Solar energy self-charging public intelligent water taking machine intelligent control method

Abstract

The application provides a solar self-charging public intelligent water taking machine intelligent control method, relates to the technical field of water taking machines, and comprises the following steps: determining Mahalanobis distance and azimuth based on historical benchmark distribution and the shape point, determining a first threshold value, a second threshold value and a third threshold value based on the historical benchmark distribution, and determining an inaccurate area in combination with the azimuth; generating corresponding severity according to the inaccurate area to which the shape point belongs, regulating the internal adjustable parameters based on the severity, repeatedly executing new parameters until a final prediction value is obtained and the process is ended when a residual sequence reaches a convergence state or a maximum iteration number is reached. The method can not only avoid model oscillation divergence caused by excessive correction, but also ensure that the method can quickly converge to a high-precision state in the early stage of failure, and finally realizes high prediction accuracy and system high reliability of the solar water taking machine in a complex outdoor environment.

Description

Technical Field

[0001] This application relates to the field of water dispenser technology, and in particular to a smart control method for a solar-powered self-charging public smart water dispenser. Background Technology

[0002] Solar-powered self-charging smart water dispensers, as public facilities integrating environmental protection, energy conservation, and intelligent services, are widely used in public places such as parks, squares, communities, and transportation hubs. Their core function is to use solar photovoltaic panels to convert light energy into electrical energy, which is then stored in batteries to drive a water pump system to provide clean drinking water or hot water to the public. However, because these devices are exposed to complex and variable outdoor environments (such as high temperature, high humidity, strong ultraviolet radiation, and sandstorms) for extended periods and operate continuously without human intervention, they are highly susceptible to various factors such as mechanical wear, electrical aging, sensor drift, and water pollution. This can lead to abnormal fluctuations in pump operating current, changes in vibration spectrum, and even serious malfunctions such as pump seizure or idling. If potential malfunctions cannot be identified and addressed in a timely manner, the water service will face the risk of interruption, directly harming the user experience and potentially causing serious accidents such as pump burnout, short circuits, and even fires, resulting in significant economic losses and safety hazards. To maintain the stable operation of water dispensers, existing technologies mostly employ data-driven fault prediction models for real-time monitoring. The typical process is as follows: First, collect the current and vibration time series data of the equipment under historical normal operating conditions as a training set, build a prediction model using neural networks or regression algorithms, and set the weight matrix and bias vector as adjustable parameters; during the equipment operation phase, input the real-time collected current and vibration data into the model, calculate the error (i.e., loss function) between the output value and the actual measured value, and then iteratively update the model parameters through the backpropagation algorithm, striving to minimize the error until convergence or reaching the preset iteration limit, and finally determine the equipment status based on the predicted value.

[0003] However, this traditional prediction method, which relies on static training or simple iterative correction, has revealed significant shortcomings in practical applications, making it difficult to meet the required prediction accuracy. On the one hand, traditional methods often adhere to fixed parameters obtained from offline training, lacking the ability to adapt to concept drift caused by component aging and environmental temperature and humidity fluctuations during long-term operation. That is, the physical characteristics of the equipment have undergone substantial changes, while the model parameters remain in the initial state, causing the prediction benchmark to fail. On the other hand, existing error correction mechanisms usually adopt a one-size-fits-all global iterative strategy, failing to finely classify and distinguish residual characteristics, making it difficult to differentiate between occasional sensor noise interference and real early fault symptoms, and easily falling into the dilemma of overfitting noise or underreporting real hidden dangers. In addition, when the model prediction error increases significantly, the lack of targeted adjustment methods often leads to parameter oscillation or even divergence, which further weakens the prediction accuracy.

[0004] The aforementioned lack of accuracy in prediction directly leads to delayed or false alarms in fault warnings. On the one hand, low accuracy may cause the system to fail to detect subtle abnormal trends in the early stages of a fault, missing the optimal maintenance window and causing minor faults to escalate into major accidents. On the other hand, frequent false alarms can trigger unnecessary downtime for maintenance, reducing the online rate and service efficiency of water intake machines and increasing operation and maintenance costs. Summary of the Invention

[0005] This application aims to at least partially address one of the technical problems in the related art.

