Heating and ventilation system early-stage fault identification method based on compressor operation characteristic mutation

Abstract

The invention discloses a heating and ventilation system early-stage fault identification method based on compressor operation characteristic mutation, and relates to the technical field of compressor fault identification control. According to the compressor operation characteristic mutation-based early fault identification method for the heating and ventilation system, operation state data and environment working condition data of a compressor are obtained and input into a pre-constructed multi-time scale dynamic reference model, a self-adaptive operation interval set is analyzed, and self-adaptive residual error enhancement processing is performed in combination with the operation state data, so that the early fault of the heating and ventilation system is identified. Obtaining an operation residual sequence, and performing projection mapping processing to obtain operation residual feature points; according to the method, grading identification operation control processing is carried out on the compressor through the operation abnormity deviation evaluation value, so that weak changes caused by normal working condition fluctuation and early faults are distinguished, the possibility of misjudgment and missed judgment is reduced, the early faults are found in time, and the reliability of the compressor is improved. And the operation reliability of the heating and ventilation system is improved.

Description

Technical Field

[0001] This invention relates to the field of compressor fault identification and control technology, specifically to a method for early fault identification of HVAC systems based on sudden changes in compressor operating characteristics. Background Technology

[0002] Heating, ventilation, and air conditioning (HVAC) systems are an indispensable core supporting facility in modern buildings, widely used in various scenarios such as residential buildings, commercial buildings, and industrial plants. Their operational stability directly affects indoor environmental comfort, building energy consumption levels, and safety. As the core power component of the HVAC system, the compressor undertakes the key functions of refrigerant compression, circulation drive, and energy conversion, essentially acting as the heart of the system. The stability of its operating status directly determines the operating efficiency and service life of the entire HVAC system. Furthermore, it is a core control target for smart city AI-optimized low-carbon HVAC and intelligent operation and maintenance platforms to achieve precise control, energy saving, and consumption reduction.

[0003] With the advancement of smart city construction, AI-optimized low-carbon HVAC and intelligent operation and maintenance platforms are gradually becoming more widespread. This places higher demands on the accurate monitoring of compressor operating status and early fault identification. These platforms must not only adapt to the intelligent and low-carbon operation needs of the platform, but also provide reliable data support for the platform's operation and maintenance decisions, ensuring the continuous and stable operation of the HVAC system and aligning with the green, low-carbon, and intelligent management of smart cities.

[0004] The limitations of existing technologies include at least the following problems: When identifying early faults in HVAC system compressors, existing technologies do not combine real-time compressor operating condition data to construct an adaptive operating range, nor do they utilize the entire process of adaptive residual enhancement, residual feature point projection mapping, evolution trajectory monitoring and mutation identification, and hierarchical operation control. This makes it difficult to accurately adapt to the dynamic changes in environmental condition data and operating status data during compressor operation, resulting in a disconnect between the operating baseline and the real-time operating status of the compressor. Either the changes in operating parameters caused by environmental condition fluctuations are misjudged as faults, or it is difficult to capture subtle operational feature mutations caused by early compressor faults, making it difficult to accurately obtain the operational anomaly deviation evaluation value, and consequently, making it difficult to achieve accurate identification and corresponding hierarchical control of early compressor faults. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an early fault identification method for HVAC systems based on sudden changes in compressor operating characteristics. This method solves the problem that existing technologies struggle to adaptively match operating conditions, making it difficult to detect subtle faults and resulting in inaccurate compressor fault identification.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an early fault identification method for HVAC systems based on abrupt changes in compressor operating characteristics, comprising the following steps: acquiring compressor operating condition data, including operating status data and environmental condition data; inputting the operating condition data into a pre-constructed multi-timescale dynamic benchmark model to analyze the compressor's adaptive operating interval set; performing adaptive residual enhancement processing on the operating status data and the adaptive operating interval set to obtain the compressor's operating residual sequence; performing projection mapping processing on the operating residual sequence to obtain the compressor's operating residual feature points; performing evolution trajectory monitoring and abrupt change identification processing on the operating residual feature points to obtain the compressor's operating abnormal deviation evaluation value; and performing graded identification and operation control processing on the compressor based on the operating abnormal deviation evaluation value.

[0007] Further, the specific steps for analyzing the adaptive operating range set of the compressor are as follows: Based on the multi-timescale dynamic benchmark model, multi-scale feature mapping and benchmark prediction processing are performed on the operating condition data to generate a multi-scale operating benchmark set of the compressor; the multi-scale operating benchmark set is fused and optimized to obtain the adaptive operating range set of the compressor.

[0008] Furthermore, the multi-timescale dynamic benchmark model includes an input matching layer, a multi-scale prediction layer, and a confidence output layer. The specific steps for generating the multi-scale operating benchmark set of the compressor are as follows: In the input matching layer, the operating condition data is matched with the historical operating condition dataset stored in the database to extract the reference operating sequence data of the compressor; in the multi-scale prediction layer, the reference operating sequence data is processed by multi-scale sliding slicing and time series mining to generate the operating prediction set of the compressor; in the confidence output layer, the operating prediction set is processed by multi-scale confidence interval construction to generate the multi-scale operating benchmark set of the compressor.

[0009] Further, the specific steps for obtaining the compressor's operating residual sequence are as follows: Dynamic deviation processing is performed on the operating state data and the adaptive operating interval set to obtain the compressor's initial operating residual sequence; based on the environmental operating condition data, operating condition interference correction processing is performed on the initial operating residual sequence to obtain the compressor's operating residual sequence.

[0010] Further, the specific steps for obtaining the compressor's operating residual feature points are as follows: obtain the historical operating residual time series sequence of the compressor, and construct the compressor's operating residual time series data by combining the operating residual sequence; perform feature extraction processing on the operating residual time series data to extract the compressor's residual feature set, including the residual amplitude feature set, the residual change rate feature set, and the residual correlation feature set; perform spatial fusion processing on the residual feature set to obtain the compressor's operating residual feature points.

[0011] Further, the specific steps for extracting the residual feature set of the compressor are as follows: perform amplitude domain statistical processing on the operating residual time series data to extract the residual amplitude feature set of the compressor; perform trend mining processing on the operating residual time series data to extract the residual change rate feature set of the compressor; and perform correlation domain coupling processing on the operating residual time series data to extract the residual correlation feature set of the compressor.

[0012] Furthermore, the specific steps of spatial fusion processing are as follows: performing condition mapping processing on the environmental operating data to generate a residual weight set for the compressor; performing fusion and coordinate mapping processing on the residual weight set and the residual feature set to generate operating residual feature points for the compressor.

[0013] Furthermore, the specific steps to obtain the compressor's abnormal deviation from the evaluation value are as follows:

[0014] The compressor's operating residual feature points are continuously acquired at multiple acquisition times to construct its operating evolution trajectory. Trend offset processing is performed on the operating evolution trajectory to extract the compressor's operating offset evaluation set. Spatial distribution mutation identification processing is performed on the operating evolution trajectory to extract the compressor's density mutation evaluation set. The operating offset evaluation set and the density mutation evaluation set are fused and evaluated to generate the compressor's operating abnormal deviation evaluation value.

[0015] Furthermore, the specific steps of the fusion evaluation process are as follows: normalize the operation deviation evaluation set and the density mutation evaluation set; obtain historical fault sample data to determine the fusion optimization weight set of the compressor; and perform comprehensive analysis based on the fusion optimization weight set and the normalization results to obtain the compressor's operation abnormality deviation evaluation value.

[0016] Furthermore, the specific steps for graded identification and operation control processing of the compressor are as follows: compare and analyze the operational abnormality deviation evaluation value with the preset operational abnormality deviation evaluation interval set; based on the comparison and analysis results, identify the compressor and take corresponding control measures.

[0017] The present invention has the following beneficial effects:

[0018] (1) The method for early fault identification of HVAC system based on compressor operating characteristic mutations collects data related to compressor operating status and environmental conditions, and analyzes them in combination with a multi-timescale dynamic benchmark model. It can automatically adjust the operating reference range under different operating conditions. In the process of generating residual sequence, it also removes the interference brought by the environment and retains only the information that truly reflects the change in the equipment's own state, making the extracted residual features more stable. This allows for a clear distinction between normal operating condition fluctuations and the slight changes brought by early faults, reducing the possibility of misjudgment and missed judgment. It enables the compressor to be detected in time when abnormal signs appear, thus improving the reliability of the entire HVAC system operation.

