, units of dimensions, and semantic meanings of different features, those skilled in the art can fully understand that these differences need to be properly handled based on their professional knowledge and past practical experience, so that the calculation result is accurate and has comparability, and logical confusion, unclear mathematical meaning, etc. are avoided.

[0142] In detail, when facing features with different numbers of dimensions, in order to accurately calculate the similarity, matching degree or feature distance between different features, those skilled in the art can use a variety of strategies, including but not limited to feature selection, feature extraction, kernel function processing.

[0143] When processing the comparison of multi-dimensional features, in order to realize the comparability alignment of the feature space, those skilled in the art can use a variety of existing general technical means, including but not limited to standardization preprocessing, mapping conversion, space projection.

[0144] In the construction process of the composite parameter (such as the loss function value), different parameter items often have different dimensions, and those skilled in the art can use normalization processing or adaptive weight distribution mechanism based on distribution characteristics.

[0145] ​The above general technical means for solving the feature matching and loss balancing problems all belong to the common knowledge in the art. These technical means have been fully verified and widely used in a large number of practical applications, and the person skilled in the art can skillfully and flexibly use these methods to deal with similar dimensional difference problems.

[0146] The formulas and calculation processes involved in the embodiments of the present application, whether for multi-dimensional feature comparison or composite loss function construction, strictly follow the dimensional correspondence principle. The variables in each formula have a clear and explicit physical meaning, and the operation logic is completely consistent with the basic mathematical and physical logic, and the operation result is necessarily the reasonable result expected by the present application. The person skilled in the art has the ability to comprehensively use the above general technical means according to the specific data situation and business requirements to effectively solve various problems caused by the number of dimensions, dimensional differences and the like in the multi-dimensional feature comparison calculation and the composite loss function construction in the embodiments, and to ensure the accuracy, reliability and implementability of the technical solutions of the present application.