[0006] To achieve the above objectives, this application proposes a method for intelligent control of a solar-powered self-charging public smart water dispenser, comprising the following steps: Step 1: Collect real-time current time-series data and real-time vibration time-series data at the current operating moment of the water pump, and perform preprocessing to obtain real-time current time-series preprocessed data and real-time vibration time-series preprocessed data; Step 2: Based on several continuous sampling points of the current running time and the previous time, the corresponding real-time current time series preprocessed data and real-time vibration time series preprocessed data are input into the fault prediction model to obtain the corresponding predicted current value and predicted vibration value; wherein, the fault prediction model includes a weight matrix and a bias vector as internal adjustable parameters. Step 3: Calculate the first difference between the real-time current time series preprocessed data and the predicted current value at each continuous sampling point, and calculate the second difference between the real-time vibration time series preprocessed data and the predicted vibration value at each continuous sampling point; all the first and second differences constitute a residual sequence; within a preset sliding time window, map the residual sequence to a set of trajectory points on a two-dimensional plane to construct a two-dimensional error trajectory; Step 4: Perform principal component analysis on the two-dimensional error trajectory to obtain the slope of the principal direction of the trajectory. Determine the geometric center based on the two-dimensional error trajectory. Calculate the average Euclidean distance from all trajectory points in the two-dimensional error trajectory to the geometric center, denoted as the trajectory distribution density index. Construct a two-dimensional space based on the horizontal axis and the vertical axis, using the slope of the principal direction of the trajectory as the horizontal axis and the trajectory distribution density index as the vertical axis. Construct data pairs based on the slope of the principal direction of the trajectory and the trajectory distribution density index, and use the data pairs in the two-dimensional space as morphological points. Step 5: Determine the Mahalanobis distance and azimuth angle based on the historical baseline distribution and the morphological point; determine the first threshold, second threshold, and third threshold based on the historical baseline distribution and combine them with the azimuth angle to determine the inaccurate region; generate the corresponding severity level according to the inaccurate region to which the morphological point belongs; Step 6: Adjust the internal adjustable parameters based on the severity, generate new parameters, and repeat Step 1 until the residual sequence reaches a convergent state or the maximum number of iterations is reached to obtain the final predicted value and end the process.

[0007] Further, the Mahalanobis distance and azimuth are determined based on the historical baseline distribution and the morphological point; a first threshold, a second threshold, and a third threshold are determined based on the historical baseline distribution, and the inaccurate region is determined in combination with the azimuth; the severity is generated according to the inaccurate region to which the morphological point belongs, including the following steps: Step 51: Obtain historical residual data, and calculate the historical baseline distribution mean vector and covariance matrix based on the historical residual data; wherein, the historical baseline distribution mean is composed of the arithmetic mean of the slope of the main direction of the trajectory of the historical residual data and the arithmetic mean of the trajectory distribution density index, and the position of the arithmetic mean of the slope of the main direction of the trajectory of the historical residual data and the arithmetic mean of the trajectory distribution density index in two-dimensional space is used as the determination origin; Step 52: Determine the Mahalanobis distance between the shape point and the determination origin; based on the shape point, the arithmetic mean of the slope of the main direction of the trajectory of the historical residual data, and the arithmetic mean of the trajectory distribution density index, determine the azimuth angle of the shape point relative to the determination origin; Step 53: Determine the weighted fluctuation radius based on the covariance matrix, and generate a first moving radius, a second moving radius, and a third moving radius based on the weighted fluctuation radius; using the determination origin as the starting point and the first moving radius, the second moving radius, and the third moving radius as the radius boundaries, divide the two-dimensional space into four concentric determination regions: accurate region, first inaccurate region, second inaccurate region, and third inaccurate region; Step 54: Based on the accurate region, the first inaccurate region, the second inaccurate region, and the third inaccurate region, determine the target region where the morphological point is located, determine the quantitative index of the morphological point within the target region, and determine the fault severity level based on the quantitative index.

[0008] Further, the weighted fluctuation radius is determined based on the covariance matrix, and a first moving radius, a second moving radius, and a third moving radius are generated based on the weighted fluctuation radius. Taking the determination origin as the starting point and the first moving radius, the second moving radius, and the third moving radius as the radius boundaries, the two-dimensional space is divided into four concentric determination regions: an accurate region, a first inaccurate region, a second inaccurate region, and a third inaccurate region. This includes the following steps: Step 531: Perform eigenvalue decomposition based on the covariance matrix to obtain the first eigenvalue and the second eigenvalue; construct the anisotropy ratio based on the first eigenvalue and the second eigenvalue. Step 532: Preset reference standard deviation, and generate standard fluctuation components based on reference standard deviation, first eigenvalue and second eigenvalue; Step 533: Obtain the Mahalanobis distance of the most recent m historical pattern points, and determine the mean of the Mahalanobis distances of the m historical pattern points as the current moving average; generate the relative fluctuation ratio based on the current moving average and the standard fluctuation component; Step 534: Based on the standard fluctuation component, the current moving average, and the relative fluctuation ratio, generate a first moving radius, a second moving radius, and a third moving radius; the first moving radius < the second moving radius < the third moving radius; Step 535: Using the determination origin as the center, and based on the position of the first moving radius along the radial direction of the center as the first radius boundary, the area formed by the first radius boundary with the determination origin is defined as the accurate region. Step 536: Using the determination origin as the center, and based on the position of the second moving radius moved radially along the center as the second radius boundary, the area enclosed by the second radius boundary and the first radius boundary is defined as the first inaccurate area; Step 537: Using the determination origin as the center, and based on the position of the third moving radius along the radial direction of the center as the third radius boundary, the area enclosed by the third radius boundary and the second radius boundary is defined as the second inaccurate area; Step 538: Define the region formed outside the third radius boundary as the third inaccurate region.

