Multi-sensor fusion heat pump full life cycle AI maintenance early warning system

Abstract

The invention discloses a multi-sensor fusion heat pump full life cycle AI maintenance early warning system, and relates to the technical field of new energy utilization, and the early warning system comprises a data collection module which obtains operation parameters in a heat pump full life cycle based on a sensor array, and constructs a data set after preprocessing the parameters; the operation parameters comprise temperature, pressure, flow and micro vibration; the data fusion module is used for extracting trend correlation characteristics and parameter coupling characteristics from temperature, pressure and flow parameters by adopting a dynamic sliding window adaptive to a working condition, and preserving core nonlinear information through KPCA dimension reduction; the micro-vibration signal extraction comprises frequency domain and time domain features. According to the method, features are extracted through a working condition adaptive dynamic sliding window, then through cross-space mapping and a life cycle-working condition double-attention mechanism, the analysis and early warning module depends on a core feature mapping library and a two-dimensional dynamic baseline, through instantaneous and accumulated deviation double judgment, abnormal accurate recognition and stage division are achieved, and early warning perspectiveness is high.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of new energy utilization, in particular to a multi-sensor fusion heat pump full life cycle AI maintenance early warning system. BACKGROUND

[0002] As a kind of efficient energy conversion equipment, heat pump is widely used in building heating, industrial waste heat recovery, hot water supply and other fields, and its running stability and service life directly affect energy utilization efficiency and user cost. With the diversification of heat pump application scenarios and the extension of operation cycle, equipment wear and tear, pipeline aging and working condition fluctuation in the full life cycle can easily cause compressor failure, heat exchanger blockage, pipeline leakage and other problems. If not timely warned and maintained, it may cause equipment downtime, energy consumption increase and even safety accidents. The existing heat pump maintenance and early warning technology relies on single sensor data or fixed threshold determination, which has obvious limitations, insufficient data support dimension, and no differentiated solution designed for the characteristics of different types of parameters. There is a complex nonlinear correlation and synergistic effect between temperature, pressure and flow parameters. Micro-vibration signals contain key information of local faults such as equipment mechanical wear and tear and pipeline loosening. After dimension reduction processing, conventional temperature, pressure and flow parameters and micro-vibration signals are often in independent feature space, lacking effective cross-space adaptive mapping mechanism. Due to the essential difference between the physical meaning and data distribution of the two types of parameters, directly splicing or simply weighting the features will cause feature conflict and information redundancy, which cannot realize the collaborative complementation of multi-source data, resulting in the loss of core fault information and the inability to provide effective support for subsequent fault identification. Therefore, there is a need for an effective multi-source data fusion method. To solve the above technical problems, a multi-sensor fusion heat pump full life cycle AI maintenance early warning system is provided. SUMMARY

[0003] To solve the above technical problems, a multi-sensor fusion heat pump full life cycle AI maintenance early warning system is provided. The technical scheme solves the problem of lacking effective cross-space adaptive mapping mechanism.

[0004] To achieve the above purposes, the technical scheme adopted by the application is as follows: a multi-sensor fusion heat pump full life cycle AI maintenance early warning system, the early warning system comprising:

[0005] A data acquisition module acquires running parameters in the full life cycle of the heat pump based on a sensor array, and constructs a data set after preprocessing the parameters. The running parameters include temperature, pressure, flow and micro-vibration.

[0006] The data fusion module extracts trend correlation features and parameter coupling features using a dynamic sliding window adapted to the working condition, and reduces the dimensionality by KPCA to retain the core nonlinear information. The micro-vibration signal extraction includes frequency domain and time domain features, which are reduced in dimension by LLE to retain the local topological structure, and the parameter features after dimensionality reduction are mapped to the same feature space. The parameter feature weight is dynamically adjusted by the life cycle and working condition double attention mechanism to highlight the strongly correlated features of the fault;

[0007] The analysis and early warning module filters the fault intensity related features, establishes a mapping library of fault types and core features, constructs a dynamic baseline in the life cycle and working condition, calculates the instantaneous and cumulative deviations of real-time features and baseline, and double determines the suspected abnormality and development stage. The analysis results are output according to the deviation degree;

[0008] The abnormality early warning module displays the early warning information based on the analysis results, outputs the maintenance plan, and establishes a health record for the entire life cycle of the heat pump.

[0009] Preferably, the sensor array is deployed in the compressor, heat exchanger, throttling device and pump group running nodes. Temperature is collected by temperature sensors, pressure is collected by piezoelectric pressure sensors, flow is collected by electromagnetic flowmeters, and micro-vibration signals are collected by piezoelectric acceleration sensors. Wavelet transform algorithm is used for pre-processing to suppress noise and eliminate irrelevant noise caused by electromagnetic interference and mechanical interference. The data is completed based on the interpolation method of time series correlation. Normalization is performed by Z-score.

[0010] Preferably, the dynamic sliding window adaptation rule for temperature, pressure and flow parameters is as follows:

[0011] Based on the load rate, start-stop signal and parameter fluctuation characteristics of the heat pump, the running state is divided into start-stop transition, steady-state operation and variable load working condition, and the temperature, pressure and flow data are grouped according to the working condition.

[0012] Calculate the indicators for each group of data, including the average parameter change rate, parameter fluctuation intensity and working condition mutation identifier, and characterize the working condition based on the indicators.

[0013] The dynamic sliding window selection adjustment is as follows: in steady-state working condition, use a large window of 80-120 sampling points and a large step of 40-60 sampling points; in variable load working condition, use a medium window of 40-80 sampling points and a medium step of 20-40 sampling points; in start-stop transition / mutation working condition, use a small window of 20-40 sampling points and a small step of 5-10 sampling points.

[0014] Preferably, for temperature, pressure and flow parameters, trend correlation features are extracted in the dynamic sliding window of the adaptive working condition; the trend correlation features include single parameter time evolution features and multi-parameter collaborative operation features; wherein the single parameter time evolution features specifically include parameter change slope, trend stability index, extreme value feature and cumulative change amount, which depict the time evolution law of a single parameter; the multi-parameter collaborative operation features specifically include parameter change direction consistency coefficient, trend synchronization delay, trend strength synergy degree and abnormal trend proportion, which represent the collaborative mechanism between multiple parameters;

[0015] For all two-parameter combinations of temperature, pressure and flow, a linear coupling feature extraction mechanism is constructed, and in the dynamic sliding window of the corresponding working condition, the linear correlation coefficient of each two-parameter combination is calculated to quantitatively represent the linear correlation degree of the time series data; combined with the consistency quantization result of the parameter change direction, the linear coupling relationship between parameters is represented, and the linear correlation of the two-parameter combination is captured.