[0019] (2) The method for early fault identification of HVAC system based on compressor operating characteristic mutations extracts multi-dimensional features from the operating residuals to comprehensively characterize the equipment status. Then, it integrates multiple features into unified operating residual feature points through spatial fusion. By continuously tracking the changes in operating residual feature points, it can simultaneously monitor the state deviation trend and spatial distribution mutations. Combined with historical data, it completes the fusion evaluation to obtain accurate abnormal deviation results. Based on the results, it automatically executes hierarchical control, provides timely warnings and adjusts control when there are minor abnormalities, and initiates protection measures when the risk is high, effectively reducing the risk of equipment damage and subsequent maintenance costs, and improving the overall operation and maintenance efficiency.

[0020] (3) The method for early fault identification of HVAC system based on compressor operating characteristic mutations can capture both rapidly occurring instantaneous mutations and slowly developing progressive anomalies by adopting a multi-timescale analysis method. It has good identification ability for various early faults that occur in the early stage of compressor. At the same time, through steps such as operating condition mapping, weight adjustment and interference correction, it can automatically adapt to changes in environmental conditions and always maintain the accuracy of fault judgment. It can be integrated into the smart city low-carbon HVAC and intelligent operation and maintenance platform, and achieve more stable and energy-saving operation goals while improving the safety level of equipment.

[0021] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0022] Figure 1 This is a flowchart of the early fault identification method for HVAC systems based on sudden changes in compressor operating characteristics, as presented in this invention.

[0023] Figure 2 This is a flowchart illustrating the specific steps involved in analyzing the adaptive operating range set of the compressor in the early fault identification method for HVAC systems based on sudden changes in compressor operating characteristics, as described in this invention.

[0024] Figure 3This is a flowchart illustrating the specific steps involved in obtaining the compressor's abnormal operation deviation assessment value in the early fault identification method for HVAC systems based on sudden changes in compressor operating characteristics, as described in this invention. Detailed Implementation

[0025] Please see Figure 1 This invention provides a technical solution: an early fault identification method for HVAC systems based on abrupt changes in compressor operating characteristics, comprising the following steps: acquiring compressor operating condition data, including operating status data and environmental condition data (of the HVAC system); inputting the operating condition data into a pre-constructed multi-timescale dynamic benchmark model to analyze the compressor's adaptive operating interval set; performing adaptive residual enhancement processing on the operating status data and the adaptive operating interval set to obtain the compressor's operating residual sequence; performing projection mapping processing on the operating residual sequence to obtain the compressor's operating residual feature points; performing evolution trajectory monitoring and abrupt change identification processing on the operating residual feature points to obtain the compressor's operating abnormal deviation evaluation value; and performing graded identification and operation control processing on the compressor based on the operating abnormal deviation evaluation value.

[0026] Specifically, such as Figure 2 As shown, the operating status data includes the instantaneous values ​​of the compressor's three-phase current, the instantaneous values ​​of the compressor's input power, the dynamic values ​​of the compressor's suction pressure, the dynamic values ​​of the compressor's discharge pressure, and three-dimensional data of the compressor casing vibration acceleration (collecting vibration amplitude values ​​in the X, Y, and Z directions). The acquisition of this operating status data can be achieved by installing current sensors, power sensors, pressure sensors, and three-dimensional vibration sensors at corresponding locations on the compressor. The environmental condition data includes outdoor temperature and humidity data (outdoor ambient temperature and relative humidity), the HVAC system's cooling / heating load rate (the cooling or heating output demand of the HVAC system), and the dynamic value of the HVAC system's pipe network resistance (collecting the flow resistance of the medium within the pipe network, reflecting the impact of the pipe network conditions on the compressor's operation). Among these environmental condition data, the outdoor temperature and humidity data can be collected by temperature and humidity sensors installed near the outdoor unit of the HVAC system; the cooling / heating load rate can be accessed in real time by the HVAC system's controller; and the dynamic value of the pipe network resistance can be calculated using flow sensors and pressure sensors installed within the pipe network, i.e., the pressure difference between the pipe network inlet and outlet / the flow rate of the pipe network medium. The specific steps for analyzing the compressor's adaptive operating range set are as follows:

[0027] Based on the multi-timescale dynamic benchmark model, multi-scale feature mapping and benchmark prediction processing are performed on the operating condition data to generate a multi-scale operating benchmark set for the compressor.

[0028] The multi-scale operating benchmark set is fused and optimized to obtain the compressor's adaptive operating interval set. Specifically, the initial fusion weights are preset as follows: short-time scale weights. =0.4 (Focusing on capturing transient, weak changes in early faults to avoid missed detections), mid-time scale weight =0.35 (to capture steady-state fluctuations and avoid misjudgments due to instantaneous interference), long-term scale weight =0.25 (adapting to the gradual changes in environmental operating conditions, correcting for baseline deviation), and satisfying And its preset process is as follows:

[0029] Collect multi-scale operating baseline data of the compressor for at least 30 days during its historical normal operation, including short-term, medium-term, and long-term forecast values ​​for each moment, as well as the corresponding actual operating values. The short-term forecast values ​​correspond to the prediction results at the second scale, the medium-term forecast values ​​correspond to the prediction results at the hour scale, and the long-term forecast values ​​correspond to the prediction results at the day scale.

[0030] Calculate the prediction error of each scale relative to the actual value. For each operating parameter, subtract the short-term prediction value from the actual operating value and take the absolute value to obtain the short-term prediction error. Similarly, subtract the medium-term prediction value from the actual operating value and take the absolute value to obtain the medium-term prediction error. Subtract the long-term prediction value from the actual operating value and take the absolute value to obtain the long-term prediction error.

[0031] The distribution characteristics of errors at various scales under different operating conditions are statistically analyzed. Historical data are classified according to the type of operating condition, such as steady-state operating condition, load change operating condition, start-stop operating condition, high temperature operating condition, low temperature operating condition, etc. For each type of operating condition, the mean of short-term prediction error, the mean of medium-term prediction error, the mean of long-term prediction error, and the corresponding standard deviation are calculated for all samples under that operating condition.

[0032] Based on the above error statistics, the entropy weight method is used to determine the basic weights of each scale under different working conditions. For a certain type of working condition, the error sequence of each scale is regarded as an evaluation index. First, the entropy value of each index is calculated. The smaller the entropy value, the smaller the uncertainty of the index and the greater the amount of information it contains, so it should be assigned a higher weight. Then, the weight is calculated based on the entropy value of each index, that is, by subtracting the entropy value from 1 and then dividing by the sum of all indices (1 minus the entropy value), to obtain the weight of each scale.

[0033] By weighting and averaging the weights of various operating conditions, the global initial weights are obtained. In this embodiment, after analyzing and calculating a large amount of historical data, the initial values ​​are determined as follows: short-term scale weight 0.4, medium-term scale weight 0.35, and long-term scale weight 0.25. This weight allocation makes the short-term scale dominate in the fusion process to prioritize capturing the transient and weak changes of early faults; the medium-term scale is second to be used to smooth transient interference; and the long-term scale is the smallest to correct the reference offset caused by gradual environmental changes.

[0034] A dynamic fine-tuning mechanism is also introduced, with the following fine-tuning rules:

[0035] When the outdoor temperature (collected by temperature and humidity sensors, accuracy ±1℃) exceeds the range of 20℃~30℃, or the cooling / heating load rate (real-time call from the controller) exceeds the range of 30%~70%, or the dynamic value of the pipeline resistance exceeds the range of 700Pa~1300Pa, and the duration is ≥5s, the long-term scale weight will be adjusted. Adjusted to 0.3, short-time scale weight Adjusted to 0.35, with a medium-time scale weight. Remain unchanged;

[0036] When the above extreme operating conditions occur simultaneously Adjusted to 0.35. Adjusted to 0.3. It remains at 0.35;

[0037] When the environmental conditions return to the above-mentioned normal range and remain for ≥10 seconds, the weights are restored to their initial values ​​to ensure that the weights are accurately adapted to the real-time conditions.

