Air conditioner fault early warning system and method based on deep learning

Abstract

The invention discloses an air conditioner fault early warning system and method based on deep learning, and the method comprises the following steps: S1, collecting data, forming current multi-channel time sequence input data, and obtaining a historical fault-free operation data sample; s2, constructing a normal state prototype sample; s3, generating prototype residual error coding features; s4, inputting the current multi-channel time sequence input data and the prototype residual encoding features into the improved multi-scale convolution branch network to obtain original feature branches and residual feature branches, and introducing a causal-guided convolution kernel selection mechanism; s5, fusing the original feature branch and the residual feature branch; s6, obtaining an abnormal score value based on fusion feature representation; s7, the abnormal score value is compared with a preset threshold value, and an air conditioner fault early warning instruction is output; and S8, when the abnormal score value does not reach the preset threshold value, updating the prototype sample in the normal state. The air conditioner fault early warning method improves the accuracy and self-adaptive capacity of air conditioner fault early warning, and is suitable for intelligent operation and maintenance scenes of an air conditioner system.

Description

Technical Field

[0001] This invention relates to the field of intelligent device fault monitoring and early warning technology, and in particular to an air conditioner fault early warning system and method based on deep learning. Background Technology

[0002] With the increasing demand for building intelligence and energy-saving management, real-time acquisition of multi-channel operating data and intelligent fault early warning technology for large air conditioning systems have received widespread attention. Existing air conditioning fault monitoring and maintenance methods mainly rely on single-parameter threshold settings, traditional statistical feature analysis, or anomaly detection based on simple neural network models. These methods generally suffer from the following problems in real-world complex environments:

[0003] The actual acquisition of multi-channel operational data is affected by factors such as equipment aging, environmental interference, and data loss. This often results in abnormal amplitudes, high noise levels, and inconsistent sampling step sizes in some sensor data, making it difficult to capture key fault symptoms such as temperature, pressure, and current anomalies in a timely and accurate manner. Overall, the data quality is low and reliability is insufficient. Existing operational data acquisition and processing mostly focus on single-channel or single-moment features, lacking continuous modeling of equipment state evolution across multiple channels and time series dimensions. Cross-channel and cross-time feature coupling relationships are difficult to reflect, affecting the sensitive detection of complex multi-source faults. For different equipment types, operating conditions, and seasonal variations, traditional methods typically use static feature templates and fixed thresholds, which are difficult to adapt to dynamic changes and new operating conditions, leading to high false alarm rates, high false negative risks, and difficulty in guaranteeing operational efficiency. Furthermore, current convolutional models mostly use preset fixed kernel structures, making it difficult to dynamically adjust the feature contributions of different branches. They lack the ability to adaptively model multi-scale features and causal relationships, making it difficult to accurately identify potential multi-type anomaly patterns, affecting the practicality and widespread adoption of intelligent early warning systems.

[0004] Therefore, how to provide an air conditioning fault early warning system and method based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based air conditioning fault early warning system and method. This invention fully utilizes deep learning techniques such as multi-channel time-series data processing, improved multi-scale convolutional branch networks, causal guided convolutional kernel selection, and prototype residual coding. It describes in detail the methods for efficient feature extraction, intelligent identification of abnormal states, and adaptive dynamic updating of health samples from air conditioning equipment operation data. It has the advantages of strong feature extraction capability, high fault early warning accuracy, and good adaptability to new operating conditions.

[0006] An air conditioning fault early warning method based on deep learning according to an embodiment of the present invention includes the following steps:

[0007] S1. Collect multi-channel operating data of the air conditioning equipment, construct the current input sample within a fixed time window, form the current multi-channel time-series input data, and use the known samples that have not experienced faults in the historical operation process as historical fault-free operating data samples.

[0008] S2. Select all or part of the historical fault-free operation data samples and define them as normal state prototype samples;

[0009] S3. Subtract the current multi-channel time-series input data from the normal state prototype sample element by element in both the channel and time dimensions to generate prototype residual coding features.

[0010] S4. Input the current multi-channel temporal input data and the prototype residual coding features into two structurally consistent improved multi-scale convolutional branch networks to obtain the original feature branch and the residual feature branch. The improved multi-scale convolutional branch network includes multiple parallel convolutional branches, each convolutional branch has a different convolutional receptive field, and a causal guided convolutional kernel selection mechanism is introduced in each parallel convolutional branch.

[0011] S5. Merge the original feature branches and the residual feature branches to obtain the fused feature representation;

[0012] S6. Obtain anomaly scores based on fused feature representations;

[0013] S7. Compare the abnormal score with the preset threshold and output the air conditioner fault warning command;

[0014] S8. When the abnormal score value does not reach the preset threshold, the current multi-channel timing input data is used as a new fault-free sample to update the normal state prototype sample.