[0016] Preferably, the KPCA dimensionality reduction step for retaining core nonlinear information is:

[0017] The extracted trend correlation features and parameter coupling features are collected and integrated to construct a high-dimensional feature unified matrix; the local outlier factor (LOF) algorithm is used to identify and correct the abnormal feature values in the high-dimensional feature unified matrix;

[0018] The radial basis function (RBF) kernel function is selected as the core mapping function for the nonlinear correlation characteristics of the adaptive parameters, and the 5-fold cross-validation and grid search parameter optimization strategy is used to iteratively select the kernel parameters of the RBF kernel function to determine the optimal kernel parameters;

[0019] Based on the optimal kernel parameters, the similarity between each sample in the high-dimensional feature unified matrix is calculated to construct a sample similarity kernel matrix; the sample similarity kernel matrix is subjected to center correction processing to obtain a standardized and centered kernel matrix;

[0020] The eigenvalue decomposition operation is performed on the centered kernel matrix to obtain an eigenvalue sequence and corresponding eigenvectors; based on the eigenvalue sequence, the cumulative contribution rate of each principal component is calculated, the cumulative contribution rate threshold is set to be greater than or equal to 85%, the effective principal components that meet the threshold requirement are selected, and the target dimensionality reduction dimension is determined; based on the selected effective principal components, a low-dimensional projection matrix is constructed to map the high-dimensional features to a low-dimensional kernel space, and the dimensionality reduction processing of the high-dimensional features is completed.

[0021] Preferably, the step of mapping the micro-vibration signal to the same feature space is:

[0022] A time-frequency domain joint feature extraction mechanism is used to extract 10-dimensional time domain features, and fast Fourier transform (FFT) and short-time Fourier transform (STFT) are performed on the micro-vibration signal to extract 10-dimensional frequency domain features;

[0023] The time domain features and the frequency domain features are integrated to construct a high-dimensional feature matrix; and the high-dimensional feature matrix is optimized to a 12-16 dimensional effective feature matrix;

[0024] The number of neighbors and the target dimension range of dimension reduction are initialized, the local neighbor set of each sample is determined through neighbor searching, and the optimal local reconstruction weight matrix is solved; the reconstruction weight relationship consistent with that in the high-dimensional space is maintained in the low-dimensional space; the final dimension reduction dimension is determined based on a screening criterion that the cumulative contribution rate is greater than or equal to 85%, and a low-dimensional feature matrix is output.

[0025] Preferably, the parameter features after dimension reduction are mapped to the same feature space through kernel space metric learning, the adaptability mapping relationship of the two types of features is learned based on the historical fault sample library of the whole life cycle of the heat pump, and the KPCA low-dimensional features and the LLE low-dimensional features are simultaneously projected into the same high-dimensional kernel space;

[0026] Preferably, the life cycle attention branch performs encoding processing on the heat pump whole life cycle stage, constructs a stage and feature correlation matrix based on the historical fault sample library of the whole life cycle, calculates the initial weight of each parameter feature based on a preset attention weight network, and the working condition attention branch performs quantitative encoding processing on the real-time running working condition to generate a dynamic feature weight matrix corresponding to the real-time working condition.

[0027] The weights output by the two branches are weighted and summed based on a preset fusion coefficient a to perform collaborative fusion of the double-branch weights.

[0028] Preferably, the analysis and early warning module analyzes and early warns through fault intensity related feature screening, quantifies the discriminability of features to the compressor wear and pipeline leakage fault types based on a random forest feature contribution degree algorithm, screens out a core feature subset with an importance score greater than or equal to 0.7, establishes a dynamic mapping library of fault types and core features, constructs a life cycle and working condition double-dimensional dynamic baseline, constructs normal feature baselines of each dimension based on historical normal samples through statistical modeling method according to the stage and working condition combined dimension, and includes mean and fluctuation range parameters; the instantaneous deviation of real-time features from the baseline is calculated by Euclidean distance, and the preliminary abnormality is determined when the deviation exceeds the 95% confidence interval; the cumulative deviation is obtained by weighted sum of the instantaneous deviation index through a sliding window to divide different stages of abnormality budding period, development period and severe period; different early warning measures are given based on different stages.

[0029] Preferably, the standard for dividing abnormal stages is:

[0030] When the cumulative deviation is greater than or equal to 1.2 times and less than 1.8 times of the upper limit of the baseline fluctuation range in the abnormality budding period, only the instantaneous deviation occasionally exceeds the threshold value, and there is no obvious sustained trend.

[0031] Cumulative deviation ≥ 1.8 times and < 3.0 times during the abnormal development period, frequent threshold crossing of instantaneous deviation, and continuously rising deviation trend;

[0032] Cumulative deviation ≥ 3.0 times and / or instantaneous deviation suddenly rises to more than 3 times the upper limit of baseline fluctuation during the abnormal severe period, continuously expanding deviation, and fault entries in the core feature matching mapping library.

[0033] Compared with the prior art, the present application has the following beneficial effects:

[0034] The present application extracts features with a working condition adaptive dynamic sliding window, combines KPCA and LLE differential dimension reduction, and then accurately refines fault strongly related features through cross-space mapping and life cycle-working condition dual attention mechanisms, avoids information loss and interference, and relies on the core feature mapping library and dual-dimensional dynamic baseline in the analysis and early warning module to realize accurate identification and stage division of abnormalities through instantaneous and cumulative deviation double determination, which has strong early warning foresight, solves the pain points of traditional early warning feature distortion, high false alarm rate, and operation and maintenance blindness, greatly improves the early warning accuracy and operation and maintenance efficiency, prolongs the service life of the equipment, and reduces energy consumption and operation and maintenance cost. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 The present application is a schematic diagram of a prediction system framework. DETAILED DESCRIPTION

[0036] The following description is used to disclose the present application so that those skilled in the art can implement the present application. The preferred embodiments in the following description are only used as examples, and other obvious variants can be thought of by those skilled in the art.