[0009] Further, based on the accurate region, the first inaccurate region, the second inaccurate region, and the third inaccurate region, the target region where the morphological point is located is determined, the quantitative index of the morphological point within the target region is determined, and the fault severity level is determined based on the quantitative index, including the following steps: Step 541: Determine the Euclidean distance between the shape point and the determination origin; obtain the radius of the target region, and obtain a quantitative index based on the Euclidean distance and the radius; Step 542: Obtain the comprehensive severity based on the quantitative indicators and Mahalanobis distance; Step 543: If the overall severity is less than the first threshold, the severity is determined to be the first severity; if the overall severity is greater than or equal to the first threshold and less than the second threshold, the severity is determined to be the second severity; if the overall severity is greater than or equal to the second threshold, the severity is determined to be the third severity.

[0010] Furthermore, the third level of severity is the highest, the first level is the lowest, and the second level is moderate.

[0011] Further, based on the severity, the internal adjustable parameters are adjusted to generate new parameters. Step 1 is repeated until the residual sequence reaches convergence or the maximum number of iterations is reached to obtain the final predicted value and the process ends. This includes the following steps: Step 61: When the severity is the first severity, update the bias vector to obtain the updated bias vector. Repeat step 1 based on the updated bias vector until the residual sequence reaches the convergence state or the maximum number of iterations is reached to obtain the final predicted value. Step 62: When the severity is the second severity, update the weight vector to obtain the updated weight vector. Repeat step 1 based on the updated weight vector until the residual sequence reaches the convergence state or the maximum number of iterations is reached to obtain the final predicted value. Step 63: When the severity level is the third most severe, execute an emergency security decision.

[0012] Further, updating the bias vector to obtain an updated bias vector includes: determining a unit direction vector based on the determination origin and shape point; and obtaining an updated bias vector based on the quantization index, the unit direction vector, and the bias vector.

[0013] Further, updating the weight vector to obtain the updated weight vector includes: constructing a decay coefficient based on the quantization index, obtaining a baseline weight vector, and obtaining the updated weight vector based on the baseline weight vector, the weight vector, and the decay coefficient.

[0014] Furthermore, the first moving radius < the second moving radius < the third moving radius.

[0015] Compared with existing technologies, the intelligent control method for a solar-powered self-charging public smart water dispenser provided in this application effectively solves the problems of single feature dimension, difficulty in distinguishing noise interference from real faults, and accuracy degradation caused by rigid model parameter update strategies in traditional fault prediction methods by constructing a two-dimensional error trajectory based on residual sequences and its morphological point analysis. By mapping the one-dimensional current and vibration residuals into a two-dimensional space composed of the slope of the main trajectory direction and the trajectory distribution density index, this scheme not only quantifies the magnitude of the error but also captures the dynamic trend of the error evolution over time (i.e., directionality). This enables it to keenly identify early fault symptoms with small numerical fluctuations but specific abnormal directions, significantly improving the robustness of feature extraction. Building upon this foundation, multiple dynamic movement radii generated from historical data are used to divide the two-dimensional space into refined judgment regions. This approach, based on Mahalanobis distance and azimuth angles, cleverly combines statistical anisotropy with the sensitivity of physical space. It adaptively adjusts the judgment boundary according to fluctuations in the current environment: automatically widening the threshold to avoid false alarms when environmental noise is high, and automatically tightening the threshold to improve sensitivity when abnormal trends occur in a specific direction. This completely overcomes the shortcomings of fixed thresholds, which cannot adapt to equipment aging and environmental drift. It abandons the traditional, coarse-grained, globally unified iterative approach, instead employing precise strategies based on the specific region where the shape point falls: fine-tuning the bias vector to eliminate systematic drift in cases of slight deviation, adjusting the weight vector to reconstruct the feature mapping relationship in cases of moderate anomalies, and directly triggering safety decisions in cases of severe faults. This progressive, hierarchical parameter self-optimization strategy avoids model oscillations and divergences caused by over-correction, and ensures rapid convergence to a high-precision state in the early stages of a fault. Ultimately, it achieves high prediction accuracy and high system reliability for the solar water pump during long-term operation in complex outdoor environments. Attached Figure Description

[0016] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart illustrating a method for intelligent control of a solar-powered self-charging public smart water dispenser provided in this application embodiment; Figure 2 This is a structural block diagram of a solar-powered self-charging public intelligent water dispenser control system provided in an embodiment of this application; Figure 3 This is a block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0017] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0018] The following describes a method for intelligent control of a solar-powered self-charging public smart water dispenser according to an embodiment of this application, with reference to the accompanying drawings.

[0019] like Figure 1 As shown, a method for intelligent control of a solar-powered self-charging public smart water dispenser includes the following steps: Step 1: Collect real-time current time-series data and real-time vibration time-series data at the current operating moment of the water pump, and perform preprocessing to obtain real-time current time-series preprocessed data and real-time vibration time-series preprocessed data.