[0015] Optionally, S1 specifically includes:

[0016] S11. During the operation of the air conditioning equipment, multi-channel operating data of the air conditioning equipment is collected through multiple sensors;

[0017] S12. Arrange the multi-channel running data in chronological order according to a uniform sampling time interval to obtain a continuous raw sampling sequence, forming multi-channel raw running data;

[0018] S13. Set a fixed time window length, segment the multi-channel raw running data according to the time window, and use the data in each time window as the current input sample to form the current multi-channel time series input data;

[0019] S14. Based on the operation and maintenance records, filter out known samples that have not experienced any faults in the historical operation data of the air conditioning equipment to obtain historical fault-free operation data samples.

[0020] Optionally, S2 specifically includes:

[0021] S21. The historical fault-free operation data sample is segmented according to the same time window length as the current multi-channel timing input data. Each segment forms a set of historical fault-free multi-channel timing input data. All segmented data form a set of historical fault-free multi-channel timing input data.

[0022] S22. For all data in the historical fault-free multi-channel timing input data set, calculate the arithmetic mean of all segments corresponding to the channels and time points at each channel and at each time point, and use it as the data value of the normal state prototype sample at that channel and time point.

[0023] S23. Combine all channels and the arithmetic mean of all time points in the original order to form a normal state prototype sample. The number of channels and the duration of the normal state prototype sample are consistent with the current multi-channel time-series input data.

[0024] Optionally, S4 specifically includes:

[0025] S41. Input the current multi-channel temporal input data and prototype residual coding features into two structurally consistent improved multi-scale convolutional branch networks. Each network contains several parallel convolutional branches, and each branch uses a different convolutional kernel size.

[0026] S42. For the current input data, construct a matrix X from the data of all channels within each time window. The number of rows in the matrix is ​​n (the number of samples), and the number of columns is d (the number of channels). Using X as input, learn the causal weight matrix A by minimizing the following optimization objective:

[0027] Minimize: L(X,A) + λ·h(A);

[0028] Where λ is the regularization coefficient, and the loss function L(X,A) is the least squares reconstruction error of all channel variables, defined as:

[0029] L(X,A)=(1 / 2n)×||XX·A|| 2 ;

[0030] The constraint term h(A) is an acyclic structural constraint on A, and it is defined as follows:

[0031]

[0032] Where tr is the trace of a matrix and exp is the exponentiation of a matrix. Represents the square of the elements of A;

[0033] S43. For each output channel, extract the elements of the corresponding row in the causal weight matrix as a branch weight vector. The length of the branch weight vector is equal to the preset number of parallel convolution branches.

[0034] S44. In each convolution branch, perform a one-dimensional convolution operation on the input data to obtain the corresponding branch convolution output;

[0035] S45. For each time point, the outputs of each convolutional branch are weighted and combined according to the weights of the branch weighting vector to obtain the final output features at that time point.

[0036] S46. Arrange the final output features of all time points in chronological order and channel order to form the feature outputs of the original feature branch and the residual feature branch respectively.

[0037] Optionally, S5 specifically includes:

[0038] S51. Arrange the original feature branches and residual feature branches in chronological order to obtain the original feature branch sequence and the residual feature branch sequence;

[0039] S52. At the same time step, the features at corresponding positions of the original feature branch sequence and the residual feature branch sequence are concatenated to form a joint feature sequence;

[0040] S53. Perform a linear transformation on the feature dimension of the joint feature sequence, and generate a gated weight vector for each time step using the sigmoid activation function.

[0041] S54. The gating weight vector is weighted element-wise with the features of the original feature branch sequence and the residual feature branch sequence at the same time step to obtain the weighted original features and the weighted residual features respectively.

[0042] S55. The weighted original features and the weighted residual features are added together at the same time step to obtain the fusion feature at that time step. The fusion features of all time steps are arranged in order to obtain the fusion feature representation.

[0043] Optionally, S6 specifically includes:

[0044] S61. Arrange the fused feature representations in chronological order to obtain the fused feature sequence;

[0045] S62. Normalize the fused feature sequence so that the value range of each feature dimension is within a preset range.

[0046] S63. Input the normalized fused feature sequence into an anomaly scoring network composed of a multi-layer fully connected neural network to extract high-dimensional feature representations;

[0047] S64. At the output of the anomaly scoring network, a sigmoid activation function is connected to map the feature representation into a single score value or a multi-dimensional score vector.

[0048] S65. Apply pooling operation to the scoring results at each time step to obtain the final anomaly score value of the current input sample.

[0049] S66. Output the abnormal rating value.

[0050] Optionally, S7 specifically includes:

[0051] S71. Compare the abnormal score value corresponding to the current input sample with the preset threshold;

[0052] S72. When the abnormal score value is greater than or equal to the preset threshold, an air conditioning fault warning command is generated.