[0037] Referring to Figure 1 As shown in the figure, a multi-sensor fusion heat pump full life cycle AI maintenance and early warning system, the early warning system comprises:

[0038] A data acquisition module acquires running parameters in the full life cycle of the heat pump based on a sensor array, and constructs a data set after preprocessing the parameters; the running parameters include temperature, pressure, flow and micro-vibration;

[0039] A data fusion module extracts trend correlation features and parameter coupling features using a working condition adaptive dynamic sliding window for temperature, pressure and flow parameters, and retains core nonlinear information through KPCA dimension reduction; micro-vibration signal extraction includes frequency domain and time domain features, and retains local topological structure through LLE dimension reduction, and maps the dimension-reduced parameter features to the same feature space; the parameter feature weight is dynamically adjusted through the life cycle and working condition dual attention mechanism, and the fault strongly related features are highlighted.

[0040] The analysis and early warning module screens the fault intensity related features, establishes a fault type and core feature mapping library, constructs a life cycle and working condition double-dimensional dynamic baseline, calculates the instantaneous and cumulative deviations of real-time features and the baseline, and double determines suspected abnormalities and development stages; and outputs analysis results according to the deviation degree.

[0041] The abnormality early warning module displays early warning information based on the analysis results, outputs maintenance schemes, and establishes a heat pump full life cycle health record.

[0042] At the data collection level of the application, a sensor array fully covers temperature, pressure, flow and micro-vibration parameters, and is matched with preprocessing technology to realize comprehensive and high-quality data capture of the macro and micro states of the equipment. In the data fusion link, features are extracted by a working condition adaptive dynamic sliding window, combined with KPCA and LLE differential dimension reduction, and then accurately refined fault intensity related features through cross-space mapping and life cycle-working condition double attention mechanism, avoiding information loss and interference. The analysis and early warning module relies on the core feature mapping library and double-dimensional dynamic baseline to realize accurate identification and stage division of abnormalities through instantaneous and cumulative deviation double determination. The early warning module has strong forward-looking, and outputs graded early warning and targeted maintenance schemes, and synchronously establishes a health record, promoting the transition of operation and maintenance from passive repair to active prevention. The system as a whole solves the pain points of traditional early warning feature distortion, high false alarm rate and blind operation and maintenance, greatly improves the accuracy of early warning and the efficiency of operation and maintenance, prolongs the service life of equipment and reduces energy consumption and operation and maintenance costs.

[0043] The sensor array is deployed in the compressor, heat exchanger, throttling device and pump group operation nodes. Temperature is collected by a temperature sensor, pressure is collected by a piezoelectric pressure sensor, flow is collected by an electromagnetic flowmeter, and micro-vibration signals are collected by a piezoelectric acceleration sensor. Wavelet transform algorithm is used for noise suppression in preprocessing to eliminate irrelevant noise of electromagnetic interference and mechanical interference. The data is completed based on the interpolation method of time series correlation. Normalization is performed by Z-score.

[0044] The application is deployed in key operation nodes such as compressors and heat exchangers to directly collect key parameters of core components, avoid signal attenuation or information distortion caused by non-key position collection, and ensure that the data can truly reflect the core operation state of the equipment. Special sensors are matched according to the characteristics of different parameters to improve the sensitivity and accuracy of parameter collection and avoid parameter capture deviation caused by general sensors.

[0045] The working condition adaptive dynamic sliding window adaptation rule for temperature, pressure and flow parameters is as follows:

[0046] A multi-dimensional working condition classification system is constructed, which is supported by real-time operation data of the heat pump, and comprehensively classifies the working conditions based on three core indexes, i.e., heat pump load rate, start-stop signal and parameter fluctuation characteristics. The load rate directly reflects the current energy supply-demand matching state of the equipment and is the core basis for determining the working condition type. The start-stop signal is an explicit identifier of working condition switching, which quickly captures the state transition nodes from shutdown to startup or from running to shutdown. The parameter fluctuation characteristics quantify the degree of parameter change over time by calculating the standard deviation and coefficient of variation of temperature, pressure and flow parameters, and assist in verifying the stability of the working condition. Based on the above three-dimensional basis, the heat pump operation state is accurately divided into three categories: start-stop transition working condition, steady-state running working condition and variable load working condition. The start-stop transition working condition corresponds to the startup or shutdown stage of the equipment, at which time the thermodynamic equilibrium of the system has not been established, and the temperature, pressure and flow parameters present a characteristic of dramatic fluctuation and nonlinear change. The steady-state running working condition corresponds to the stage where the heat pump load demand is stable and the energy conversion efficiency is in a stable interval, and the parameter changes smoothly with small fluctuation amplitude and stable trend. The variable load working condition corresponds to the stage of dynamic adjustment of external energy demand, and the load rate presents a continuous change trend, and the temperature, pressure and flow parameters present a gradual fluctuation with a change rate between the start-stop transition and the steady-state running.

[0047] After the working condition classification is completed, the temperature, pressure and flow data under the three types of working conditions are stored in groups, and then three types of key characteristic indexes are calculated for each group of data to construct the working condition feature vector. First, the average parameter change rate is calculated, which quantitatively reflects the overall change rate of the parameter in this working condition by calculating the average change amplitude of the parameter per unit time. For example, the average change rate of the start-stop transition working condition is significantly higher than that of the steady-state working condition. Second, the parameter fluctuation intensity is calculated by using the weighted combination of variance and range to characterize the dispersion degree of the parameter in this working condition, which directly reflects the parameter stability. Third, the working condition mutation identifier is set to capture the instantaneous nodes of working condition switching by monitoring the start-stop signal jump and the average parameter change rate sudden increase / sudden decrease characteristics, ensuring the continuity and accuracy of working condition classification. Through the synergistic effect of the three types of indexes, the running characteristics of each working condition are accurately quantified and characterized, providing data support for the subsequent dynamic adjustment of the sliding window.

[0048] Based on the characterization results of each working condition, adaptive selection and adjustment of dynamic sliding window are performed. For steady-state operating conditions, due to the gentle change and stable trend of parameters without obvious mutation characteristics, a large window size of 80-120 sampling points is adopted, combined with a large step size of 40-60 sampling points, which can fully capture the long-term trend correlation characteristics of parameters, effectively reduce the redundant data processing amount, and improve the feature extraction efficiency. For variable load conditions, the parameters gradually fluctuate with the change of load, and the change rate and fluctuation amplitude are between the two, so a medium window size of 40-80 sampling points and a medium step size of 20-40 sampling points are selected, which can balance the accuracy and efficiency of feature extraction while ensuring the capture of the dynamic change trend of parameters. For start-stop transition conditions and sudden change scenarios, the parameters have characteristics of violent fluctuation and instantaneous mutation, so a small window size of 20-40 sampling points and a small step size of 5-10 sampling points are required to maximize the retention of key detail information at the moment of parameter mutation, avoid the loss of key information due to excessive window size, and ensure that the trend correlation characteristics and parameter coupling characteristics of temperature, pressure and flow parameters under different operating conditions can be truly and completely extracted, laying a solid foundation for subsequent data fusion and fault feature recognition.