[0020] First, at the physical sensing layer, a high-precision Hall effect current sensor (sampling frequency set to 2kHz to cover high-frequency harmonics) is connected in series in the three-phase power supply circuit of the water pump motor. High-sensitivity piezoelectric acceleration vibration sensors (sampling frequency synchronously set to 2kHz) are installed at key resonance points on the pump body bearing housing and motor housing. The raw analog signal is then transmitted to the edge computing gateway in real time via a wireless transmission module. Second, in the data access and cleaning stage, a sliding window algorithm is used to remove outliers from continuously acquired time-series data. The Laida criterion is used to identify and correct spike noise caused by electromagnetic interference or momentary sensor contact problems. Simultaneously, wavelet threshold denoising technology is combined to effectively filter out power frequency interference, environmental wind vibration, and other noise. Random white noise generated by the shutdown action is removed, preserving characteristic components that reflect the true operating state of the equipment. Subsequently, data normalization and alignment are performed. The Z-Score normalization method is used to eliminate the influence of dimensions on the current and vibration sequences, making their mean 0 and variance 1 respectively. A timestamp alignment mechanism is used to ensure that the current data and vibration data are strictly synchronized at the microsecond level, solving the data misalignment problem caused by sensor response delay. Finally, feature enhancement preprocessing is performed. An overlapping sliding window is constructed on the cleaned long time series data to transform the original one-dimensional time series data into a two-dimensional matrix form containing information on local trends, periodic fluctuations, and transient changes. The output is real-time current time series preprocessed data and real-time vibration time series preprocessed data.

[0021] Step 2: Based on several consecutive sampling points of the current running time and the previous time, the corresponding real-time current time series preprocessed data and real-time vibration time series preprocessed data are input into the fault prediction model to obtain the corresponding predicted current value and predicted vibration value; wherein, the fault prediction model includes a weight matrix and a bias vector as internal adjustable parameters.

[0022] In this embodiment, the fault prediction model employs a Long Short-Term Memory (LSTM) network. Its core architecture includes an input layer, a bidirectional LSTM hidden layer, and a fully connected output layer. The internal adjustable parameters specifically refer to the gated weight matrix and bias vector within the LSTM unit, as well as the linear transformation weight matrix and bias vector of the output layer. During model training, a training set is first constructed using massive amounts of current and vibration time-series data collected before the equipment leaves the factory and under historical normal operating conditions. A continuous sampling point sequence (e.g., a time step of length 64) is used as input, and the corresponding actual measurement value at the next moment is used as the label. The training process uses mean squared error (MSE) as the loss function. The weight matrix and bias vector are iteratively updated using a backpropagation algorithm combined with the Adam optimizer. An L2 regularization term is introduced to prevent overfitting, and a dynamic learning rate decay strategy is set until the loss function on the training set converges or reaches the preset maximum number of iterations, thereby obtaining the initialized fault prediction model. In the real-time prediction phase, real-time current time-series preprocessed data and real-time vibration time-series preprocessed data of the current operating moment and several consecutive sampling points preceding it (such as the most recent 32 time steps) are acquired in real time. These data are concatenated into a multi-dimensional feature tensor and input into the trained fault prediction model. The data flows sequentially through the input layer for mapping and the LSTM hidden layer for extraction of time-dependent features and state memory updates (with effective information filtered through a gating mechanism). Finally, the fully connected layer outputs the predicted current value and predicted vibration value at the corresponding moment. This prediction process is streaming, meaning that for each new sampling point acquired, the model slides the window forward one step, generating an accurate prediction of the equipment's operating state in real time.

[0023] Step 3: Calculate the first difference between the real-time current time series preprocessed data and the predicted current value at each continuous sampling point, and calculate the second difference between the real-time vibration time series preprocessed data and the predicted vibration value at each continuous sampling point; all the first and second differences constitute a residual sequence; within a preset sliding time window, map the residual sequence to a set of trajectory points on a two-dimensional plane to construct a two-dimensional error trajectory.

[0024] In this embodiment, the residual sequence construction process is first performed: each continuous sampling point within the current sliding window is acquired in real time, and the real-time current time series preprocessed data at that moment is subtracted point by point from the predicted current value output by the fault prediction model to obtain the first difference sequence, which reflects the instantaneous deviation of the current amplitude; at the same time, the real-time vibration time series preprocessed data at the same moment is subtracted point by point from the predicted vibration value to obtain the second difference sequence, which characterizes the abnormal fluctuation of the vibration signal in the time domain; then, the first difference of the corresponding time step in the two sequences is used as the coordinate component and arranged in chronological order to form a two-dimensional vector set containing N (N is greater than 5) points, which is the residual sequence. To capture the dynamic evolution of errors rather than simply their numerical magnitude, a geometric mapping is performed on the residual sequence within a preset sliding time window (e.g., the most recent 64 sampling points): each point in the aforementioned two-dimensional vector set is projected onto a two-dimensional Cartesian plane with the first difference as the horizontal axis (X-axis) and the second difference as the vertical axis (Y-axis), forming a set of discrete trajectory points. By connecting these trajectory points arranged in chronological order, a two-dimensional error trajectory is constructed. This trajectory visually demonstrates the movement path, distribution pattern, and evolution trend of current and vibration errors in the composite space. If the equipment is operating normally, the trajectory points will be tightly clustered near the origin, forming a compact circular or elliptical cloud. If a specific type of fault occurs (e.g., bearing wear leading to increased vibration but with a lag in current change), the trajectory points will exhibit obvious elongation, deflection, or divergence, thus providing an intuitive geometric basis for subsequently extracting the slope and distribution density indicators of the main trajectory direction.