[0053] S73, Output the air conditioner fault warning command.

[0054] Optionally, S8 specifically includes:

[0055] S81. When the abnormal score value is less than the preset threshold, the current multi-channel timing input data will be used as a new fault-free sample.

[0056] S82. Add the newly added fault-free samples to the historical fault-free operation data sample set;

[0057] S83. Based on the updated set of historical fault-free operation data samples, reselect all or part of the samples and update the normal state prototype samples.

[0058] Optionally, a deep learning-based air conditioning fault early warning system includes the following modules:

[0059] The data acquisition module is used to collect multi-channel operating data of the air conditioning equipment, construct the current multi-channel time sequence input data within a fixed time window, and generate historical fault-free operating data samples.

[0060] The prototype sample construction module is used to select all or part of the historical fault-free operation data samples and define them as normal state prototype samples.

[0061] The prototype residual coding module is used to subtract the current multi-channel time-series input data from the normal state prototype sample element by element in both the channel and time dimensions to generate prototype residual coding features.

[0062] The feature extraction module is used to input the current multi-channel temporal input data and the prototype residual coding features into two structurally consistent improved multi-scale convolutional branch networks to obtain the original feature branch and the residual feature branch, and introduces a causal guided convolution kernel selection mechanism in each parallel convolutional branch;

[0063] The feature fusion module is used to fuse the original feature branches and the residual feature branches to obtain the fused feature representation;

[0064] Anomaly scoring module, used to obtain anomaly scores based on fused feature representations;

[0065] The fault diagnosis module is used to compare the abnormal score value with the preset threshold and output the air conditioner fault warning command.

[0066] The sample update module is used to add new fault-free samples as the current multi-channel time-series input data when the abnormal score value does not reach the preset threshold, and to update the normal state prototype sample.

[0067] The beneficial effects of this invention are:

[0068] (1) By constructing a dual-branch input of multi-channel time-series input data and prototype residual coding features, it is possible to capture the normal and abnormal differences in the operation of air conditioning equipment more comprehensively and improve the richness of feature expression.

[0069] (2) An improved multi-scale convolutional branch network is adopted, and a causal guided convolutional kernel selection mechanism is introduced in each branch to achieve adaptive extraction and combination of multi-scale features, effectively enhancing the ability to identify complex working conditions and diverse abnormal patterns.

[0070] (3) Utilize historical fault-free operation data samples to dynamically construct normal state prototype samples, and combine them with the sample update mechanism to enable the model to adaptively track the health status of the equipment, thereby improving its adaptability and generalization ability to new operating conditions and potential anomalies.

[0071] (4) The feature fusion method innovatively fuses the original feature branches and the residual feature branches, further improving the accuracy and robustness of anomaly detection;

[0072] (5) The anomaly scoring and early warning instruction output process based on deep learning network can realize end-to-end intelligent fault detection and efficient early warning, reduce the risk of false alarms and missed alarms, and improve the operation safety and maintenance efficiency of air conditioning system. Attached Figure Description

[0073] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0074] Figure 1 This is an overall flowchart of a deep learning-based air conditioning fault early warning method proposed in this invention;

[0075] Figure 2 This is a schematic diagram of the structure of an air conditioning fault early warning system based on deep learning proposed in this invention. Detailed Implementation

[0076] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0077] refer to Figure 1 A deep learning-based method for early warning of air conditioning faults includes the following steps:

[0078] S1. Collect multi-channel operating data of the air conditioning equipment, construct the current input sample within a fixed time window, form the current multi-channel time-series input data, and use the known samples that have not experienced faults in the historical operation process as historical fault-free operating data samples.

[0079] S2. Select all or part of the historical fault-free operation data samples and define them as normal state prototype samples;

[0080] S3. Subtract the current multi-channel time-series input data from the normal state prototype sample element by element in both the channel and time dimensions to generate prototype residual coding features.

[0081] S4. Input the current multi-channel temporal input data and the prototype residual coding features into two structurally consistent improved multi-scale convolutional branch networks to obtain the original feature branch and the residual feature branch. The improved multi-scale convolutional branch network includes multiple parallel convolutional branches, each convolutional branch has a different convolutional receptive field, and a causal guided convolutional kernel selection mechanism is introduced in each parallel convolutional branch.

[0082] S5. Merge the original feature branches and the residual feature branches to obtain the fused feature representation;

[0083] S6. Obtain anomaly scores based on fused feature representations;

[0084] S7. Compare the abnormal score with the preset threshold and output the air conditioner fault warning command;

[0085] S8. When the abnormal score value does not reach the preset threshold, the current multi-channel timing input data is used as a new fault-free sample to update the normal state prototype sample.