[0038] After the weights are dynamically fine-tuned, the upper and lower limits of the benchmarks at short-term, medium-term, and long-term scales in the multi-scale operating benchmark set are fused and calculated respectively. The fusion formula is as follows (taking the k-th operating state parameter as an example):

[0039] ;

[0040] ;

[0041] in, , , These are the baseline upper limits of the kth operational status parameter at short-time, medium-time, and long-time scales, respectively, with units consistent with the corresponding operational status parameter. , , These are the baseline lower limits of the kth operating status parameter at short-time, medium-time, and long-time scales, respectively, with units consistent with the corresponding operating status parameter.

[0042] This yields the dynamic normal value range for each parameter (in the running status data). = That is, to form the adaptive operating range set .

[0043] The multi-timescale dynamic benchmark model includes an input matching layer, a multi-scale prediction layer, and a confidence output layer. The specific steps for generating the multi-scale operating benchmark set for the compressor are as follows:

[0044] In the input matching layer, the operating condition data is compared with the historical operating condition dataset stored in the database (including...). The historical operating condition data of the compressor is matched and processed to extract the reference operating sequence data, which is as follows:

[0045] The current operating condition data and historical operating condition data are normalized using a min-max method to obtain the current standardized operating condition feature vector. = (Corresponding to parameters in the operational status data and environmental condition data, 7 operational parameters and 3 environmental parameters) and historical operational condition dataset. = ;

[0046] Accurate matching of operating conditions: Utilizing a cosine similarity algorithm to calculate... and Each set of historical data similarity Its value ranges from [0, 1]. The closer it is to 1, the higher the similarity of the working conditions, and the better the corresponding historical operating status can reflect the current normal operating pattern.

[0047] Set similarity threshold (e.g., 0.85), filter ≥ Historical data groups; if the number of groups is less than 5, the threshold is lowered to 0.8 for re-filtering; if there are still fewer than 5 groups, the 5 groups with the highest similarity are selected as matching samples. The unnormalized historical operating condition data and operating condition data corresponding to the matching samples are arranged in chronological order of collection time and aligned by time (based on the current time) to obtain reference operating sequence data. = ( Let be the value of the k-th parameter (in the running status data) at time t, and (corresponding to 7 operating status parameters).

[0048] The steps for setting the similarity threshold are as follows:

[0049] Historical operating condition data stored in the smart city AI-optimized low-carbon HVAC and intelligent operation and maintenance platform are collected, covering typical operating conditions under different seasons, time periods and load conditions throughout the year, forming a historical operating condition sample library. Each set of data in this sample library contains complete operating status parameters (three-phase current, input power, intake and exhaust pressure, three-dimensional vibration) and corresponding environmental operating condition parameters (outdoor temperature and humidity, system load rate, and pipeline resistance).

[0050] The data in the historical working condition sample library is preprocessed, and the min-max normalization method is used to map all working condition parameters to the [0, 1] interval to eliminate the influence of different dimensions and obtain a standardized historical working condition feature vector set.

[0051] Several typical working conditions are randomly selected from the historical working condition sample database as test samples. For each test sample, the cosine similarity between it and all other working conditions in the sample database is calculated, and the mean of the similarity values ​​is used as the similarity threshold.

[0052] In the multi-scale prediction layer, multi-scale sliding slicing and time-series mining are performed on the reference runtime data to generate a compressor runtime prediction set, specifically as follows:

[0053] Reference runtime sequence data Sliding window slicing was performed at short, medium, and long time scales, with slicing parameters adapted to the acquisition frequency and length of the reference runtime data. For example, the short-term window size could be 10 acquisition times (corresponding to 100ms), with a sliding step of 1 acquisition time; the medium-term window size could be 600 acquisition times (corresponding to 1min), with a sliding step of 60 acquisition times; and the long-term window size could be 36,000 acquisition times (corresponding to 1h), with a sliding step of 3,600 acquisition times. Short-term slice sets were generated for each scale. Medium-scale slice set Long-term scale slice set ;

[0054] Each scale slice set is input into the corresponding scale temporal prediction model. The temporal prediction model adopts an LSTM and GRU cascade structure. The network parameters are customized according to the feature differences of different scale slice sets. For example, the input layer dimension is consistent with the window size of the corresponding scale, one LSTM layer is set to extract temporal features and initially learn long-term dependencies, one GRU layer is set to strengthen dependency learning and avoid gradient vanishing, a fully connected layer is used for feature mapping, and the output layer activation function is linear.

[0055] The future state of slice sets at each scale is predicted by a time series prediction model, and the output of the predicted time series of running state parameters is consistent with the slice length at each scale. The prediction time series length is 10 acquisition time for short time scale, 600 acquisition time for medium time scale, and 36,000 acquisition time for long time scale.

[0056] The prediction time series at the three scales mentioned above are time-aligned with the current acquisition time as the reference, and invalid values ​​in the prediction time series are filtered out (if a prediction value exceeds the physical value range of the corresponding operating state parameter, it is filled by linear interpolation of adjacent valid prediction values). The integrated data form the compressor operation prediction set. = ,in For short-term scale prediction sets, For mesoscale prediction sets, It is a long-term prediction set, and each prediction set corresponds to a sequence of predicted values ​​for 7 operating state parameters;

[0057] In the confidence output layer, the operational prediction set is processed to construct multi-scale confidence intervals, generating a multi-scale operational benchmark set for the compressor, specifically as follows:

[0058] For running prediction set = The predicted value sequences at various scales were statistically analyzed, and the predicted mean value of each operational state parameter at each acquisition time was calculated. With the predicted standard deviation d=1, 2, 3 correspond to short-term, medium-term, and long-term scales, respectively;

[0059] Combined with the normal distribution quantiles at the 95% confidence level =1.96, calculate the upper and lower confidence limits for each scale, parameter, and time point, and construct a multi-scale confidence interval. The calculation formula is as follows:

[0060] ;

[0061] ;

[0062] in, , These are the upper and lower confidence limits for the d-th and k-th operating parameters at time t, respectively, with units consistent with the corresponding operating state parameters. If the calculated lower confidence limit is less than the physical minimum value of the parameter (e.g., current less than 0A), then the physical minimum value of the parameter is taken as the lower confidence limit to ensure the rationality of the confidence interval.

[0063] The confidence interval data at each scale are integrated to generate a multi-scale operating benchmark set for the compressor. = ,in, For short-time scale reference sets, For mesoscale reference sets, For long-term scale reference sets, each reference set contains the upper and lower confidence limits of 7 operating status parameters at each acquisition time at the corresponding scale.

[0064] The pre-training steps for the multi-timescale dynamic benchmark model are as follows:

[0065] Construct a training dataset by extracting at least one year of historical normal operation data from the smart city AI-optimized low-carbon HVAC and intelligent operation and maintenance platform. Ensure that the data covers typical operating conditions under different seasons, time periods, and load conditions throughout the year. Arrange the extracted data in chronological order to form a continuous historical operation sequence dataset. Each data point contains complete operating status parameters (three-phase current, input power, intake pressure, exhaust pressure, three-dimensional vibration) and corresponding environmental operating condition parameters (outdoor temperature and humidity, system load rate, and pipeline resistance).

[0066] Historical runtime sequence data is sliced ​​using a sliding window to construct training sample pairs. For the short-term prediction model, the input window length is set to 10 acquisition times, the output window length is also set to 10 acquisition times, and the sliding step is set to 1 acquisition time. This generates a short-term training sample set. Each sample pair contains the operating condition data and running status data within the input window, as well as the running status data within the corresponding output window, as the prediction target. For the medium-term prediction model, the input window length is set to 600 acquisition times, the output window length is set to 600 acquisition times, and the sliding step is set to 60 acquisition times. This generates a medium-term training sample set. For the long-term prediction model, the input window length is set to 36,000 acquisition times (corresponding to 1 hour), the output window length is set to 36,000 acquisition times, and the sliding step is set to 3,600 acquisition times. This generates a long-term training sample set.