[0025] Step 4: Perform principal component analysis on the two-dimensional error trajectory to obtain the slope of the principal direction of the trajectory. Determine the geometric center based on the two-dimensional error trajectory. Calculate the average Euclidean distance from all trajectory points in the two-dimensional error trajectory to the geometric center, denoted as the trajectory distribution density index. Construct a two-dimensional space based on the horizontal axis and the vertical axis, using the slope of the principal direction of the trajectory as the horizontal axis and the trajectory distribution density index as the vertical axis. Construct data pairs based on the slope of the principal direction of the trajectory and the trajectory distribution density index, and use the data pairs in the two-dimensional space as morphological points.

[0026] First, rigorous statistical calculations are performed on the two-dimensional error trajectory point set within the preset sliding time window to obtain the covariance matrix: All sampling points within the sliding time window are traversed, and the arithmetic mean of the current residual data and the arithmetic mean of the vibration residual data are calculated respectively to determine the geometric center coordinates of the trajectory; for each specific sampling point, the following quantization operations are performed to obtain the three core elements constituting the standard covariance matrix of the two-dimensional error trajectory point set: First, the calculation method for current fluctuation (i.e., the variance of current residual) is as follows: first, calculate the difference between the current value at the point and the average current value, then square the difference, then sum the squared differences of all sampling points and divide by the number of samples minus one. The result is the current fluctuation, which reflects the degree of dispersion of the current error around the mean.

[0027] Second, the method for calculating the vibration fluctuation amount (i.e. the variance of the vibration residual) is as follows: first, calculate the difference between the vibration value at the point and the average vibration value, then square the difference, then sum the squared differences of all sampling points and divide by the sample size minus one. The result is the vibration fluctuation amount, which reflects the degree of dispersion of the vibration error around the mean.

[0028] Third, the calculation method of the linkage change law quantity (i.e. the covariance of current and vibration) is as follows: For each sampling point, the difference between its current residual and the average value is multiplied by the difference between the vibration residual and the average value at that point to obtain a pair of "product terms". Then, these product terms of all sampling points are summed and divided by the number of samples minus one. The result is the linkage change law quantity, which characterizes the correlation between current error and vibration error, whether they increase synchronously, decrease synchronously, or cancel each other out.

[0029] Finally, the specific values ​​obtained from the above three calculations are filled into a 2×2 symmetric matrix: the current fluctuation is filled into the upper left corner, the vibration fluctuation is filled into the lower right corner, and the linkage change law is filled into both the upper right and lower left corners (to quantify the linear correlation and coupling direction between the two), thus completely constructing the standard covariance matrix describing the two-dimensional error trajectory point set.

[0030] Eigenvalue decomposition is performed on the standard covariance matrix of the two-dimensional error trajectory point set. The characteristic equation of the matrix is ​​solved to obtain two eigenvalues ​​and their corresponding eigenvectors. The unit eigenvector corresponding to the eigenvalue with the largest value is selected as the first principal component vector. The slope of this vector is further calculated as the trajectory principal direction slope (the ratio of the longitudinal component to the lateral component in the first principal component vector), thereby accurately capturing the principal axis direction of the error distribution. At the same time, the geometric center of the trajectory is directly locked based on the arithmetic mean calculated above. All N trajectory points are retraced, and the distance from each point to the center is calculated using the Euclidean distance formula. Finally, the arithmetic mean of these distances is calculated and normalized to serve as the trajectory distribution density index, which characterizes the dispersion range of the error in the vertical direction of the principal axis. Finally, the calculated slope and density index are combined into a data pair and mapped to a two-dimensional feature space with the trajectory principal direction slope as the horizontal axis and the trajectory distribution density index as the vertical axis, forming a unique morphological point. This point completely encodes the error correlation pattern and fluctuation intensity of the equipment in the current operating state, providing a highly robust geometric feature input for subsequent fault diagnosis.

[0031] Step 5: Determine the Mahalanobis distance and azimuth angle based on the historical baseline distribution and the morphological point; determine the first threshold, second threshold, and third threshold based on the historical baseline distribution and combine them with the azimuth angle to determine the inaccurate region; generate the corresponding severity level according to the inaccurate region to which the morphological point belongs.

[0032] Step 51: Obtain historical residual data, and calculate the historical baseline distribution mean vector and covariance matrix based on the historical residual data; wherein, the historical baseline distribution mean is composed of the arithmetic mean of the trajectory principal direction slope and the arithmetic mean of the trajectory distribution density index of the historical residual data, and the position of the arithmetic mean of the trajectory principal direction slope and the arithmetic mean of the trajectory distribution density index of the historical residual data in two-dimensional space is used as the determination origin.