[0049] Trend correlation feature extraction is divided into single parameter and multi-parameter two categories. On the single parameter level, the change slope is obtained by linear fitting to reflect the parameter trend direction and rate, and adapt to the differences of different working conditions. The deviation dispersion of data and fitting line is calculated to obtain the trend stability index, which quantifies the parameter change stability. The local extreme value is selected and the related data is counted to capture the parameter fluctuation amplitude and mutation characteristics. The cumulative change value is obtained by accumulating the absolute change value of adjacent parameters, avoiding the randomness of single time data. On the multi-parameter level, the direction consistency coefficient is obtained by counting the proportion of time when the change direction of two parameters is the same, which represents the change consistency. The trend synchronization delay is calculated by cross-correlation analysis to reflect the parameter response synchronization. The synergy degree is calculated based on the trend intensity difference to reflect the intensity matching degree. The abnormal parameter proportion is compared with the historical normal range to judge the synergy imbalance. These features can show obvious abnormalities when faults occur;

[0050] Linear coupling feature extraction is for parameter combination. First, the linear correlation strength between parameters is calculated to quantify the linear correlation tightness. Then, the comprehensive linear coupling coefficient is obtained by multiplying the strength and the change direction consistency coefficient to avoid false coupling and accurately capture the linear correlation between parameters.

[0051] The overall extraction method is based on the dynamic sliding window adapted to the working condition. Through the multi-layer design of single parameter time sequence rule, multi-parameter synergy mechanism and linear coupling relationship, it not only adapts to different operating conditions, but also fully describes the parameter change and parameter interaction, providing high-quality core features for subsequent processing.

[0052] The KPCA dimensionality reduction step for retaining core nonlinear information is:

[0053] The extracted trend-related features and parameter-coupled features are collected and integrated to construct a high-dimensional feature uniform matrix. The local outlier factor (LOF) algorithm is used to identify and correct abnormal feature values in the high-dimensional feature uniform matrix.

[0054] The radial basis function (RBF) kernel function with adaptive parameter nonlinear correlation characteristics is selected as the core mapping function. The 5-fold cross-validation and grid search combination parameter optimization strategy is used to iteratively select the kernel parameters of the RBF kernel function and determine the optimal kernel parameters.

[0055] Based on the optimal kernel parameters, the similarity between each sample in the high-dimensional feature uniform matrix is calculated to construct a sample similarity kernel matrix. The sample similarity kernel matrix is subjected to center correction processing to obtain a standardized and centered kernel matrix.

[0056] The centered kernel matrix is subjected to eigenvalue decomposition to obtain a sequence of eigenvalues and corresponding eigenvectors. The cumulative contribution rate of each principal component is calculated based on the sequence of eigenvalues. A cumulative contribution rate threshold of ≥85% is set to select effective principal components that meet the threshold requirement and determine the target dimensionality reduction dimension. Based on the selected effective principal components, a low-dimensional projection matrix is constructed to map to a low-dimensional kernel space, completing the dimensionality reduction of high-dimensional features. The cumulative contribution rate threshold of ≥85% is determined based on expert analysis.

[0057] The local outlier factor (LOF) algorithm is the core of data preprocessing and anomaly correction. After constructing the high-dimensional feature uniform matrix, the LOF algorithm is used to calculate the local density deviation of each feature sample, accurately identify abnormal feature values caused by data noise, working condition mutations, and other factors, and process the abnormal values through interpolation replacement or correction strategies to ensure the reliability of high-dimensional feature data and lay a foundation for subsequent dimensionality reduction.

[0058] The radial basis function (RBF) kernel function is specially adapted to the complex nonlinear correlation characteristics between temperature, pressure, and flow parameters. Its core function is to map the nonlinear relationships in high-dimensional feature space, which are difficult to handle directly, to a higher-dimensional kernel space, converting them into linearly processable problems, thereby breaking through the limitations of traditional linear dimensionality reduction methods that cannot capture nonlinear information.

[0059] The 5-fold cross-validation and grid search combination algorithm is used for parameter optimization of the RBF kernel function. Grid search first presets a series of kernel parameter candidate values, then splits the data set into 5 groups through 5-fold cross-validation, and iteratively verifies the dimensionality reduction effect of different candidate parameters using 4 groups as the training set and 1 group as the validation set. Finally, the kernel parameter that optimizes the feature retention rate and model generalization ability is selected to avoid parameter selection blindness.

[0060] The kernel matrix centering algorithm normalizes the constructed sample similarity kernel matrix, eliminates the mean shift problem that may occur in the kernel function mapping process, ensures that the kernel matrix satisfies the symmetry and normalization of data distribution, and provides standardized data for subsequent eigenvalue decomposition;

[0061] The eigenvalue decomposition algorithm performs decomposition operation on the centralized kernel matrix, separates the eigenvalue sequence representing the importance of the features and the corresponding eigenvectors, and the eigenvalue size directly reflects the information carrying capacity of the corresponding principal component.

[0062] The cumulative contribution rate screening algorithm calculates the cumulative contribution rate of each principal component based on the eigenvalue sequence, sets a threshold of ≥ 85%, screens the effective principal components carrying core information, determines a reasonable dimension reduction dimension, and maximizes the retention of the nonlinear core features between parameters while eliminating redundant information. Finally, the dimension reduction mapping of high-dimensional features is completed by constructing a low-dimensional projection matrix.

[0063] The core of the cumulative contribution rate screening algorithm is to quantify the information carrying capacity of each principal component and screen the effective principal components that retain the core nonlinear information. The contribution rate formula of a single principal component is:

[0064]

[0065] Where η i is the contribution rate of the i-th principal component, reflecting the proportion of the principal component carrying the original high-dimensional feature information; λ i is the i-th eigenvalue obtained after the eigenvalue decomposition of the centralized kernel matrix, and the larger the eigenvalue, the stronger the information carrying capacity of the corresponding principal component; m is the total number of eigenvalues of the centralized kernel matrix, which is related to the dimension of the original high-dimensional feature matrix; is the sum of all principal component eigenvalues representing the total amount of information of the original high-dimensional feature; and the cumulative contribution rate formula of the first k principal components is:

[0066]

[0067] Γ k is the cumulative contribution rate of the first k principal components, reflecting the total proportion of the first k principal components carrying the original high-dimensional feature information; and k is the number of principal components to be screened.