[0086] In this embodiment, S1 specifically includes:

[0087] S11. During the operation of the air conditioning equipment, multi-channel operating data of the air conditioning equipment is collected through multiple sensors;

[0088] S12. Arrange the multi-channel running data in chronological order according to a uniform sampling time interval to obtain a continuous raw sampling sequence, forming multi-channel raw running data;

[0089] S13. Set a fixed time window length, segment the multi-channel raw running data according to the time window, and use the data in each time window as the current input sample to form the current multi-channel time series input data;

[0090] S14. Based on the operation and maintenance records, filter out known samples that have not experienced any faults in the historical operation data of the air conditioning equipment to obtain historical fault-free operation data samples.

[0091] This implementation method collects multi-channel operating data and filters historical fault-free samples to form current multi-channel timing input data and historical fault-free operating data samples, thereby improving the completeness and timing consistency of air conditioning operating status information and ensuring the accuracy and reliability of the data.

[0092] In this embodiment, S2 specifically includes:

[0093] S21. The historical fault-free operation data sample is segmented according to the same time window length as the current multi-channel timing input data. Each segment forms a set of historical fault-free multi-channel timing input data. All segmented data form a set of historical fault-free multi-channel timing input data.

[0094] S22. For all data in the historical fault-free multi-channel timing input data set, calculate the arithmetic mean of all segments corresponding to the channels and time points at each channel and at each time point, and use it as the data value of the normal state prototype sample at that channel and time point.

[0095] S23. Combine all channels and the arithmetic mean of all time points in the original order to form a normal state prototype sample. The number of channels and the duration of the normal state prototype sample are consistent with the current multi-channel time-series input data.

[0096] This implementation method segments the data into time windows to obtain a historical set of fault-free multi-channel time-series input data, and calculates the arithmetic mean of each segment's data at each channel and each time point. This approach fully extracts and integrates common features from a large amount of healthy operation data, effectively eliminating the randomness and noise interference of individual samples, thereby improving the sensitivity and robustness of the early warning method to abnormal changes.

[0097] In this embodiment, S4 specifically includes:

[0098] S41. Input the current multi-channel temporal input data and prototype residual coding features into two structurally consistent improved multi-scale convolutional branch networks. Each network contains several parallel convolutional branches, and each branch uses a different convolutional kernel size.

[0099] S42. For the current input data, construct a matrix X from the data of all channels within each time window. The number of rows in the matrix is ​​n (the number of samples), and the number of columns is d (the number of channels). Using X as input, learn the causal weight matrix A by minimizing the following optimization objective:

[0100] Minimize: L(X,A) + λ·h(A);

[0101] Where λ is the regularization coefficient, and the loss function L(X,A) is the least squares reconstruction error of all channel variables, defined as:

[0102] L(X,A)=(1 / 2n)×||XX·A|| 2 ;

[0103] The constraint term h(A) is an acyclic structural constraint on A, and it is defined as follows:

[0104]

[0105] Where tr is the trace of a matrix and exp is the exponentiation of a matrix. Represents the square of the elements of A;

[0106] S43. For each output channel, extract the elements of the corresponding row in the causal weight matrix as a branch weight vector. The length of the branch weight vector is equal to the preset number of parallel convolution branches.

[0107] S44. In each convolution branch, perform a one-dimensional convolution operation on the input data to obtain the corresponding branch convolution output;

[0108] S45. For each time point, the outputs of each convolutional branch are weighted and combined according to the weights of the branch weighting vector to obtain the final output features at that time point.

[0109] S46. Arrange the final output features of all time points in chronological order and channel order to form the feature outputs of the original feature branch and the residual feature branch respectively.

[0110] This implementation introduces a causal-guided convolution kernel selection mechanism into each parallel convolution branch, achieving efficient extraction and dynamic weighted combination of multi-scale temporal features. It not only adapts to the complex correlations between different time scales and channels in air conditioning equipment operation data, but also utilizes causal relationships to guide convolution kernel selection, improving the effectiveness of feature branches.

[0111] In this embodiment, S5 specifically includes:

[0112] S51. Arrange the original feature branches and residual feature branches in chronological order to obtain the original feature branch sequence and the residual feature branch sequence;

[0113] S52. At the same time step, the features at corresponding positions of the original feature branch sequence and the residual feature branch sequence are concatenated to form a joint feature sequence;

[0114] S53. Perform a linear transformation on the feature dimension of the joint feature sequence, and generate a gated weight vector for each time step using the sigmoid activation function.

[0115] S54. The gating weight vector is weighted element-wise with the features of the original feature branch sequence and the residual feature branch sequence at the same time step to obtain the weighted original features and the weighted residual features respectively.

[0116] S55. The weighted original features and the weighted residual features are added together at the same time step to obtain the fusion feature at that time step. The fusion features of all time steps are arranged in order to obtain the fusion feature representation.