[0033] In this embodiment, firstly, multiple sets of residual data samples under stable and normal equipment conditions are screened and extracted from historical operation records. For each set of historical residual data, the aforementioned statistical analysis process is performed: calculating the geometric center within its sliding window, constructing the covariance matrix, performing eigenvalue decomposition to obtain the first principal component vector, and deriving the trajectory principal direction slope. Simultaneously, the average distance from each point to the center is calculated as the trajectory distribution density index. Subsequently, the "trajectory principal direction slope" values ​​corresponding to all collected historical samples are summed and their arithmetic mean is calculated. Similarly, the trajectory distribution density index values ​​are summed and their arithmetic mean is calculated. These two statistical means together constitute a two-dimensional coordinate point, which is the historical baseline distribution mean vector, representing the equipment's healthy state. The typical morphological characteristics of the error distribution are obtained (the acquisition process is the same as in step 4, and will not be described in detail in this embodiment); then, the position of the mean vector is located in the preset two-dimensional feature space and defined as the judgment origin. This origin not only includes the center value (slope mean) of the correlation trend between current and vibration error, but also includes the center value (density mean) of the dispersion of error fluctuation; based on this judgment origin, the covariance matrix calculated using historical data is further used to determine the confidence ellipse boundary or Gaussian distribution range under normal conditions, thereby establishing a dynamic fault diagnosis benchmark model. When the new morphological point generated by the real-time collected error data deviates significantly from the judgment origin or exceeds the preset statistical threshold, it can be determined that the equipment operating status has abnormally drifted.

[0034] Step 52: Determine the Mahalanobis distance between the morphological point and the determination origin; based on the morphological point, the arithmetic mean of the slope of the main trajectory direction of the historical residual data, and the arithmetic mean of the trajectory distribution density index, determine the azimuth angle of the morphological point relative to the determination origin.

[0035] In this embodiment, the Mahalanobis distance is first quantitatively calculated to assess the statistical deviation of the morphological point from the normal baseline: the real-time generated morphological point coordinates (composed of the current trajectory principal direction slope and trajectory distribution density index) are regarded as a two-dimensional vector, and the previously determined judgment origin coordinates (i.e., the historical baseline distribution mean vector) are subtracted from it to obtain the deviation vector; the deviation vector is then weighted using the inverse matrix of the covariance matrix constructed under the historical normal operating conditions. Specifically, the deviation vector is first multiplied on the left by the inverse matrix of the covariance matrix, then multiplied on the right by the transpose of the deviation vector, and finally the square root of the quadratic form is taken to obtain the Mahalanobis distance that accurately reflects the correlation of multidimensional data distribution (the method of calculating the Mahalanobis distance based on the vector and the deviation matrix is ​​a conventional technique, which will not be described in detail in this embodiment). This distance value eliminates the dimensional difference between the current residual and the vibration residual. Anomalies and related interferences can objectively measure the degree of abnormality of the current state relative to the healthy state. The relative azimuth of the morphological point is determined based on geometric relationships: by calculating the line connecting the coordinates of the morphological point and the coordinates of the judgment origin in the two-dimensional feature space, the arctangent function (arctan2) is used to solve the angle between the line and the horizontal axis (i.e., the slope axis of the main trajectory direction). The increment of the horizontal axis is defined as the difference between the slope of the main trajectory direction of the current morphological point and the historical average slope, and the increment of the vertical axis is defined as the difference between the trajectory distribution density index of the current morphological point and the historical average density index. The azimuth obtained from this not only indicates the specific direction of the equipment state deviation (e.g., whether the slope increases or the density diverges), but also implies the trend characteristics of fault evolution. By combining the magnitude of the Mahalanobis distance with the direction of the azimuth, the accurate location of the equipment fault type and the severity classification can be achieved.

[0036] Step 53: Determine the weighted fluctuation radius based on the covariance matrix, and generate a first moving radius, a second moving radius, and a third moving radius based on the weighted fluctuation radius; using the determination origin as the starting point and the first moving radius, the second moving radius, and the third moving radius as the radius boundary, divide the two-dimensional space into four concentric determination regions: accurate region, first inaccurate region, second inaccurate region, and third inaccurate region.

[0037] Step 531: Perform eigenvalue decomposition based on the covariance matrix to obtain the first eigenvalue and the second eigenvalue; construct the anisotropy ratio based on the first eigenvalue and the second eigenvalue.

[0038] A rigorous eigenvalue decomposition operation is performed on the covariance matrix to solve its characteristic equation, thereby calculating the first and second eigenvalues, both of which are greater than 0. The first eigenvalue corresponds to the major axis of the data distribution ellipse, representing the direction of the most significant principal component of signal fluctuation (i.e., the main trajectory direction), and its magnitude reflects the energy intensity or dispersion amplitude in that direction. The second eigenvalue corresponds to the minor axis of the data distribution ellipse, representing the secondary fluctuation direction (i.e., the secondary trajectory direction), and its magnitude reflects the degree of fluctuation constraint perpendicular to the principal direction. Based on these two eigenvalues, an anisotropy ratio is further constructed, i.e., the first eigenvalue divided by the second eigenvalue, describing the geometric flattening or elongation of the data distribution. When the anisotropy ratio is close to 1, it indicates that the data points are approximately circularly distributed in space, meaning that fluctuations in all directions are uniform and isotropic. However, when the anisotropy ratio is much greater than 1, it indicates that the data points are highly concentrated on a thin, elongated elliptical trajectory, exhibiting strong anisotropy, where fluctuations in the principal direction dominate the dynamic characteristics of the entire system.