[0068] Effective principal component screening and dimension reduction dimension determination logic

[0069] The eigenvalue sequence obtained by decomposing the centralized kernel matrix is sorted in descending order;

[0070] The cumulative contribution rates Γ k of the first 1, the first 2, …, and the first k principal components are calculated in turn according to the above formula.

[0071] Setting cumulative contribution rate threshold Γ th ≥ 85%, find Γ k ≥ Γ th The minimum k value, which is the final dimensionality reduction target dimension;

[0072] Select the top k principal component corresponding eigenvector, construct low-dimensional projection matrix, map the original high-dimensional feature matrix to low-dimensional kernel space, and complete dimensionality reduction.

[0073] The micro-vibration signal is mapped to the same feature space step is:

[0074] A time-frequency domain joint feature extraction mechanism is adopted to extract 10-dimensional time domain features, and fast Fourier transform (FFT) and short-time Fourier transform (STFT) are performed on the micro-vibration signal to extract 10-dimensional frequency domain features.

[0075] The time domain features and frequency domain features are integrated to construct a high-dimensional feature matrix, which is optimized to a 12-16 dimensional effective feature matrix.

[0076] The number of neighbors and the target dimensionality reduction range are initialized, the local neighbor set of each sample is determined by neighbor search, and the optimal local reconstruction weight matrix is solved. The reconstruction weight relationship consistent with the high-dimensional space is maintained in the low-dimensional space. The final dimensionality reduction dimension is determined based on the screening criterion of cumulative contribution rate ≥ 85%, and the low-dimensional feature matrix is output.

[0077] The time domain feature extraction algorithm extracts 10-dimensional basic features representing the mechanical state from the time domain waveform of the micro-vibration signal. The statistical feature extraction algorithm is mainly used to calculate the mean, variance, peak value, peak factor, kurtosis, skewness, pulse factor, kurtosis factor, waveform factor, and mean frequency statistics to quantify the time domain distribution characteristics, fluctuation intensity, and pulse characteristics of the signal, and fully capture the core information of the signal in the time dimension.

[0078] For the frequency distribution characteristics of the micro-vibration signal, two types of Fourier transform algorithms are used to extract frequency domain features:

[0079] The fast Fourier transform (FFT) algorithm converts the time domain micro-vibration signal from the time domain to the frequency domain to obtain the frequency spectrum distribution of the signal, extracts the frequency spectrum peak value, main frequency, frequency spectrum barycenter, and harmonic amplitude ratio frequency domain features, and captures the fixed characteristic frequency corresponding to the device fault.

[0080] The short-time Fourier transform (STFT) algorithm processes the non-stationary micro-vibration signal by sliding time window, and then performs Fourier transform on each segment of the signal to obtain a time-frequency matrix. The 10-dimensional features of time-frequency entropy, time-frequency peak energy, and characteristic frequency duration are extracted to solve the limitation of FFT that cannot capture time-varying frequency characteristics.

[0081] The high-dimensional feature optimization algorithm is used to optimize the high-dimensional feature matrix with 20 dimensions in time domain and frequency domain into a 12-16 dimensional effective feature matrix, and the core adopts a feature screening / redundancy elimination algorithm: by calculating the variance contribution of the features, the linear correlation between the features or the feature importance ranking based on the fault correlation degree, the core fault information is retained while the feature dimension is compressed, and the subsequent dimension reduction calculation complexity is reduced;

[0082] The local linear embedding LLE dimension reduction algorithm takes the LLE algorithm as the core of the whole dimension reduction process, and realizes the dimension reduction target of retaining the local topological structure through multiple sub-algorithms. Specifically, it includes:

[0083] The neighbor searching algorithm: based on the initialized number of neighbors, K-nearest neighbor (KNN) or ball tree search method is used to find the nearest local neighbor set for each high-dimensional feature sample, to ensure that the subsequent reconstruction only depends on the local samples and avoids global nonlinear interference.

[0084] The local reconstruction weight solving algorithm: by constructing an optimization objective through the least square method, the local neighbor weight that can optimally reconstruct each sample in the high-dimensional space is solved. The core logic is to minimize the linear combination error of the sample and the neighbor, and to capture the local correlation between the samples.

[0085] The low-dimensional space preservation algorithm: within the preset dimension range of the dimension reduction target, the low-dimensional feature vector is constructed, and the local reconstruction weight of the sample in the low-dimensional space is ensured to be completely consistent with that in the high-dimensional space through the constraint condition, so as to completely retain the local topological structure of the micro-vibration signal features.

[0086] The cumulative contribution rate screening algorithm is consistent with the screening logic in KPCA dimension reduction, and is used to determine the final dimension reduction dimension: the information contribution proportion of each principal component in the LLE dimension reduction process is calculated, a threshold of cumulative contribution rate ≥ 85% is set, and the effective dimension that can carry the core local feature information is selected, which can guarantee the dimension reduction effect and avoid the loss of core fault features, and finally the low-dimensional feature matrix is output; the contribution rate ≥ 85% is determined based on the expert analysis method.

[0087] First, the multi-dimensional features of the micro-vibration signal are extracted through the time domain unified algorithm and the FFT / STFT frequency domain transformation algorithm; then the feature optimization algorithm is used to eliminate redundancy and simplify the feature dimension; taking the LLE algorithm as the core, combined with the neighbor searching and weight solving sub-algorithms, the local topological structure of the high-dimensional features is retained in the low-dimensional space, and finally the optimal dimension reduction dimension is determined through the cumulative contribution rate screening, which lays a foundation for the subsequent mapping of the temperature, pressure and flow parameter features into the same space.

[0088] The parameter features after dimension reduction are mapped into the same feature space through the kernel space metric learning method, based on the historical fault sample library of the whole life cycle of the heat pump, the adaptive mapping relationship of the two types of features is learned, and the KPCA low-dimensional features and the LLE low-dimensional features are projected into the same high-dimensional kernel space.

[0089] The life cycle attention branch performs encoding processing on the heat pump full life cycle stage, constructs a stage and feature correlation matrix based on a full life cycle historical fault sample library, calculates an initial weight of each parameter feature based on a preset attention weight network, and the working condition attention branch performs quantitative encoding processing on a real-time running working condition to generate a dynamic feature weight matrix corresponding to the real-time working condition.

[0090] The two types of branch outputs are weighted and summed based on a preset fusion coefficient a to perform collaborative fusion of the double-branch weights.