[0117] This implementation uses a gating mechanism to perform weighted fusion of joint feature sequences, which fully combines the original features and residual information, effectively suppresses the interference of irrelevant or redundant features, highlights key abnormal features, and further improves the discrimination ability of fused features, thereby enhancing the accuracy of subsequent anomaly scoring and the adaptability of the fault early warning model to complex scenarios.

[0118] In this embodiment, S6 specifically includes:

[0119] S61. Arrange the fused feature representations in chronological order to obtain the fused feature sequence;

[0120] S62. Normalize the fused feature sequence so that the value range of each feature dimension is within a preset range.

[0121] S63. Input the normalized fused feature sequence into an anomaly scoring network composed of a multi-layer fully connected neural network to extract high-dimensional feature representations;

[0122] S64. At the output of the anomaly scoring network, a sigmoid activation function is connected to map the feature representation into a single score value or a multi-dimensional score vector.

[0123] S65. Apply pooling operation to the scoring results at each time step to obtain the final anomaly score value of the current input sample.

[0124] S66. Output the abnormal rating value.

[0125] This implementation improves the model's ability to distinguish abnormal states by performing deep characterization and nonlinear mapping on the fused features. Combined with normalization and pooling, it effectively reduces the impact of data scale differences and local anomaly fluctuations on the results, thereby making the anomaly score more stable and reliable.

[0126] In this embodiment, S7 specifically includes:

[0127] S71. Compare the abnormal score value corresponding to the current input sample with the preset threshold;

[0128] S72. When the abnormal score value is greater than or equal to the preset threshold, an air conditioning fault warning command is generated.

[0129] S73, Output the air conditioner fault warning command.

[0130] This implementation method enables automatic anomaly detection and fault warning, allowing for rapid response and intelligent alerts to potential faults. This improves the safety and reliability of the air conditioning system, effectively reducing equipment damage and maintenance costs caused by faults, and providing users with efficient and intelligent fault management methods.

[0131] In this embodiment, S8 specifically includes:

[0132] S81. When the abnormal score value is less than the preset threshold, the current multi-channel timing input data will be used as a new fault-free sample.

[0133] S82. Add the newly added fault-free samples to the historical fault-free operation data sample set;

[0134] S83. Based on the updated set of historical fault-free operation data samples, reselect all or part of the samples and update the normal state prototype samples.

[0135] This implementation method dynamically updates the prototype samples of the normal state, enabling the model to adaptively learn the health state and continuously improve the sample library. This allows the system to reflect the latest changes in equipment operating conditions in a timely manner, enhances its adaptability to new operating conditions, and effectively improves the generalization and long-term stability of the fault early warning method.

[0136] refer to Figure 2 A deep learning-based air conditioning fault early warning system includes the following modules:

[0137] The data acquisition module is used to collect multi-channel operating data of the air conditioning equipment, construct the current multi-channel time sequence input data within a fixed time window, and generate historical fault-free operating data samples.

[0138] The prototype sample construction module is used to select all or part of the historical fault-free operation data samples and define them as normal state prototype samples.

[0139] The prototype residual coding module is used to subtract the current multi-channel time-series input data from the normal state prototype sample element by element in both the channel and time dimensions to generate prototype residual coding features.

[0140] The feature extraction module is used to input the current multi-channel temporal input data and the prototype residual coding features into two structurally consistent improved multi-scale convolutional branch networks to obtain the original feature branch and the residual feature branch, and introduces a causal guided convolution kernel selection mechanism in each parallel convolutional branch;

[0141] The feature fusion module is used to fuse the original feature branches and the residual feature branches to obtain the fused feature representation;

[0142] Anomaly scoring module, used to obtain anomaly scores based on fused feature representations;

[0143] The fault diagnosis module is used to compare the abnormal score value with the preset threshold and output the air conditioner fault warning command.

[0144] The sample update module is used to add new fault-free samples as the current multi-channel time-series input data when the abnormal score value does not reach the preset threshold, and to update the normal state prototype sample.

[0145] By integrating eight modules—data acquisition, prototype sample construction, prototype residual coding, feature extraction, feature fusion, anomaly scoring, fault identification, and sample updating—a complete deep learning-based air conditioning fault early warning system has been built. This system enables efficient acquisition and intelligent processing of air conditioning equipment operating data, dynamically generates and updates health status baselines, fully mines multi-channel time-series features and anomaly patterns, and enhances feature representation and anomaly identification capabilities through causal guidance and multi-branch fusion mechanisms, achieving real-time and accurate early warning of air conditioning faults. The overall solution boasts a clear structure, high level of intelligence, strong adaptability, and excellent early warning effect, effectively improving the intelligent operation and maintenance level and operational safety of air conditioning systems.