[0067] Three time-series prediction models were trained separately. Each scale's prediction model employed a cascaded LSTM and GRU structure, with the specific network configuration consistent with the inference phase. During training, historical data within the input window was used as model input, and actual running data within the output window was used as supervision labels. Mean squared error was used as the loss function to measure the difference between the model's predicted values ​​and the true values. Network parameters were optimized using the backpropagation algorithm, with an initial learning rate of 0.001. Iterative training was performed using the Adam optimizer, with each scale's model trained independently for 100 epochs. An early stopping mechanism was employed to prevent overfitting. Convergence occurred when the validation set loss decreased by less than 1e-5 for 10 consecutive epochs.

[0068] Model validation and error statistics were performed. The trained models at three scales were tested using validation set data, and the error between the predicted and actual values ​​for each model at each time step was collected. For the short-term model, the mean and standard deviation of the prediction error for each running parameter were calculated across all validation samples; similarly, the mean and standard deviation of the prediction error for each parameter were calculated for the medium- and long-term models. These statistics will serve as the baseline parameters for subsequent confidence interval construction.

[0069] Save the pre-trained model parameters and error statistics at each scale to form a deployable multi-timescale dynamic benchmark model.

[0070] In this implementation plan, a multi-timescale dynamic benchmark model is used to complete operating condition matching, time series prediction, and confidence interval generation. Then, an adaptive operating interval is obtained by combining dynamically adjusted fusion weights. This allows for real-time correction of judgment criteria in response to changes in environment and load. In the data processing stage, historical reference data that fits the current scenario is first screened through standardization and similarity matching. Then, multi-scale time series prediction is used to obtain operating trends under different time dimensions. Combined with confidence analysis, a stable and reliable multi-scale benchmark is formed. During the fusion process, transient mutation capture, steady-state interference filtering, and environmental gradual change correction are taken into account. At the same time, the contribution ratio of each scale is automatically adjusted according to changes in operating conditions, so that the final operating interval always maintains a high degree of matching with the current operating conditions. This improves the fit and accuracy of fault judgment, making the entire fault identification process closer to actual operating needs.

[0071] Specifically, the steps to obtain the compressor's operating residual sequence are as follows: Dynamic deviation processing is performed on the operating state data and the adaptive operating interval set to obtain the compressor's initial operating residual sequence, which is as follows:

[0072] Obtain the real-time values ​​of each operating parameter in the current data acquisition status. (such as instantaneous values ​​of three-phase current, instantaneous values ​​of input power, etc.), and the dynamic normal value range of the corresponding parameters. = Calculate the deviation between the real-time value of each operating parameter and the center value of the dynamic normal value interval, i.e., the initial residual of each operating parameter. :

[0073] ;

[0074] By combining the initial residuals of all operating parameters, we obtain the initial residual sequence. ;

[0075] Based on environmental operating condition data, the initial operating residual sequence is subjected to operating condition disturbance correction processing to obtain the compressor's operating residual sequence, which is as follows:

[0076] Read environmental condition data , is represented as: , In order, the values ​​are: outdoor ambient temperature (typically ranging from -20℃ to 45℃), outdoor relative humidity (ranging from 0% to 100%), HVAC system cooling / heating load rate (ranging from 0% to 100%), and dynamic value of pipe network resistance (ranging from 0 to 500 kPa).

[0077] And environmental operating condition data Normalization is performed (using the min-max normalization method; the minimum and maximum values ​​of each parameter in historical data can be obtained by statistically analyzing historical normal operation data) to obtain normalized environmental operating condition data. , ;

[0078] Based on the above normalized environmental condition data To establish a mapping model between environmental operating conditions and residual disturbances, a multiple linear regression model can be used to analyze the residual disturbance component of each parameter caused by fluctuations in environmental operating conditions. Make an estimate:

[0079] ;

[0080] in, For the regression intercept term, , , , These are the regression coefficients corresponding to the temperature, humidity, load rate, and pipeline resistance of the k-th operating parameter, respectively.

[0081] The initial residual sequence is corrected for operating condition disturbances to obtain the corrected residual values. :

[0082] ;

[0083] By combining all the corrected residual values, the operating residual sequence of the compressor is obtained. ;

[0084] ;

[0085] And the above , , , , The steps to obtain it are as follows:

[0086] First, data from N sampling times during the historical normal operation period is collected. Each sampling time contains two sets of key information:

[0087] The first set consists of normalized environmental condition data. For each sampling time N, the original environmental condition parameters (outdoor temperature, relative humidity, system load rate, and pipeline resistance) are min-max normalized to obtain normalized temperature, humidity, load rate, and pipeline resistance values. To include the intercept term of the regression model, a constant term 1 is added before the normalized data to form a complete input vector.

[0088] The second set is the actual initial residual value of each operating parameter at this moment. This value is calculated from the operating status data at the current moment and the set center value of the adaptive operating interval, reflecting the actual deviation of the equipment operation under this environmental condition.

[0089] Arrange the input vectors of all N sampling times in rows to construct a design matrix. This matrix has N rows and 5 columns, with each row corresponding to an input vector at one sampling time.

[0090] Meanwhile, for each operating parameter K, the actual initial residual values ​​of the parameter at all N sampling times are arranged in the corresponding order to construct a target vector. The length of the vector is N, and each element is the actual initial residual value at the corresponding time.

[0091] Then, the least squares method is used to solve for the regression coefficients. The mathematical principle of this method is to find a set of regression coefficients that minimizes the sum of squared errors between the predicted residual disturbance components and the actually observed initial residual values. The regression coefficients are then solved through matrix operations to obtain the aforementioned... , , , , .

[0092] In this implementation scheme, the initial residual is obtained through dynamic deviation calculation, and then interference correction is completed by combining environmental operating condition data, which effectively eliminates the impact of environmental fluctuations. By establishing a mapping relationship between environmental parameters and residual interference, the impact components of various operating condition changes on operating parameters can be accurately estimated. Then, this part of the interference component is removed from the initial residual, so that the final residual sequence only retains the information brought about by the changes in the equipment's own state. The processed residual sequence is cleaner and more stable, which can provide a reliable data foundation for subsequent feature extraction and anomaly identification, improve the accuracy of subsequent fault judgment, and clearly distinguish between environmental influences and equipment anomalies.

[0093] Specifically, the steps to obtain the compressor's operating residual feature points are as follows: Obtain the historical operating residual time series sequence of the compressor, and combine it with the operating residual sequence to construct the compressor's operating residual time series data. Specifically, obtain the operating residual sequences of the most recent L acquisition times from the historical database. The residual sequence at each time point is an n-dimensional vector (n=7 in this embodiment), corresponding to the residual values ​​of each operating parameter. Let the current time be t, then obtain the historical operating residual sequences at times t−1, t−2, ..., t−L. , ,..., ;

[0094] The running residual sequence at the current moment By concatenating the above historical residual sequence in chronological order, a time series of running residual data with a time window length of L+1 is constructed;

[0095] Feature extraction processing is performed on the operating residual time series data to extract the residual feature set of the compressor, including the residual amplitude feature set, the residual change rate feature set, and the residual correlation feature set;

[0096] Spatial fusion processing is performed on the residual feature set to obtain the operating residual feature points of the compressor.

[0097] The specific steps for extracting the residual feature set of the compressor are as follows: Perform amplitude domain statistical processing on the operating residual time series data to extract the residual amplitude feature set of the compressor. Specifically, for each operating parameter K (K=1, 2, ..., 7), extract the corresponding residual time series sequence from the operating residual time series data. Calculate the following statistics:

[0098] residual mean : Reflects the average degree to which the parameter deviates from the center value within the time window;

[0099] Peak Residual : Reflects the maximum deviation of the parameter within the time window (absolute value), used to capture instantaneous large fluctuations;

[0100] Residual fluctuation amplitude Standard deviation is used to measure the dispersion of residuals within a window, reflecting the stability of the parameters;

[0101] By combining the three statistics of each parameter, the residual amplitude feature set is obtained. :

[0102] ;

[0103] Trend mining is performed on the operating residual time series data to extract the residual change rate feature set of the compressor. Specifically, based on the residual time series of the same operating parameter, three features that can reflect the trend of the parameter change over time are calculated, including first-order difference, second-order difference and linear fitting slope.