[0091] The collaborative fusion algorithm formula is:

[0092] W final = a * W L + (1-a) * W v

[0093] W final is the feature weight matrix after final collaborative fusion, with a dimension of "feature quantity x 1", which is directly used to adjust the feature weight after KPCA and LLE fusion to highlight the fault strongly related features; a is a preset fusion coefficient, with a value range of a ∈ [0, 1], which is used to balance the weight contribution degree of the two types of attention branches; it is determined by offline training and optimization of the full life cycle historical fault sample library; W L The initial weight matrix output by the life cycle attention branch has a dimension of "feature quantity x 1"; it is calculated by a preset attention weight network based on the full life cycle stage encoding and feature correlation matrix, and reflects the fault correlation importance of each feature under different life cycle stages; W v The dynamic feature weight matrix output by the working condition attention branch has a dimension of "feature quantity x 1"; it is generated based on the quantitative encoding result of the real-time running working condition, and reflects the sensitivity of each feature to fault identification under the current working condition;

[0094] The fusion coefficient a is a key adjustment parameter: when a is close to 1, more attention is paid to the feature weight of the life cycle stage; when a is close to 0, more attention is paid to the feature weight of the real-time working condition; the weighted sum operation is adopted to realize the collaboration of the double-branch weights: both the long-term correlation rule of the features and the faults under the full life cycle dimension and the dynamic change of the feature importance under the real-time working condition are adapted, and the finally output W final can dynamically and accurately adjust the feature weight to maximize the recognition degree of the fault strongly related features.

[0095] The analysis and early warning module performs a precise screening operation of the fault intensity correlation characteristics. The module takes the unified feature set generated after data fusion as an input basis, adopts a random forest feature contribution degree algorithm, and uses the technical characteristics of integrating multiple decision trees of the algorithm to realize precise evaluation of feature importance by quantifying the classification and differentiation ability of each feature for the two types of faults, i.e., compressor wear and pipeline leakage. The algorithm obtains the importance score of each feature through the two-dimensional weighted calculation of the Gini coefficient reduction and the out-of-bag data (OOB) error rate increase. After verification of the historical fault sample set throughout the life cycle, a core feature subset with an importance score not less than 0.7 is finally selected. The core feature subset has a strong fault recognition degree. For example, the frequency domain peak value feature of the micro-vibration signal corresponding to the compressor wear fault and the linear coupling feature of the pressure-flow parameter corresponding to the pipeline leakage fault provide high-recognition feature support for subsequent fault precise identification.

[0096] A dynamic mapping library of fault types and core features is constructed. Based on the core feature subset obtained through the above screening, different fault types are associated and mapped with corresponding high-sensitivity core features to form an initial mapping relationship. At the same time, in the process of system operation, with the continuous accumulation of new fault sample data and maintenance feedback data, the mapping library is dynamically updated and iterated to optimize the matching accuracy of features and fault types, ensuring that the mapping relationship can adapt to the dynamic evolution of the equipment throughout the life cycle, and providing a reliable technical basis for rapid matching of fault types.

[0097] On this basis, a dual-dimensional dynamic baseline of life cycle and working condition is built. The present scheme breaks through the technical bottleneck of traditional single fixed baseline, divides the baseline construction scene according to the combination dimension of life cycle stage and operating condition, and takes each combination dimension as an independent baseline construction unit. For each baseline construction unit, a normal feature baseline is constructed based on the historical normal running sample set using statistical modeling means. The baseline includes the mean value parameter and the fluctuation interval parameter of the core feature, ensuring that the baseline can accurately adapt to the normal running characteristics of the equipment under different life cycle stages and different operating conditions.

[0098] The dual judgment mechanism of instantaneous deviation calculation + cumulative deviation calculation realizes accurate discrimination. On the one hand, the instantaneous deviation value of the real-time core feature and the corresponding two-dimensional dynamic baseline is calculated based on the Euclidean distance algorithm, which directly quantifies the deviation degree of the real-time feature from the normal baseline. The 95% confidence interval is set as the preliminary abnormality judgment threshold. When the instantaneous deviation value exceeds the confidence interval, the preliminary abnormality alarm is triggered, which indicates that the equipment has potential failure risk, and the high-intensity maintenance process is not directly started to avoid excessive operation and maintenance problems caused by false positives. On the other hand, the sliding window mechanism is introduced to perform weighted cumulative operation on the instantaneous deviation index to obtain the cumulative deviation value. The weight configuration of the sliding window follows the principle of giving higher weight to recent deviations, which can effectively capture the trend of fault development and avoid accidental interference of single instantaneous deviation. Combined with the numerical range and trend of the cumulative deviation value, three fault evolution stages are divided: abnormal budding period, abnormal development period and abnormal severe period. The abnormal budding period corresponds to the cumulative deviation value being greater than or equal to 1.2 times and less than 1.8 times of the upper limit of the baseline fluctuation. At this time, only the instantaneous deviation occasionally exceeds the threshold without obvious continuous trend. The abnormal development period corresponds to the cumulative deviation value being greater than or equal to 1.8 times and less than 3.0 times. The instantaneous deviation frequently exceeds the threshold, and the deviation shows a continuous upward trend. The abnormal severe period corresponds to the cumulative deviation value being greater than or equal to 3.0 times, or the instantaneous deviation suddenly rising to more than 3 times of the upper limit of the baseline fluctuation. At this time, the deviation continues to expand, and the core feature is highly matched with the fault items in the dynamic mapping library.

[0099] Finally, different early warning and disposal schemes are output based on different fault evolution stages. For the abnormal budding period, a mild early warning instruction is output, and preventive measures such as parameter calibration and increasing real-time monitoring frequency are recommended to suppress further evolution of the fault. For the abnormal development period, a moderate early warning instruction is output, prompting the operation and maintenance personnel to perform shutdown detection during non-peak hours, focusing on the equipment components corresponding to the core features associated with the dynamic mapping library, and carrying out local repair or component maintenance. For the abnormal severe period, a severe early warning instruction is output, triggering an emergency shutdown instruction, and pushing an emergency maintenance scheme, which clearly specifies the fault component replacement list, repair process and safety operation specification, to minimize equipment damage and operation loss caused by fault expansion. The whole analysis and early warning process not only guarantees the accuracy of fault identification, but also realizes the adaptability of early warning and disposal schemes, providing scientific and efficient technical support for operation and maintenance decision-making.