[0146] Example 1:

[0147] To verify the feasibility of this invention in practice, it was applied to the central air conditioning system of a large commercial complex. This system comprises multiple large chillers, cooling towers, circulating water pumps, and terminal fan coil units, involving various operating parameters such as temperature, pressure, current, and valve opening. System operational stability is crucial for building energy consumption and comfort; failure to detect potential equipment problems in a timely manner could lead to a surge in energy consumption, uncontrolled regional temperatures, or even equipment damage.

[0148] Firstly, in this embodiment, the system installs various sensors, such as temperature sensors, pressure transmitters, and current transformers, on key components of the air conditioning equipment to achieve real-time acquisition of more than ten key parameters, including inlet and outlet water temperature, chilled water pressure, compressor current, ambient temperature, condensing pressure, and cooling water temperature. All sensors are connected to the building automation system, and data is uploaded to the central server at a uniform sampling frequency (e.g., once every 5 minutes). To ensure data quality, the system has built-in mechanisms for outlier removal and missing value imputation. After preprocessing, approximately 288 sets (24 hours × 12 times / hour) of multi-channel operating data can be obtained daily.

[0149] During the historical data accumulation phase, the operator retrieved operation records from the past two years that had not experienced any major failures, removed data from periods of manual maintenance or sudden anomalies, and finally compiled approximately 170,000 sets of multi-channel fault-free sample data over 600 days as the historical fault-free operation data sample of the system.

[0150] To improve the representativeness and generalization ability of the samples, the system automatically segments the above historical samples into fixed time windows (e.g., hourly), calculates the arithmetic mean of each channel at each time point, and forms the normal state prototype samples of the core equipment such as the chiller units in this project. Taking 10 key parameters as an example, the normal state prototype sample contains the mean of 60 sampling points for each parameter within a one-hour window, forming a 10×60 prototype feature matrix.

[0151] During the actual operation of the system, the real-time collected multi-channel time-series input data is subtracted element-by-element from the normal state prototype sample to generate prototype residual coding features. Taking the compressor outlet water temperature of a chiller unit as an example, the difference between the currently collected 60-minute temperature sequence and the temperature sequence at the same position in the prototype sample is used to obtain the residual curve, which reflects the subtle differences between the equipment operation and normal state.

[0152] The system's built-in improved multi-scale convolutional branch network inputs the current multi-channel temporal input data and prototype residual coding features into two independent branch networks. Each branch contains three sets of convolutional kernels with different receptive fields (e.g., 3, 5, and 7 sampling points) to extract short-term, periodic, and trend features, respectively. Within each parallel convolutional branch, a causal-guided convolutional kernel selection mechanism is introduced. This mechanism dynamically allocates the participation of different convolutional branches in feature fusion by modeling the causal weights of the input data, achieving adaptive combination of multi-channel and multi-timescale information.

[0153] After the outputs of the two feature branches are concatenated at the same time step, the system employs a gated fusion method. Gating weights are generated using sigmoid activation, which weights the original and residual features, automatically emphasizing the feature combination that best reflects the anomaly. The fused temporal features are fed into a multi-layer fully connected neural network, and after normalization and sigmoid activation, the anomaly score for each time step is output. Finally, max pooling is applied to all scores to output the final anomaly score for the current input sample.

[0154] To achieve full lifecycle adaptability, when the system determines that the current abnormal score is lower than the set threshold, it automatically includes the current input data into the historical fault-free sample library and dynamically updates the normal state prototype sample accordingly, so that the model can adapt to seasonal changes, environmental changes and equipment aging in a timely manner.

[0155] Taking the operation of a chiller unit during the high temperature and high load period in July and August as an example, the return water temperature data of the air conditioning unit collected in the first month after the system went online is shown in the table below (some key parameters):

[0156] Table 1. Statistical Table of Typical Operating Parameters and Fault Early Warning Results of Chiller Units

[0157] Run Date Average return water temperature (°C) Average current (A) Abnormal score Early warning judgment 2023-07-01 12.3 78.1 0.11 normal 2023-07-02 12.4 77.9 0.12 normal 2023-07-10 12.9 81.3 0.19 normal 2023-07-19 14.7 91.8 0.42 Warning 2023-07-20 14.8 92.2 0.45 Warning 2023-07-21 15.1 93.0 0.48 Warning 2023-07-22 12.5 78.0 0.14 normal

[0158] Between July 19th and 21st, 2023, the chiller unit experienced a significant increase in return water temperature and current, with an anomaly score exceeding 0.4 (the system threshold was 0.3). The system automatically issued a fault warning, prompting maintenance personnel to conduct timely inspections. Inspection revealed scaling on the condenser heat exchange surface, reducing cooling efficiency. After timely treatment, the return water temperature and current returned to normal, the anomaly score decreased, and the warning was lifted. Under normal circumstances, the anomaly score remains stable between 0.1 and 0.2, effectively preventing false alarms.