[0104] The first-order difference is calculated, reflecting the change in residuals between adjacent time points, i.e., the instantaneous rate of change of the residuals. For each operating parameter k, there are L+1 residual values ​​within the time window. The difference between the residuals at adjacent time points is calculated to obtain the first-order difference sequence. Positive values ​​of the first-order difference indicate an increase in residuals, while negative values ​​indicate a decrease. The larger the absolute value, the more drastic the change. Since the residual change at the current time t best reflects the latest dynamic trend, the first-order difference at the current time is used in this embodiment. As an eigenvalue;

[0105] Calculate the second difference, which is the rate of change of the first difference, i.e., the rate of change of the residual. It is used to identify trend inflection points or acceleration / deceleration characteristics. The first-difference sequence is then differenced again. A positive second difference indicates an increasing rate of change (acceleration), while a negative value indicates a decreasing rate of change (deceleration). A larger absolute value indicates more drastic fluctuations in the rate of change. Similarly, the second difference at the current moment is taken. As an eigenvalue;

[0106] To calculate the slope of the linear fit, for the residual time series of the same operating parameter, a linear fit is performed. The least squares method is used to estimate the overall trend of the residual over time, reflecting the long-term monotonic change direction. This involves assigning a time index to each moment within the time window in chronological order. For each operating parameter, the residual value of that parameter at each moment is used as the dependent variable, and the corresponding time index is used as the independent variable. A univariate linear regression model is established, and the least squares fit is performed on this model. The slope of the best-fit line is the desired linear fit slope. The sign of the slope indicates the overall direction of change of the residuals; a positive value indicates an increasing trend, and a negative value indicates a decreasing trend. The magnitude of the absolute value of the slope indicates the strength of the trend; the larger the absolute value, the more obvious the trend.

[0107] The first difference of each of the above operating parameters Second-order difference linear fitting slope Combining these features yields the residual rate of change feature set. ;

[0108] ;

[0109] The residual time series data is subjected to correlation domain coupling processing to extract the residual correlation feature set of the compressor, specifically as follows:

[0110] For any two running parameters and , Read the corresponding residual time series and analyze the corresponding correlation coefficient based on the Pearson correlation coefficient method. ;

[0111] The residual values ​​of each operating parameter are sorted by numerical value and then evenly divided into V intervals (e.g., V=10). Each interval contains the same number of residual values. After discretization, each residual value is mapped to a discrete symbol. ∈{1, 2, ..., V};

[0112] Let the discretized symbol sequence of the running parameter i be... The symbol sequence of the running parameter j is , statistically calculate the joint probability distribution , that is, the frequency of the symbol pair ; at the same time, statistically calculate the marginal probability distributions and , that is, the frequency of each symbol, then the mutual information has the following calculation formula:

[0113] ;

[0114] The mutual information is non - negative, with the unit of nat (if taking the logarithm with base 2, the unit is bit), = 0 indicates that the two parameters are independent of each other; The larger it is, the stronger the dependence relationship between the two parameters;

[0115] The residual co - deviation value is used to measure the degree to which two operating parameters deviate from the normal level simultaneously at the current moment. The definition of this index is based on the following observation: Many fault modes will cause multiple parameters to deviate abnormally at the same time. For example, refrigerant leakage will cause coordinated changes in multiple parameters such as reduced suction pressure, reduced discharge pressure, and reduced current. Therefore, capturing this co - deviation helps in early fault identification. Its calculation formula is:

[0116] ;

[0117] Among them, and are the residual values of the operating parameters i and j at the current moment respectively, and are the standard deviations of the residual time - series sequences of the operating parameters i and j respectively;

[0118] For all parameter pairs that satisfy i < j, combine the above - mentioned correlation coefficient , mutual information , residual co - deviation value to obtain the residual correlation feature set of the compressor:

[0119] .

[0120] The specific steps of spatial fusion processing are as follows: Perform condition mapping processing on the environmental condition data to generate the residual weight set of the compressor. Specifically: Normalize the environmental condition data at the current acquisition moment to obtain the normalized environmental condition data , ;

[0121] A mapping model from operating conditions to weights is constructed. This embodiment adopts a data-driven method based on historical fault sensitivity analysis, enabling weight allocation to adapt to the sensitivity of each feature to faults under different operating conditions. The specific steps are as follows:

[0122] A historical sensitivity sample library is constructed, collecting historical fault sample data. Each sample contains a sequence of environmental operating condition data at Y time points before the fault occurred, as well as fault sensitivity labels for each operating parameter across various feature dimensions. The calculation method is as follows: statistically analyze the change of parameter k in feature dimension b over a period of time before the failure occurs (e.g., the change of the mean residual for the amplitude feature, the mean of the first difference for the rate of change feature, and the change of the correlation coefficient for the correlation feature), divide it by the standard deviation of the fluctuation of the feature when the parameter is operating normally, and obtain the dimensionless sensitivity index. The operating condition data of all samples (taking the operating condition at the last moment before the failure as representative) and the corresponding sensitivity labels are combined to form a training dataset.

[0123] A Multilayer Perceptron (MLP) is used as the weight prediction model. Its network structure is as follows: the input layer contains 4 nodes (corresponding to the 4 normalized environmental condition parameters); the hidden layer has two layers, with 64 nodes in the first layer and 32 nodes in the second layer, using ReLU activation; the output layer contains W×3=27 nodes (corresponding to the weights of the 7 operating parameters on 3 feature dimensions); and a Softmax normalization layer follows the output layer to ensure that the sum of the three weights for each parameter is 1. The objective function of network training is to minimize the mean squared error between the predicted weights and the sensitivity labels. Backpropagation is used to optimize the network parameters, enabling the model to learn the optimal fusion weights for each feature dimension under different operating conditions. The MLP weight prediction model can be trained for 50 epochs, and an early stopping mechanism is employed to prevent overfitting.

[0124] The current normalized operating condition vector environment operating condition data The trained weight prediction network is input, and the output vector is calculated through forward propagation. This vector is then reshaped into a k×3 weight matrix. :

[0125] ;

[0126] Each row corresponds to an operating parameter, and each column corresponds to a feature dimension (b=1 for amplitude feature, b=2 for rate of change feature, and b=3 for correlation feature). This weight matrix is ​​the residual weight set of the compressor.

[0127] The residual weight set and residual feature set are fused and mapped to generate the compressor's operating residual feature points, specifically:

[0128] Read the mean residual value of each operating parameter in the residual magnitude feature set. residual peak residual fluctuation amplitude And perform normalization processing (such as z-score normalization);

[0129] And the mean residual of each operating parameter after normalization. residual peak residual fluctuation amplitude Weighted processing is performed (and the weighting coefficients in the weighting process are preset according to the contribution of each indicator to the fault; in this embodiment, the mean of the residuals is used). residual peak residual fluctuation amplitude The corresponding weighting coefficients are 0.3, 0.4, and 0.3 (to emphasize the important role of peak values ​​in capturing transient anomalies), resulting in the comprehensive amplitude characteristics of each operating parameter. ;

[0130] Read the first difference of each running parameter in the residual rate of change feature set Second-order difference linear fitting slope The data is then normalized and then weighted (the weighting coefficients are preset based on the contribution of each indicator to the fault; in this embodiment, the first-order difference is used). Second-order difference linear fitting slope The corresponding weighting coefficients are 0.4, 0.3, and 0.3 (to highlight the ability of the first-order difference at the current moment to represent the latest trend of change), thus obtaining the comprehensive change characteristics of each operating parameter. ;

[0131] Read the correlation coefficient of each parameter pair in the residual correlation feature set. Mutual information residual collaborative deviation value For the correlation coefficient Perform transformation processing (i.e., 1- ), to obtain the relevant deviation. The correlation deviation for each parameter pair Mutual information residual collaborative deviation value Normalization is performed, followed by weighting (the weighting coefficients are preset based on the contribution of each indicator to the fault; in this embodiment, the first-order difference is used). Second-order difference linear fitting slope The corresponding weighting coefficients are 0.3, 0.3, and 0.4 (to emphasize the ability of the co-deviation index to represent simultaneous anomalies at the current time), thus obtaining the comprehensive correlation features of each parameter pair. ;

[0132] Comprehensive correlation features for each parameter pair Aggregation processing is performed to obtain the comprehensive residual correlation features corresponding to each operating parameter. (It reflects the overall coupling relationship between the operating parameter k and all other operating parameters; the larger the value, the more obvious the abnormal coupling relationship between this parameter and other parameters.)