[0100] The acquisition logic of the core features with importance score ≥ 0.7 in the present application first takes the fused uniform feature set as input, constructs a training set labeled with normal operation, compressor wear and pipeline leakage based on the full life cycle historical fault sample library, and trains a random forest model composed of multiple decision trees; The importance of the feature is calculated by two methods: one is the Gini coefficient reduction method, which quantifies the reduction of data impurity after feature splitting node; The second is the out-of-bag data OOB error rate increase method, which measures the increase of classification error after the feature is disturbed, and the average value of the two is taken as the initial score; The initial score is normalized to the interval [0, 1], and the fault recognition effect of different thresholds is verified through historical samples to determine that ≥ 0.7 is the screening standard. This threshold can accurately retain the exclusive distinguishing features of the two types of faults, balance the recognition accuracy and feature size, and finally screen out the core features that play a key role in distinguishing faults.

[0101] The standard for dividing the abnormal stage is:

[0102] When the cumulative deviation is ≥ 1.2 times and < 1.8 times the upper limit of the baseline fluctuation, only the instantaneous deviation occasionally exceeds the threshold, and there is no obvious sustained trend.

[0103] When the cumulative deviation is ≥ 1.8 times and < 3.0 times, the instantaneous deviation frequently exceeds the threshold, and the deviation shows a sustained upward trend.

[0104] When the cumulative deviation is ≥ 3.0 times or the instantaneous deviation suddenly rises to more than 3 times the upper limit of the baseline fluctuation, the deviation continues to expand, and there is a fault entry in the core feature matching mapping library.

[0105] Prior art cases

[0106] Case 1: Intelligent monitoring system of ground source heat pump in a large commercial complex

[0107] System architecture: Deploy temperature (16 points), pressure (8 points), flow (4 points) and vibration (4 points) sensor matrix, sampling frequency 100 Hz; Use PCA dimensionality reduction technology to process multi-dimensional data, build a fault detection model, and realize fault prediction based on LSTM neural network, which can warn 95% potential faults 72 hours in advance.

[0108] Case 2: Intelligent diagnosis system of air source heat pump group in a regional heating station

[0109] Core technology: Multi-sensor data spatio-temporal correlation analysis: Accurately distinguish normal start-up from locked-rotor failure by combining current surge with temperature rise rate; Fault grading warning establishes a three-level warning mechanism for graded response from minor anomaly to emergency fault; Fault tree diagnosis: Correlate alarm information with maintenance knowledge base to provide accurate maintenance guidance.

[0110] Limitations of existing technologies

[0111] Data fusion is not comprehensive: most systems simply superimpose sensor data, lacking deep correlation analysis between parameters; feature extraction is static: unable to adapt to different life cycle stages and operating conditions of heat pumps; baseline model is single: lacking dynamic baseline considering equipment aging and operating condition changes; early warning dimension is one-sided: relying on single parameter threshold judgment, easy to miss or misreport; maintenance plan lacks pertinence: unable to provide precise maintenance suggestions matching fault type and severity.

[0112] Technical advantages of multi-sensor fusion heat pump full life cycle AI maintenance and early warning system.

[0113] Data acquisition module, comprehensive perception and accurate capture;

[0114] Micro-vibration monitoring, for the first time, micro-vibration is included in the full life cycle monitoring of heat pumps, capturing early abnormal signals of mechanical parts;

[0115] Multi-parameter collaborative sampling, realizing accurate timestamp synchronization and spatial coordinate calibration of temperature, pressure, flow and vibration data;

[0116] Full life cycle data archiving, establishing complete operating parameter archives from equipment commissioning to decommissioning.

[0117] Advantages, compared with traditional systems, more than 30% of early mechanical failures can be detected, and the early warning time is 7-30 days earlier.

[0118] The above shows and describes the basic principles, main features and advantages of the present application. It should be understood by those skilled in the art that the present application is not limited by the above examples, and the above examples and descriptions in the specification are only the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the present application. The scope of protection required by the present application is defined by the appended claims and their equivalents.

Claims

1. A multi-sensor fusion heat pump full life cycle AI maintenance early warning system, characterized in that, The early warning system comprises: a data acquisition module, which acquires operating parameters of the heat pump in a whole life cycle based on a sensor array, and constructs a data set after preprocessing the parameters; the operating parameters include temperature, pressure, flow and micro-vibration; a data fusion module, which extracts trend correlation features and parameter coupling features from the temperature, pressure and flow parameters by using a dynamic sliding window adapted to the working condition, reduces dimensions by KPCA to retain core nonlinear information; the micro-vibration signal extraction includes frequency domain and time domain features, reduces dimensions by LLE to retain local topological structure, and maps the parameter features after dimension reduction to the same feature space; the parameter feature weight is dynamically adjusted by a life cycle and working condition double attention mechanism to highlight the strongly related features of faults; an analysis and early warning module, which screens the strongly related features of faults, establishes a mapping library of fault types and core features, constructs a dynamic baseline in the life cycle and working condition, calculates the instantaneous and cumulative deviations of real-time features and the baseline, and double-determines suspected abnormalities and development stages; and outputs analysis results according to the deviation degree; an abnormal early warning module, which displays early warning information, outputs maintenance schemes, and establishes a health record of the heat pump in a whole life cycle based on the analysis results.

2. The multi-sensor fusion heat pump full life cycle AI maintenance early warning system according to claim 1, characterized in that: The sensor array is deployed in the operating nodes of the compressor, heat exchanger, throttling device and pump group, the temperature is collected by a temperature sensor, the pressure is collected by a piezoelectric pressure sensor, the flow is collected by an electromagnetic flowmeter, and the micro-vibration signal is collected by a piezoelectric acceleration sensor; preprocessing is performed by using a wavelet transform algorithm for noise suppression to eliminate irrelevant noise caused by electromagnetic interference and mechanical interference; the data is completed by using an interpolation method based on time sequence correlation; and normalization is performed by using Z-score.

3. The multi-sensor fusion heat pump full life cycle AI maintenance early warning system according to claim 1, characterized in that, The dynamic sliding window adaptation rule for temperature, pressure and flow parameters is as follows: based on the load rate, start-stop signal and parameter fluctuation characteristics of the heat pump, the operating state is divided into start-stop transition, steady-state operation and variable load working condition, and the temperature, pressure and flow data are grouped according to the working condition; calculate indicators for each group of data, including parameter change rate average, parameter fluctuation intensity and working condition mutation identification, and characterize the working condition based on the indicators; wherein the dynamic sliding window selection adjustment is as follows: in the steady-state working condition, a large window of 80-120 sampling points and a large step of 40-60 sampling points are used; in the variable load working condition, a medium window of 40-80 sampling points and a medium step of 20-40 sampling points are used; and in the start-stop transition / mutation working condition, a small window of 20-40 sampling points and a small step of 5-10 sampling points are used.