[0159] Statistical analysis of operational data from the first six months after the system's launch revealed that the model can accurately detect various faults in chiller units, including condenser scaling, circulating pump malfunctions, and sensor failures. Furthermore, it maintains a high recognition rate even under new operating conditions such as changes in ambient temperature and load fluctuations. Compared to traditional threshold methods and single-feature detection methods, the deep learning model's average early warning accuracy has increased to 97.6%, the false negative rate has decreased to 0.8%, and the false positive rate has decreased to 1.6%. The system can also adaptively update using historical health samples, ensuring broad adaptability to different seasons and equipment operating cycles.

[0160] Furthermore, the system can simultaneously monitor the operating data of multiple chillers and terminal fan coil units, and output a batch of anomaly warning lists, greatly improving the troubleshooting efficiency of maintenance personnel. On-site surveys show that under the traditional model, maintenance personnel need to conduct inspections every 2 hours; after the fully automated warning system is implemented, the inspection frequency can be reduced by more than 50%, significantly reducing labor costs.

[0161] As demonstrated in this embodiment, the deep learning-based air conditioning fault early warning system can effectively integrate multi-channel time-series information, historical health data, and causal guidance mechanisms, fully adapting to large-scale, multi-type, and variable operating conditions in real-world scenarios. Its main beneficial effects are reflected in the following aspects: it can significantly improve the accuracy and response speed of fault early warning, significantly reduce the risk of false alarms and missed alarms, enable the model to continuously adapt to new operating conditions, improve the safety and energy efficiency of the air conditioning system, and significantly optimize the operation and maintenance management process and human resource allocation, promoting the digital and intelligent upgrade of building intelligent operation and maintenance.

[0162] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based method for early warning of air conditioning faults, characterized in that, Includes the following steps: S1. Collect multi-channel operating data of the air conditioning equipment, construct the current input sample within a fixed time window, form the current multi-channel time-series input data, and use the known samples that have not experienced faults in the historical operation process as historical fault-free operating data samples. S2. Select all or part of the historical fault-free operation data samples and define them as normal state prototype samples; S3. Subtract the current multi-channel time-series input data from the normal state prototype sample element by element in both the channel and time dimensions to generate prototype residual coding features. S4. Input the current multi-channel temporal input data and the prototype residual coding features into two structurally consistent improved multi-scale convolutional branch networks to obtain the original feature branch and the residual feature branch. The improved multi-scale convolutional branch network includes multiple parallel convolutional branches, each convolutional branch has a different convolutional receptive field, and a causal guided convolutional kernel selection mechanism is introduced in each parallel convolutional branch. S5. Merge the original feature branches and the residual feature branches to obtain the fused feature representation; S6. Obtain anomaly scores based on fused feature representations; S7. Compare the abnormal score with the preset threshold and output the air conditioner fault warning command; S8. When the abnormal score value does not reach the preset threshold, the current multi-channel timing input data is used as a new fault-free sample to update the normal state prototype sample.

2. The air conditioning fault early warning method based on deep learning according to claim 1, characterized in that, S1 specifically includes: S11. During the operation of the air conditioning equipment, multi-channel operating data of the air conditioning equipment is collected through multiple sensors; S12. Arrange the multi-channel running data in chronological order according to a uniform sampling time interval to obtain a continuous raw sampling sequence, forming multi-channel raw running data; S13. Set a fixed time window length, segment the multi-channel raw running data according to the time window, and use the data in each time window as the current input sample to form the current multi-channel time series input data; S14. Based on the operation and maintenance records, filter out known samples that have not experienced any faults in the historical operation data of the air conditioning equipment to obtain historical fault-free operation data samples.

3. The air conditioning fault early warning method based on deep learning according to claim 1, characterized in that, S2 specifically includes: S21. The historical fault-free operation data sample is segmented according to the same time window length as the current multi-channel timing input data. Each segment forms a set of historical fault-free multi-channel timing input data. All segmented data form a set of historical fault-free multi-channel timing input data. S22. For all data in the historical fault-free multi-channel timing input data set, calculate the arithmetic mean of all segments corresponding to the channels and time points at each channel and at each time point, and use it as the data value of the normal state prototype sample at that channel and time point. S23. Combine all channels and the arithmetic mean of all time points in the original order to form a normal state prototype sample. The number of channels and the duration of the normal state prototype sample are consistent with the current multi-channel time-series input data.