[0133] ;in, This represents the total number of running parameters.

[0134] Using residual weight set The weights in the model are then used for cross-parameter weighted fusion to obtain the components of the three coordinate axes in the residual feature space.

[0135] Amplitude axis coordinates The calculation formula is:

[0136] ;

[0137] This coordinate component integrates the magnitude characteristics of all parameters, with weights. The larger the value, the greater the contribution of the magnitude characteristic of parameter k to fault identification under the current operating conditions;

[0138] Rate of change axis coordinates The calculation formula is:

[0139] ;

[0140] This coordinate component integrates the rate of change characteristics of all parameters, reflecting the overall dynamic trend of change;

[0141] Related axis coordinates The calculation formula is:

[0142] ;

[0143] This coordinate component integrates the coupling relationship characteristics of all parameters, reflecting the overall coordination state of the system;

[0144] After standardization, the coordinate components are mapped to a preset three-dimensional feature space to obtain the operating residual feature points representing the current operating state of the compressor. .

[0145] The preset steps for the three-dimensional feature space are as follows:

[0146] Define three coordinate axes in the three-dimensional feature space. This space is a three-dimensional Cartesian coordinate system. The first coordinate axis is the amplitude axis, used to characterize the degree to which each operating parameter deviates from the normal level; the second coordinate axis is the rate of change axis, used to characterize the drastic degree and trend direction of each operating parameter's change; the third coordinate axis is the correlation axis, used to characterize the coupling and coordination state between each operating parameter.

[0147] Historical normal operation data is collected as a benchmark for spatial establishment; at least 30 days of normal operation data are extracted from the compressor historical operation database, and each collection moment is processed according to the above method to obtain three comprehensive feature values ​​corresponding to each moment: amplitude comprehensive feature, rate of change comprehensive feature, and correlation comprehensive feature;

[0148] Calculate the historical statistical characteristics of each parameter in each dimension. For each operating parameter k, calculate the mean and standard deviation of its comprehensive amplitude characteristics, the mean and standard deviation of its comprehensive rate of change characteristics, and the mean and standard deviation of its comprehensive correlation characteristics during the historical normal operation period, so as to form the benchmark parameters for subsequent standardization.

[0149] Define the measurement standards for each coordinate axis, and standardize the comprehensive amplitude characteristics calculated at any time (using the above-mentioned benchmark parameters). The same applies to the comprehensive characteristics of the rate of change and the comprehensive correlation characteristics.

[0150] The standardized parameters and features of each dimension are weighted and fused with the generated residual weight set to obtain the specific values ​​of three coordinate components: amplitude axis coordinate, rate of change axis coordinate, and correlation axis coordinate. These three coordinate components together constitute the running residual feature points in the preset three-dimensional feature space at the current moment. This feature space is based on the statistical distribution of historical normal operating states. 0 points represent the average level of the normal state, positive and negative values ​​represent the direction of deviation, and the magnitude of the absolute value represents the degree of deviation.

[0151] In this implementation plan, time-series data is constructed by combining historical residual sequences. Features are extracted from three perspectives: amplitude, trend of change, and coupling relationship between parameters. Abnormal information during equipment operation is fully preserved. Then, feature weights are automatically assigned according to environmental conditions, allowing features that are more sensitive to faults to play a greater role and improving the targeting of anomaly identification. Through weighted fusion, multiple types of features are transformed into coordinate points in a three-dimensional feature space, simplifying complex multi-parameter changes into intuitive spatial locations, which facilitates subsequent trajectory monitoring and mutation identification. This allows for the full mining of hidden information in the data, while reducing the impact of interference factors, and thus enabling clearer identification of various abnormal signs that appear in the early stages of equipment operation.

[0152] Specifically, such as Figure 3As shown, the specific steps to obtain the compressor's abnormal deviation evaluation value are as follows: Continuously acquire the compressor's operating residual feature points at multiple acquisition times, and connect them in chronological order to construct the compressor's operating evolution trajectory. Specifically, following the above steps, continuously acquire the compressor's operating status data within a preset time period, generating an operating residual feature point corresponding to each acquisition time. , To determine the number of data acquisition timestamps, the residual feature points from all consecutive acquisition timestamps are connected sequentially according to their acquisition timestamps to form the compressor's operational evolution trajectory. ;

[0153] Trend offset processing is performed on the operational evolution trajectory to extract the compressor's operational offset evaluation set. Specifically, this involves calculating the displacement vector between operational residual feature points at adjacent acquisition times on the operational evolution trajectory.

[0154] ;

[0155] in, For from the first The feature point points to the first The displacement vector of each feature point reflects the direction and magnitude of the change in the compressor's operating state in the residual feature space between adjacent time points;

[0156] The sliding window method is used to analyze the directional consistency of displacement vectors. The length of the sliding window is set (5 in this embodiment) and the step size is 1. The displacement vector sequence is scanned. For each sliding window, the average direction vector of all displacement vectors in the window is calculated, and the angle between each displacement vector in the window and the average direction vector is calculated (the vector angle formula can be used for calculation).

[0157] If in multiple consecutive windows (3 in this embodiment), all included angles in each window are less than a preset angle threshold (e.g., 15), then it is determined that the compressor operating state has a continuous unidirectional deviation.

[0158] The steps for setting the angle threshold are as follows:

[0159] Collect historical normal operation data stored in the smart city AI-optimized low-carbon HVAC and intelligent operation and maintenance platform, extract at least 30 days of continuous operation data, generate operation residual feature points for each moment according to the above method, and construct the operation evolution trajectory;

[0160] A sliding window analysis is performed on the evolution trajectory under normal operating conditions. The window length is set to be consistent with that used in real-time monitoring (5 displacement vectors in this embodiment), and the step size is 1. The entire normal operating trajectory is scanned. For each window, the average direction vector of each displacement vector within the window is calculated, and the angle between each displacement vector and the average direction vector is calculated.

[0161] The distribution characteristics of the included angles within all windows under normal operating conditions were statistically analyzed. Under normal operating conditions, due to random fluctuations, the directions of each displacement vector are usually quite dispersed, and the included angles are generally large. The 90th percentile of the included angles within all windows was calculated, meaning that 90% of the included angles under normal operating conditions are greater than this value. Considering the need to ensure the detection sensitivity for continuous unidirectional displacement, and to avoid misjudging normal random fluctuations as trend displacements, the angle threshold was set to a value slightly lower than this quantile, such as 15 degrees.

[0162] The starting index of the recorded offset is the starting index of the first window that meets the above conditions. This indicates that the device has a progressive failure trend, so the offset magnitude is calculated. (Using the offset starting point as the starting point and the latest running residual feature point as the ending point, calculate the Euclidean distance between the two points in the three-dimensional residual feature space to obtain...) ) and offset duration (Count the number of acquisition times elapsed from the initial offset time to the latest time, multiply by the time interval between adjacent acquisition times, and obtain...) ), constitute the runtime offset evaluation set ;

[0163] Spatial distribution abrupt changes are identified in the operational evolution trajectory, and a density abrupt change assessment set for the compressor is extracted. Specifically, a baseline density distribution is established using feature point data from historical normal operation periods. At least 30 days of normal operation data are extracted from the historical database, and feature points are collected evenly at several times each day to form a baseline sample set. The probability density function in the feature space is calculated using the kernel density estimation method. In this embodiment, a Gaussian kernel function is selected.

[0164] ;

[0165] Where g is any point in the feature space, and hx is the bandwidth parameter, which is automatically determined by the Silverman rule: , The density function is the mean of the standard deviations of each dimension of the sample. It reflects the spatial distribution of feature points under normal operating conditions. The density is high in the normal operating area and low in the sparse area.