4. The multi-sensor fusion heat pump full life cycle AI maintenance early warning system according to claim 1, characterized in that: Trend correlation features are extracted in the dynamic sliding window adapted to the working condition for temperature, pressure and flow parameters; the trend correlation features include single-parameter time evolution features and multi-parameter collaborative operation features; wherein the single-parameter time evolution features specifically include parameter change slope, trend stability index, extreme value feature and cumulative change amount, which depict the time evolution law of a single parameter; the multi-parameter collaborative operation features specifically include parameter change direction consistency coefficient, trend synchronization delay amount, trend intensity synergy degree and abnormal trend proportion, which represent the collaborative mechanism among multiple parameters; For all two two combinations of temperature, pressure and flow, a linear coupling feature extraction mechanism is constructed. In the dynamic sliding window of the corresponding working condition, the linear correlation coefficient of each two two parameter combination is calculated to quantitatively represent the linear correlation of the time series data. Combined with the consistency quantization result of parameter change direction, the linear coupling relationship between parameters is represented, and the linear correlation of two two parameter combinations is captured.

5. The multi-sensor fusion heat pump full life cycle AI maintenance early warning system according to claim 1, characterized in that: The KPCA dimensionality reduction step for retaining core nonlinear information is: The extracted trend correlation features and parameter coupling features are collected and integrated to construct a high-dimensional feature unified matrix. The local outlier factor (LOF) algorithm is used to identify and correct the abnormal feature values in the high-dimensional feature unified matrix. Select the radial basis function (RBF) kernel function as the core mapping function, which adapts to the nonlinear correlation characteristics of the parameters. Use the 5-fold cross-validation and grid search parameter optimization strategy to iteratively select the kernel parameters of the RBF kernel function and determine the optimal kernel parameters. Based on the optimal kernel parameters, the similarity between each sample in the high-dimensional feature unified matrix is calculated to construct a sample similarity kernel matrix. The sample similarity kernel matrix is centered and corrected to obtain a standardized and centered kernel matrix. Perform eigenvalue decomposition on the centered kernel matrix to obtain the eigenvalue sequence and the corresponding eigenvectors. Based on the eigenvalue sequence, calculate the cumulative contribution rate of each principal component, set the cumulative contribution rate threshold to be greater than or equal to 85%, and select the effective principal components that meet the threshold requirement to determine the target dimensionality reduction dimension. Based on the selected effective principal components, construct a low-dimensional projection matrix to map the high-dimensional features to a low-dimensional kernel space, and complete the dimensionality reduction of the high-dimensional features.

6. The multi-sensor fusion heat pump full life cycle AI maintenance early warning system according to claim 1, characterized in that, The micro-vibration signal mapping to the same feature space step is: Use the time-frequency domain joint feature extraction mechanism to extract 10-dimensional time domain features. Perform fast Fourier transform (FFT) and short-time Fourier transform (STFT) on the micro-vibration signal to extract 10-dimensional frequency domain features. Collect and integrate the time domain features and frequency domain features to construct a high-dimensional feature matrix. Optimize the high-dimensional feature matrix to a 12-16 dimensional effective feature matrix. Initialize the neighbor number and the target dimensionality reduction range, determine the local neighbor set of each sample through neighbor search, and solve the optimal local reconstruction weight matrix. Maintain the same reconstruction weight relationship in the low-dimensional space as in the high-dimensional space. Determine the final dimensionality reduction dimension based on the cumulative contribution rate greater than or equal to 85%, and output the low-dimensional feature matrix.

7. The multi-sensor fusion heat pump full life cycle AI maintenance early warning system according to claim 1, characterized in that: Map the reduced parameter features to the same feature space using kernel space metric learning. Based on the historical fault sample library of the whole life cycle of the heat pump, learn the adaptive mapping relationship between the two types of features, and project the KPCA low-dimensional features and LLE low-dimensional features to the same high-dimensional kernel space.

8. The multi-sensor fusion heat pump full life cycle AI maintenance early warning system according to claim 1, characterized in that: The life cycle attention branch encodes the heat pump whole life cycle stage, and based on the whole life cycle historical fault sample library, a stage and feature correlation matrix is constructed. Based on the preset attention weight network, the initial weights of each parameter feature are calculated. The working condition attention branch quantitatively encodes the real-time running working condition to generate a dynamic feature weight matrix corresponding to the real-time working condition. The two branch outputs are subjected to a weighted sum operation based on a preset fusion coefficient alpha to perform a synergistic fusion of the weights of the two branches.

9. The multi-sensor fusion heat pump full life cycle AI maintenance early warning system according to claim 1, characterized in that: The analysis and early warning module analyzes and early warns by screening the fault intensity related features, quantifies the feature contribution degree algorithm based on the random forest, quantifies the feature contribution degree algorithm based on the random forest, and quantifies the feature contribution degree algorithm based on the random forest. The importance score of the core feature subset is greater than or equal to 0.

7. A dynamic mapping library of fault types and core features is established, a double-dimensional dynamic baseline of life cycle and working condition is constructed, a core feature subset with an importance score greater than or equal to 0.7 is selected, and a double-dimensional dynamic baseline of life cycle and working condition is constructed. Based on historical normal samples, the normal feature baseline of each dimension is constructed by a statistical modeling method, including mean and fluctuation range parameters; the instantaneous deviation of real-time features and baseline is calculated by Euclidean distance, and the preliminary abnormality is determined when the deviation exceeds 95% confidence interval; the cumulative deviation is obtained by weighted sum of instantaneous deviation index through sliding window, and different stages of abnormal budding period, development period and severe period are divided; different early warning measures are given based on different stages.

10. The multi-sensor fusion heat pump full life cycle AI maintenance early warning system according to claim 9, characterized in that, The standards for dividing abnormal stages are as follows: When the cumulative deviation is greater than or equal to 1.2 times and less than 1.8 times of the upper limit of the baseline fluctuation, only the instantaneous deviation occasionally exceeds the threshold value, and there is no obvious sustained trend; When the cumulative deviation is greater than or equal to 1.8 times and less than 3.0 times, the instantaneous deviation frequently exceeds the threshold value, and the deviation shows a sustained upward trend; When the cumulative deviation is greater than or equal to 3.0 times or the instantaneous deviation suddenly rises to more than 3 times of the upper limit of the baseline fluctuation, the deviation continuously expands and the fault entry in the core feature matching mapping library.

Images