Claims

1. A method for identifying anomalies in building equipment based on machine learning, characterized in that, The method comprises: obtaining a continuous operation data set of a target building equipment, the continuous operation data set containing a plurality of equipment state record units marked with time stamps; performing time-frequency domain feature extraction processing on the continuous operation data set to obtain a time-frequency domain feature set of the equipment state record units; calling a pre-constructed hybrid machine learning model to perform anomaly detection processing on the time-frequency domain feature set to generate an anomaly identification result of the equipment state record units; determining an anomaly type of the target building equipment and distribution feature information of the anomaly type in the time dimension according to the anomaly identification result; generating a target early warning instruction containing a time positioning identifier based on the anomaly type and the time distribution feature information, and sending the equipment early warning instruction to a target equipment management terminal. 2.The machine learning based building device anomaly identification method of claim 1, wherein, The time-frequency domain feature extraction processing on the continuous operation data set to obtain the time-frequency domain feature set of the equipment state record units comprises: performing time window division processing on the continuous operation data set to obtain a plurality of data segment units with continuous time sequence relationship, each data segment unit corresponding to a fixed length of equipment state record; performing time domain feature extraction processing on each data segment unit to extract the fluctuation amplitude feature, the change rate feature and the periodic fluctuation feature of the equipment state parameter in the data segment unit as a time domain feature subset; performing frequency domain conversion processing on each data segment unit to convert the time sequence signal of the data segment unit into a frequency domain signal representation; performing energy distribution analysis processing on the frequency domain signal representation to extract the energy proportion feature, the main frequency component feature and the frequency component stability feature of different frequency intervals as a frequency domain feature subset; performing association matching processing on the time domain feature subset and the features of the corresponding data segment unit in the frequency domain feature subset to generate a time-frequency domain feature set of each equipment state record unit. 3.The machine learning based building device anomaly identification method of claim 2, wherein, The time domain feature extraction processing on each data segment unit to extract the fluctuation amplitude feature, the change rate feature and the periodic fluctuation feature of the equipment state parameter in the data segment unit as a time domain feature subset comprises: calculating the difference between the maximum value and the minimum value of the equipment state parameter in the data segment unit to obtain the fluctuation amplitude feature of the data segment unit; calculating the ratio of the difference value of the equipment state parameter at adjacent time points in the data segment unit to the time interval to obtain the change rate feature of the data segment unit; performing autocorrelation analysis processing on the sequence of the equipment state parameter of the data segment unit to identify the mode period length repeatedly appearing in the sequence to obtain the periodic fluctuation feature of the data segment unit; performing normalization processing on the fluctuation amplitude feature, the change rate feature and the periodic fluctuation feature to eliminate the dimensional difference between different parameters; combining the normalized fluctuation amplitude feature, change rate feature and periodic fluctuation feature into the time domain feature subset of the data segment unit; The frequency domain conversion processing on each data segment unit to convert the time sequence signal of the data segment unit into a frequency domain signal representation comprises: Windowing processing is performed on the time series signal of the data segmentation unit to suppress the spectral leakage phenomenon of the signal edge, to obtain a windowed time series signal; The windowed time series signal is processed based on a fast Fourier transform algorithm to convert the time domain signal into a frequency domain signal; The real part and imaginary part information of the frequency domain signal is extracted to generate a frequency distribution list containing frequency values and corresponding amplitude values; The frequency distribution list is subjected to smoothing filter processing to eliminate high-frequency noise interference, to obtain a smoothed frequency domain signal representation; The smoothed frequency domain signal representation is stored in association with the timestamp information of the data segmentation unit to form a frequency domain signal representation having a time-frequency correspondence. 4.The machine learning based building device anomaly identification method of claim 1, wherein, The pre-constructed hybrid machine learning model is called to perform anomaly detection processing on the time-frequency domain feature set to generate an anomaly recognition result of the device state record unit, including: 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 having time sequence association information; The sequence feature representation is input into the isolation 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 mode; Contextual association analysis is performed on the abnormal feature points in combination with the feature information of adjacent device state record units in the time-frequency domain feature set to determine whether the abnormal feature points are persistent anomalies or sporadic anomalies; If so, the abnormal degree of the abnormal feature points is quantitatively evaluated to generate an abnormal feature descriptor containing an anomaly confidence; The abnormal feature descriptor is associated with the timestamp information of the device state record unit to generate an anomaly recognition result of the device state record unit. 5.The machine learning based building device anomaly identification method of claim 4, wherein, 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 having time sequence association information, including: The time-frequency domain feature set is subjected to feature dimension alignment processing to unify the number of feature dimensions of different device state record 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 an initial feature input vector; The initial feature input vector is subjected to historical information memory processing by the memory unit of the LSTM neural network to retain the feature information of the previous device state record unit to obtain historical information; The historical information is filtered by the forgetting gate of the LSTM neural network to filter noise historical feature information to obtain filtered historical information; The filtered historical information and the current feature input vector are fused by the output gate of the LSTM neural network to generate a sequence feature representation containing time sequence association information. 