[0166] For the benchmark sample set For each feature point in the dataset, calculate its density value under the baseline density distribution to obtain a set of density values. Calculate the 5th percentile of this set. This serves as the lower confidence limit for density (i.e., the lower baseline limit), meaning that under normal operating conditions, only 5% of feature point densities are below this value.

[0167] On the trajectory of operation Real-time monitoring is performed for the latest feature points. Calculate its current density under the baseline density distribution. Simultaneously, the average density of feature points over a number of recent time periods (10 in this embodiment) is calculated. As a local reference;

[0168] If both of the following conditions are met simultaneously, a sudden change in spatial distribution density is determined to have occurred:

[0169] The current density is below the lower limit of the baseline: < ;

[0170] The decrease in current density relative to local average density exceeds a preset threshold. (This embodiment takes) =0.5): ;

[0171] in, The default steps are as follows:

[0172] Collect historical normal operation data, extract at least 30 days of continuous operation data, and generate operation residual feature points for each moment according to the above method to form a normal operation sample set;

[0173] A sliding window analysis was performed on the normal operating sample set. The window length was 10 time points and the step size was 1. For each window, the average density of feature points within the window and the baseline density of the last feature point in the window were calculated, and the relative decrease was calculated. The value was recorded only when the baseline density was lower than the baseline lower limit.

[0174] Then, the 95th percentile of all relative decreases is calculated. This value represents the density decrease that occurs in 95% of cases under normal operating conditions. Set a value slightly below the 95th percentile, such as =0.5;

[0175] When a density mutation is detected, the mutation start time is recorded as the moment when the density first falls below the baseline lower limit. The following two indicators are calculated to form a density mutation assessment set. :

[0176] Density decrease The difference between the lower limit of the baseline and the current density. If the current density is higher than the lower limit of the baseline, it is set to 0. This value reflects the degree to which the current state deviates from the normal density region.

[0177] Mutation duration ψ represents the length of time elapsed from the start of the mutation to the current time, reflecting the duration of the abnormal state;

[0178] The operating deviation assessment set and the density mutation assessment set are fused and assessed to generate the compressor's operating abnormal deviation assessment value.

[0179] The specific steps of the fusion evaluation process are as follows: Normalization is performed on the runtime offset evaluation set and the density mutation evaluation set. Specifically, this involves: normalizing the runtime offset evaluation set... Obtain the offset magnitude and offset duration From the density mutation assessment set Obtain the density decrease rate and mutation duration ;

[0180] The min-max normalization method is used to map each indicator to the interval [0, 1]. The minimum and maximum values ​​of each indicator in historical normal operation data and historical fault data are pre-calculated, and min-max normalization is performed on them to obtain the normalized offset. , duration of offset Density decrease Duration of mutation ;

[0181] Furthermore, historical fault sample data is acquired to determine the fusion optimization weight set for the compressor. Specifically, this involves collecting historical fault event data and, for each fault, extracting the normalized index vector from a period prior to the fault occurrence (e.g., 30 minutes before the fault). ,in The system indexes fault samples and assigns a fault severity label to each sample based on the severity of the consequences of the fault (such as equipment downtime, maintenance costs, energy efficiency loss, etc.). The value ranges from [0, 10], and the larger the value, the more severe the fault.

[0182] The training dataset is composed of the indicator vectors of all failure samples and their corresponding severity labels. ,in =1, 2, ..., , The total number of samples;

[0183] A fusion model is established using multiple linear regression, with the fusion evaluation value defined as a linear combination of the indicators:

[0184] ;

[0185] in , , , The desired fusion weights are determined by minimizing the predicted values. Compared to actual severity The optimal weights are determined by the mean square error between them, i.e., the following optimization problem is solved:

[0186] ;

[0187] This optimization problem can be solved directly using the least squares method to obtain the optimal weight vector. The weights are then normalized to sum to 1 to obtain the fused optimized weight set. ;

[0188] Based on the fusion optimization weight set and combined with the normalization results, a comprehensive analysis is performed to obtain the compressor's operational anomaly deviation assessment value. Specifically, the initial anomaly deviation assessment value is obtained by weighted summation of the four normalized indicators according to the fusion optimization weight set. (This value combines anomalies from both trend shifts and density abrupt changes, reflecting the overall degree of anomaly in the current equipment status):

[0189] ;

[0190] Initial abnormal deviation assessment value The mapping process is performed to obtain the runtime anomaly deviation evaluation value. : ;

[0191] in, The steepness coefficient is a preset value that controls the slope of the mapping curve; in this embodiment, it can be set to 8. The center point parameter represents the initial value when the evaluation value reaches 50. In this embodiment, it can be taken as 0.5.

[0192] Among them, steepness coefficient The default steps are as follows:

[0193] Collect historical fault sample data and corresponding normal operation sample data stored in the smart city AI-optimized low-carbon HVAC and intelligent operation and maintenance platform. For each sample, calculate its initial abnormal deviation evaluation value according to the above method to obtain the initial evaluation value set of normal operation samples and the initial evaluation value set of fault samples.

[0194] Analyze the distribution characteristics of the two sets, and statistically analyze the mean and standard deviation of the initial evaluation values ​​of the normal operation sample and the initial evaluation values ​​of the fault sample.

[0195] Define a discrimination index, using the ratio of the distance between the means of two classes of samples to the sum of their standard deviations as the discrimination index:

[0196] The optimal steepness coefficient is determined using a grid search method, and the following settings are made: The search range is [1, 15], with a step size of 0.1. For each candidate... Value, calculated as described above The formula maps the initial evaluation values ​​of the two types of samples and calculates the corresponding discrimination index, where We will temporarily use 0.5 as the center point parameter;

[0197] Traverse all candidates After setting the value, select the one that maximizes the discrimination index. As the optimal steepness coefficient;

[0198] In this implementation plan, feature points from multiple moments are combined in chronological order to form an operational evolution trajectory. Then, the continuous deviation of the trajectory and the sudden change in spatial density are analyzed separately. This allows for the simultaneous capture of both gradual and sudden changes in equipment status. During the evaluation process, various abnormal indicators are first standardized and then combined with historical fault data to determine reasonable fusion weights. This allows the two types of information, deviation and sudden change, to be reasonably combined according to the actual impact of the fault, in order to obtain a comprehensive abnormal deviation evaluation value. This value reflects the current level of equipment abnormality, enabling timely detection and appropriate handling of equipment abnormalities.

[0199] Specifically, the steps for graded identification and operation control processing of the compressor are as follows: The operational anomaly deviation assessment value is compared and analyzed with a preset set of operational anomaly deviation assessment intervals. The steps for setting the operational anomaly deviation assessment interval set are as follows:

[0200] Collect normal operation samples and fault samples from historical operation data, and calculate the operation abnormality deviation evaluation value for each sample according to the above method to obtain the set of normal operation evaluation values ​​and the set of fault sample evaluation values;

[0201] The distribution characteristics of the set of normal operation assessment values ​​are statistically analyzed, their mean and standard deviation are calculated, and the upper limit of the assessment values ​​for normal operation is determined. = Mean + Adjustment coefficient (3 in this example) × Standard deviation, which is the upper confidence limit for the normal state;

[0202] Analyze the distribution of the fault sample evaluation value set, calculate its lower quartile and median, and set the upper limit of the evaluation value. The median of the quartiles and the median is used as the first-level warning threshold Th1, the median of the quartiles and the median is used as the second-level warning threshold Th2, and the median is used as the third-level warning threshold Th3.

[0203] Based on the above thresholds, the abnormal deviation from the evaluation value is divided into four intervals:

[0204] Normal range: [0, Th1);

[0205] Level 1 warning range: [Th1, Th2);

[0206] Level II warning range: [Th2, Th3);

[0207] Level 3 warning range: [Th3, 100];

[0208] Based on the comparative analysis results, the compressor was identified, and corresponding control measures were taken. Specifically, when the compressor's operation deviates abnormally from the assessed value... When the compressor is in the normal range, it is determined that the compressor is in normal operation. The control measures can be as follows: maintain the existing operating mode, continue to maintain the regular monitoring frequency, leave the control command empty, do not make any adjustments to the compressor operating parameters, and the HVAC system status indicator light will be green, indicating that the operation is normal.