4. The air conditioning fault early warning method based on deep learning according to claim 1, characterized in that, S4 specifically includes: S41. Input the current multi-channel temporal input data and prototype residual coding features into two structurally consistent improved multi-scale convolutional branch networks. Each network contains several parallel convolutional branches, and each branch uses a different convolutional kernel size. S42. For the current input data, construct a matrix X from the data of all channels within each time window. The number of rows in the matrix is ​​n (the number of samples), and the number of columns is d (the number of channels). Using X as input, learn the causal weight matrix A by minimizing the following optimization objective: Minimize: L(X,A) + λ·h(A); Where λ is the regularization coefficient, and the loss function L(X,A) is the least squares reconstruction error of all channel variables, defined as: L(X,A)=(1 / 2n)×||X-X·A|| 2 ; The constraint term h(A) is an acyclic structural constraint on A, and it is defined as follows: Where tr is the trace of a matrix and exp is the exponentiation of a matrix. This represents the square of an element of A. S43. For each output channel, extract the elements of the corresponding row in the causal weight matrix as a branch weight vector. The length of the branch weight vector is equal to the preset number of parallel convolution branches. S44. In each convolution branch, perform a one-dimensional convolution operation on the input data to obtain the corresponding branch convolution output; S45. For each time point, the outputs of each convolutional branch are weighted and combined according to the weights of the branch weighting vector to obtain the final output features at that time point. S46. Arrange the final output features of all time points in chronological order and channel order to form the feature outputs of the original feature branch and the residual feature branch respectively.

5. The air conditioning fault early warning method based on deep learning according to claim 1, characterized in that, S5 specifically includes: S51. Arrange the original feature branches and residual feature branches in chronological order to obtain the original feature branch sequence and the residual feature branch sequence; S52. At the same time step, the features at corresponding positions of the original feature branch sequence and the residual feature branch sequence are concatenated to form a joint feature sequence; S53. Perform a linear transformation on the feature dimension of the joint feature sequence, and generate a gated weight vector for each time step using the sigmoid activation function. S54. The gating weight vector is weighted element-wise with the features of the original feature branch sequence and the residual feature branch sequence at the same time step to obtain the weighted original features and the weighted residual features respectively. S55. The weighted original features and the weighted residual features are added together at the same time step to obtain the fusion feature at that time step. The fusion features of all time steps are arranged in order to obtain the fusion feature representation.

6. The air conditioning fault early warning method based on deep learning according to claim 1, characterized in that, S6 specifically includes: S61. Arrange the fused feature representations in chronological order to obtain the fused feature sequence; S62. Normalize the fused feature sequence so that the value range of each feature dimension is within a preset range. S63. Input the normalized fused feature sequence into an anomaly scoring network composed of a multi-layer fully connected neural network to extract high-dimensional feature representations; S64. At the output of the anomaly scoring network, a sigmoid activation function is connected to map the feature representation into a single score value or a multi-dimensional score vector. S65. Apply pooling operation to the scoring results at each time step to obtain the final anomaly score value of the current input sample. S66. Output the abnormal rating value.

7. The air conditioning fault early warning method based on deep learning according to claim 1, characterized in that, Specifically, S7 includes: S71. Compare the abnormal score value corresponding to the current input sample with the preset threshold; S72. When the abnormal score value is greater than or equal to the preset threshold, an air conditioning fault warning command is generated. S73, Output the air conditioner fault warning command.

8. The air conditioning fault early warning method based on deep learning according to claim 1, characterized in that, S8 specifically includes: S81. When the abnormal score value is less than the preset threshold, the current multi-channel timing input data will be used as a new fault-free sample. S82. Add the newly added fault-free samples to the historical fault-free operation data sample set; S83. Based on the updated set of historical fault-free operation data samples, reselect all or part of the samples and update the normal state prototype samples.

9. A deep learning-based air conditioning fault early warning system, applied to the deep learning-based air conditioning fault early warning method according to any one of claims 1-8, characterized in that, Includes the following modules: The data acquisition module is used to collect multi-channel operating data of the air conditioning equipment, construct the current multi-channel time sequence input data within a fixed time window, and generate historical fault-free operating data samples. The prototype sample construction module is used to select all or part of the historical fault-free operation data samples and define them as normal state prototype samples. The prototype residual coding module is used to subtract the current multi-channel time-series input data from the normal state prototype sample element by element in both the channel and time dimensions to generate prototype residual coding features. The feature extraction module is used to input the current multi-channel temporal input data and the prototype residual coding features into two structurally consistent improved multi-scale convolutional branch networks to obtain the original feature branch and the residual feature branch, and introduces a causal guided convolution kernel selection mechanism in each parallel convolutional branch; The feature fusion module is used to fuse the original feature branches and the residual feature branches to obtain the fused feature representation; Anomaly scoring module, used to obtain anomaly scores based on fused feature representations; The fault diagnosis module is used to compare the abnormal score value with the preset threshold and output the air conditioner fault warning command. The sample update module is used to add new fault-free samples as the current multi-channel time-series input data when the abnormal score value does not reach the preset threshold, and to update the normal state prototype sample.

Images