[0039] Step 532: Preset reference standard deviation, and generate standard fluctuation components based on the reference standard deviation, the first eigenvalue, and the second eigenvalue.

[0040] First, a preset reference standard deviation is determined by statistical characteristics under historical normal conditions. The average of the square roots of the eigenvalues ​​of the covariance matrix of a large number of sample data collected during the long-term stable operation of the equipment is selected, representing the inherent noise level or baseline fluctuation amplitude of the specific equipment under ideal health conditions. Then, using the first and second eigenvalues ​​calculated in the previous steps and the preset reference standard deviation, a normalization operation is performed to generate a standard fluctuation component. The specific calculation process is to divide the arithmetic mean of the two eigenvalues ​​by the square of the reference standard deviation, and then take the square root of the resulting ratio. This value intuitively reflects how many times the overall dispersion of the data distribution at the current moment is relative to the historical baseline fluctuation.

[0041] Step 533: Obtain the Mahalanobis distance of the most recent m historical pattern points, and determine the mean of the Mahalanobis distances of the m historical pattern points as the current moving average; generate the relative fluctuation ratio based on the current moving average and the standard fluctuation component.

[0042] A sliding time window of length m (m ≥ 3) is maintained. The m nearest historical morphological points within this window are continuously collected. Using the constructed covariance matrix and its inverse matrix, the Mahalanobis distance of each morphological point relative to the distribution center is calculated one by one. Then, the arithmetic mean of these m Mahalanobis distances is calculated to obtain the current moving average, which intuitively represents the actual dispersion of the data stream in the statistical space at the current moment. If the current moving average increases, it indicates that the fluctuation range of the current data point is expanding. Next, using the standard fluctuation component generated in the previous steps as a baseline, the ratio of the current moving average to the standard fluctuation component is calculated to generate the relative fluctuation ratio. This quantifies the deviation factor of the currently observed overall fluctuation level from the expected fluctuation level under the historical baseline: when the relative fluctuation ratio is 1... When the relative fluctuation ratio is greater than 1, it indicates that the current environment is in a normal and stable state, and there is no need to make significant adjustments to the threshold. When the relative fluctuation ratio is greater than 1, it means that the current data distribution is more dispersed than the historical baseline. Based on this, the radius of the dynamically generated threshold will be automatically increased to prevent false alarms caused by the general increase in environmental noise. Conversely, when the relative fluctuation ratio is less than 1, it indicates that the data distribution is more concentrated, and the threshold can be appropriately tightened to improve the sensitivity to minor anomalies.

[0043] Step 534: Based on the standard fluctuation component, the current moving average, and the relative fluctuation ratio, generate a first moving radius, a second moving radius, and a third moving radius; the first moving radius < the second moving radius < the third moving radius.

[0044] First, calculate a comprehensive reference value, which is composed of the product of the first weight coefficient and the current moving average, plus the product of the second weight coefficient and the standard fluctuation component, so as to not only retain the real-time fluctuation characteristics of the data at the current moment but also take into account the stability of the historical reference. The first weight coefficient is 0.4 and the second weight coefficient is 0.6. Subsequently, multiply this comprehensive reference value by the relative fluctuation ratio to eliminate the interference of environmental noise and obtain the basic scaling radius. On this basis, introduce three preset non-negative hierarchical expansion coefficients k1, k2, k3 (satisfying 0 < k1 < k2 < k3), and linearly amplify the basic scaling radius respectively, that is, calculate three dynamic threshold radii through the formula (comprehensive reference value × relative fluctuation ratio) × (1 + expansion coefficient). Among them, the first moving radius corresponds to the smallest expansion coefficient k1, which delimits the boundary of the safe area for equipment operation, indicating that the data points within this range belong to the category of normal fluctuations; the second moving radius corresponds to the medium expansion coefficient k2, which delimits the boundary of the early warning area. When the data points exceed the first moving radius but are within the second moving radius, an early warning signal is triggered to prompt potential risks; the third moving radius corresponds to the largest expansion coefficient k3, which delimits the boundary of the fault / danger area. Once the distance of the data points exceeds the third moving radius, it is determined as a serious abnormality and an emergency shutdown or alarm mechanism is triggered. These three radii together constitute a concentric circular shape space that adapts and expands according to the working conditions, ensuring that regardless of how the current environmental noise level changes, different risk-level determination areas can be accurately delimited based on the real-time statistical distribution characteristics.

[0045] The core purpose of the above processing flow in the solution of this application is to construct a dynamic and adaptive fault diagnosis space to overcome the defect that the traditional fixed threshold method cannot adapt to equipment aging and environmental drift. The construction of this space cleverly uses the coordinate transformation system of physical space - standardized space - physical space.

[0046] First, in the physical space (i.e., the two-dimensional feature space composed of the slope of the main direction of the trajectory and the trajectory distribution density index), the distribution of historical normal data usually presents as an ellipse. This ellipse is defined by the historical reference covariance matrix, and its shape (major axis, minor axis, direction) precisely depicts the correlation between the data in the two dimensions and their respective fluctuation intensities (i.e., anisotropy). This is a real distribution form that objectively reflects the statistical characteristics of the true operating state of the equipment.