[0209] When the operation deviates abnormally from the evaluation value When the compressor is in the Level 1 warning zone, it is determined to be in a Level 1 warning state (monitoring required). The control measures can be as follows: generate a Level 1 warning signal and push it to the monitoring interface through the operation and maintenance platform to prompt operation and maintenance personnel to pay attention to the equipment status; limit the maximum operating frequency of the compressor to no more than 90% of the rated frequency to avoid high-load operation of the equipment under potential abnormal conditions; at the same time, start the operation data cache, mark the data of this period as data of concern and store it in the historical database for subsequent analysis; the HVAC system status indicator light will turn yellow, indicating that monitoring is required.

[0210] When the operation deviates abnormally from the evaluation value When the compressor is in the Level 2 warning zone, it is determined that the compressor is in a Level 2 warning state (inspection needs to be arranged). The control measures can be as follows: generate a Level 2 warning signal, push it to the monitoring interface through the operation and maintenance platform, and send an SMS notification to the on-duty engineer; further limit the maximum operating frequency of the compressor to no more than 70% of the rated frequency, for example, appropriately increase the evaporation temperature setpoint to reduce the compression ratio; the HVAC system status indicator light will turn orange, indicating that inspection needs to be arranged as soon as possible.

[0211] When the operation deviates abnormally from the evaluation value When the compressor is in a Level 3 warning zone, it is determined that the compressor is in a Level 3 warning state (immediate shutdown and maintenance are recommended). The control measures can be as follows: generate a Level 3 warning signal, push it to the monitoring interface through the operation and maintenance platform, and simultaneously send SMS and telephone notifications to the on-duty engineer and operation and maintenance supervisor; trigger active load reduction protection to smoothly reduce the compressor load rate to below 50% within 30 seconds; if the load rate is difficult to reduce to a safe range or the assessed value continues to rise, execute the emergency shutdown procedure; increase the chilled water outlet temperature setpoint by 2-3°C to reduce the overall system load demand; start the backup compressor (if any) to share the load and ensure the basic cooling / heating needs of the terminal; the HVAC system status indicator light will turn red and issue an audible and visual alarm to prompt immediate action.

[0212] This implementation plan quickly determines the current operating status of equipment by comparing the real-time calculated abnormal deviation assessment value with a preset range. Clear and definite judgment criteria exist for everything from normal operation to different levels of warnings. Corresponding measures are taken at different abnormality levels, avoiding unnecessary intervention for minor anomalies and promptly reducing equipment load when risks escalate to prevent escalation of faults. While ensuring equipment safety, it maintains continuous and stable system operation as much as possible, effectively reducing the impact of sudden shutdowns, lowering subsequent maintenance costs, and improving the overall reliability of the HVAC system.

[0213] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0214] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for early fault identification in HVAC systems based on abrupt changes in compressor operating characteristics, characterized in that, Includes the following steps: Acquire the compressor's operating condition data, which includes operating status data and environmental condition data; The operating condition data is input into a pre-constructed multi-timescale dynamic benchmark model to analyze the compressor's adaptive operating range set; Adaptive residual enhancement processing is performed on the operating status data and the adaptive operating interval set to obtain the compressor's operating residual sequence; The operating residual sequence is subjected to projection mapping processing to obtain the operating residual feature points of the compressor; The evolution trajectory monitoring and mutation identification processing of the operation residual feature points are performed to obtain the compressor operation abnormal deviation evaluation value; Based on the aforementioned operational anomaly deviation assessment value, the compressor is subjected to graded identification and operational control processing.

2. The method for early fault identification of HVAC systems based on sudden changes in compressor operating characteristics according to claim 1, characterized in that, The specific steps for analyzing the adaptive operating range set of the compressor are as follows: Based on the multi-timescale dynamic benchmark model, the operating condition data is processed by multi-scale feature mapping and benchmark prediction to generate a multi-scale operating benchmark set for the compressor. The multi-scale operating benchmark set is fused and optimized to obtain the adaptive operating interval set of the compressor.

3. The method for early fault identification of HVAC systems based on sudden changes in compressor operating characteristics according to claim 2, characterized in that, The multi-timescale dynamic benchmark model includes an input matching layer, a multi-scale prediction layer, and a confidence output layer. The specific steps for generating the multi-scale operating benchmark set for the compressor are as follows: In the input matching layer, the operating condition data is matched with the historical operating condition dataset stored in the database to extract the compressor's reference operating sequence data; In the multi-scale prediction layer, the reference runtime sequence data is subjected to multi-scale sliding slicing and time-series mining processing to generate a compressor runtime prediction set. In the confidence output layer, the operation prediction set is processed to construct multi-scale confidence intervals to generate a multi-scale operation benchmark set for the compressor.

4. The method for early fault identification of HVAC systems based on sudden changes in compressor operating characteristics according to claim 1, characterized in that, The specific steps to obtain the compressor's operating residual sequence are as follows: The operating status data and the adaptive operating interval set are dynamically deviated to obtain the compressor's initial operating residual sequence; Based on the environmental operating condition data, the operating condition interference correction process is performed on the initial operating residual sequence to obtain the operating residual sequence of the compressor.

5. The method for early fault identification of HVAC systems based on sudden changes in compressor operating characteristics according to claim 1, characterized in that, The specific steps to obtain the operating residual characteristic points of the compressor are as follows: Obtain the historical operating residual time series sequence of the compressor, and combine the operating residual sequence to construct the operating residual time series data of the compressor; The operating residual time series data is subjected to feature extraction processing to extract the residual feature set of the compressor, including the residual amplitude feature set, the residual change rate feature set, and the residual correlation feature set; Spatial fusion processing is performed on the residual feature set to obtain the operating residual feature points of the compressor.

6. The method for early fault identification of HVAC systems based on sudden changes in compressor operating characteristics according to claim 5, characterized in that, The specific steps for extracting the residual feature set of the compressor are as follows: The residual time series data of the operation is subjected to amplitude domain statistical processing to extract the residual amplitude feature set of the compressor; The operating residual time series data is subjected to trend mining processing to extract the residual change rate feature set of the compressor; The residual time series data is subjected to correlation domain coupling processing to extract the residual correlation feature set of the compressor.

7. The method for early fault identification of HVAC systems based on sudden changes in compressor operating characteristics according to claim 5, characterized in that, The specific steps of spatial fusion processing are as follows: The environmental operating condition data is subjected to operating condition mapping processing to generate a residual weight set for the compressor; The residual weight set and the residual feature set are fused and mapped to generate the compressor's operating residual feature points.

8. The method for early fault identification of HVAC systems based on sudden changes in compressor operating characteristics according to claim 1, characterized in that, The specific steps to obtain the compressor's abnormal operation deviation assessment value are as follows: By continuously acquiring the operating residual feature points of the compressor at multiple acquisition times, the operating evolution trajectory of the compressor is constructed. The operational evolution trajectory is subjected to trend offset processing to extract the compressor's operational offset evaluation set; Spatial distribution mutation identification processing is performed on the aforementioned operational evolution trajectory to extract the compressor density mutation evaluation set; The operating offset evaluation set and the density mutation evaluation set are fused and evaluated to generate the compressor's operating abnormality deviation evaluation value.

9. The method for early fault identification of HVAC systems based on sudden changes in compressor operating characteristics according to claim 8, characterized in that, The specific steps of the fusion assessment process are as follows: The running offset evaluation set and the density mutation evaluation set are normalized. And acquire historical fault sample data to determine the fusion optimization weight set for the compressor; Based on the fusion optimization weight set and combined with the normalization processing results, a comprehensive analysis is performed to obtain the compressor's abnormal operation deviation evaluation value.

10. The method for early fault identification of HVAC systems based on sudden changes in compressor operating characteristics according to claim 1, characterized in that, The specific steps for classifying and identifying the operation and control of the compressor are as follows: The operational anomaly deviation assessment value is compared and analyzed with a preset operational anomaly deviation assessment interval set; Based on the comparison and analysis results, the compressor is identified, and corresponding control measures are taken.

Images