6.The machine learning based building device anomaly identification method of claim 1, wherein, The abnormal type of the target building device and the distribution feature information of the abnormal type in the time dimension are determined according to the anomaly recognition result, including: Analyzing an abnormal feature descriptor in the abnormality recognition result, extracting a time-frequency domain feature mode corresponding to the abnormal feature descriptor; Matching the time-frequency domain feature mode with a preset abnormal type feature library to determine an abnormal type corresponding to the abnormal feature descriptor; Extracting timestamp information of the abnormal feature descriptor in the abnormality recognition result, and counting occurrence frequencies of each abnormal type in different time intervals; Performing time series analysis on the occurrence frequencies of the abnormal types to identify active and inactive time periods of the abnormal types in a time dimension; Constructing a time distribution curve of the abnormal type based on the active and inactive time periods, and generating time distribution feature information including time intervals and corresponding frequencies; The time series analysis on the occurrence frequencies of the abnormal types to identify active and inactive time periods of the abnormal types in a time dimension, includes: Performing sliding window statistical processing on the time series of the occurrence frequencies of the abnormal types, and calculating average occurrence frequencies in each time window; Comparing the average occurrence frequencies of adjacent time windows to identify a time point of sudden increase in occurrence frequency as a starting point of an active period, and a time point of sudden decrease in occurrence frequency as an ending point of the active period; In the inactive period, counting time intervals with occurrence frequencies lower than a preset frequency as the inactive period; Performing boundary smoothing processing on the active and inactive time periods to eliminate period division errors caused by accidental fluctuations, and obtaining smoothed active and inactive time periods; Storing the smoothed active and inactive time periods in association with the abnormal type to form the time distribution feature information of the abnormal type. 7.The machine learning based building device anomaly identification method of claim 1, wherein, The target early warning instruction including time positioning identification is generated based on the abnormal type and the time distribution feature information, including: Analyzing a preset early warning rule library corresponding to the abnormal type, and extracting an early warning priority identifier and a maintenance strategy code associated with the abnormal type; Extracting the active and inactive time periods in the time distribution feature information to determine a current active time interval of the abnormal type; Calculating a re-occurrence probability of the abnormal type in a specified subsequent period according to the current active time interval; Performing associated weighting processing on the re-occurrence probability and the early warning priority identifier of the abnormal 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 including time positioning identification; Performing format standardization processing on the device early warning instruction to obtain a standardized early warning instruction; the standardized early warning instruction matches the communication protocol requirements of the target device management terminal. 8.The machine learning based building device anomaly identification method of claim 7, wherein, The 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 including time positioning identification, includes: Assigning a corresponding early warning identifier to the dynamic early warning level parameter, and the early warning identifier is positively correlated with the early warning level parameter; Converting the maintenance strategy code into a maintenance operation description text including maintenance step descriptions; and convert the current active time interval into a time range descriptor, the time range descriptor comprising a start time point and an end time point; structurally combine the pre-warning identifier, the maintenance operation description text and the time range descriptor to generate a device pre-warning instruction body comprising a time positioning identifier; add device identification information to the device pre-warning instruction body, the device identification information being used to uniquely identify the target building device; perform error checking coding on the device pre-warning instruction body and the device identification information to generate a target pre-warning instruction with error checking function. 9.The machine learning based building device anomaly identification method of claim 8, wherein, The converting the maintenance strategy code into a maintenance operation description text comprising maintenance step descriptions includes: obtaining a multi-segment coding structure of the maintenance strategy code, the multi-segment coding structure being composed of a maintenance action type identifier segment, an action object identifier segment and an operation condition identifier segment arranged in sequence; calling a pre-constructed maintenance action knowledge base, the maintenance action knowledge base storing a basic step description unit corresponding to the maintenance action type identifier segment, a device component positioning description unit corresponding to the action object identifier segment, and an execution constraint description unit corresponding to the operation condition identifier segment; performing segmented analysis processing on the multi-segment coding structure to extract coding values of the maintenance action type identifier segment, the action object identifier segment and the operation condition identifier segment; based on the coding value of the maintenance action type identifier segment, matching and obtaining the corresponding basic step description unit from the maintenance action knowledge base, the basic step description unit comprising a standard action verb and general operation direction information; based on the coding value of the action object identifier segment, matching and obtaining the corresponding device component positioning description unit from the maintenance action knowledge base, the device component positioning description unit comprising component name and spatial position information; based on the coding value of the operation condition identifier segment, matching and obtaining the corresponding execution constraint description unit from the maintenance action knowledge base, the execution constraint description unit comprising environmental parameter range requirement and safety precautions information; performing semantic fusion processing on the standard action verb of the basic step description unit and the component name and spatial position information of the device component positioning description unit to generate an action description clause comprising an operation object; performing conditional association processing on the action description clause and the environmental parameter range requirement and safety precautions information of the execution constraint description unit to generate a step constraint description clause comprising an operation prerequisite condition; performing logical concatenation processing on the action description clause and the step constraint description clause in the arrangement order of the multi-segment coding structure to generate a maintenance step description sequence with sequential association; performing natural language fluency optimization processing on the maintenance step description sequence to adjust the grammatical structure and conjunction usage of the sentences to generate a maintenance operation description text comprising complete operation logic.

10. A building equipment abnormality identification system characterized by comprising: It comprises a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method in any one of claims 1-9.

Images