[0047] However, directly setting the judgment boundary in this physical space presents a problem of dimension inconsistency. Since slope and density are two completely different physical quantities, their numerical ranges and fluctuation scales can differ significantly, causing one dimension to dominate the result during distance calculation, thus masking anomalies in the other dimension. To address this issue, this scheme introduces an intermediate step: standardization space. By using the standard deviations of the horizontal and vertical axes as dynamic normalization scaling factors, the ellipse in the physical space is stretched or compressed, transforming it into a dimensionless, isotropic circular space. In this standardized space, a unit distance in any direction represents an equal degree of statistical deviation, thereby achieving fair and accurate distance measurement and threshold setting.

[0048] Finally, to ensure that the final judgment result can be intuitively mapped back to the physical world that engineers can understand and to conform to the actual distribution of data, the circular judgment boundaries (i.e., the first, second, and third moving radii) set in the standardized space are stretched back to the original physical space through inverse coordinate transformation, forming the final judgment region boundary. This resulting boundary inherits the advantages of isotropy and fair measurement in the standardized space, while perfectly conforming to the actual distribution of data (ellipse) in the physical space, thus achieving high-precision and robust fault diagnosis.

[0049] In summary, this solution successfully constructs an intelligent diagnostic framework that can both eliminate the influence of dimensions and adapt to the data distribution pattern through the transformation process of ellipse (physical) - circle (standard) - ellipse (physical). This is the key technology for achieving high accuracy and high reliability in this invention.

[0050] Step 535: Using the determination origin as the center, and based on the position of the first moving radius along the radial direction of the center as the first radius boundary, the area formed by the first radius boundary with the determination origin is defined as the accurate region.

[0051] Step 536: Using the determination origin as the center, and based on the position of the second moving radius moved radially along the center as the second radius boundary, the area enclosed by the second radius boundary and the first radius boundary is defined as the first inaccurate area; Step 537: Using the determination origin as the center, and based on the position of the third moving radius along the radial direction of the center as the third radius boundary, the area enclosed by the third radius boundary and the second radius boundary is defined as the second inaccurate area; Step 538: Define the region formed outside the third radius boundary as the third inaccurate region.

[0052] In this embodiment, in order to map the first, second, and third moving radii to a two-dimensional space with the trajectory principal direction slope (horizontal axis) and trajectory distribution density index (vertical axis) as coordinate axes, statistical preprocessing is performed to construct feature scale factors for each dimension, and then the final geometric boundaries are established through a coordinate transformation mechanism: the first radius boundary, the second radius boundary, and the third radius boundary.

[0053] First, calculate the standard deviation of the horizontal and vertical axes in two-dimensional space relative to the origin: Based on historically collected normal operating condition sample data, extract the set of trajectory principal direction slope values ​​and the set of trajectory distribution density indices for each moment. If the origin is not a zero coordinate, perform a decentering translation to return it to the origin before calculating the standard deviation. Calculate the standard deviation of these two sets relative to the zero coordinate using the standard deviation formula, thus obtaining the standard deviation of the horizontal and vertical axes, i.e., the standard deviation of the trajectory principal direction slope and the standard deviation of the trajectory distribution density index.

[0054] Then, the standard deviation of the slope of the principal direction of the trajectory and the standard deviation of the trajectory distribution density index are used as dynamic normalized scaling factors for the horizontal and vertical axes. These two factors represent the physical magnitude corresponding to a one-unit standard fluctuation in the slope and density indices under the current equipment characteristics, and they act as a bridge connecting the abstract statistical space and the concrete physical space.

[0055] Finally, a coordinate transformation is performed based on the three moving radii to generate the radius boundary in two-dimensional space: the first, second, and third moving radii are considered as geometric radii in standardized Euclidean space, where the origin is determined to be at (0,0), and the unit length of the horizontal and vertical axes has been unified to a standard deviation of their respective dimensions. At this point, a circular trajectory with radii equal to the first, second, and third moving radii is drawn within the standardized space, and a basic circle equation is constructed for this trajectory. Several angle points are uniformly selected along the circumference, and the coordinates in the standardized space are calculated by combining the basic circle equation with these angle points. Using a dynamic normalization scaling factor, the coordinate points in the standardized space are linearly stretched back to the original physical coordinate system. For any point on the normalized circular boundary, its actual physical coordinates on the original two-dimensional plane (slope-density plane) are determined by the following relationship: the physical coordinate on the horizontal axis equals the normalized horizontal coordinate multiplied by the dynamic normalization scaling factor of the horizontal axis; the physical coordinate on the vertical axis equals the normalized vertical coordinate multiplied by the dynamic normalization scaling factor of the vertical axis, thus obtaining the physical coordinates of the horizontal and vertical axes in two-dimensional space, which is the true position of the point in the original physical coordinate system (slope-density plane). Connecting all the transformed points in sequence forms the final first radius boundary, second radius boundary, and third radius boundary.

[0056] This embodiment defines a sensitive angle range, which is a pre-defined azimuth angle range most sensitive to measurement errors, based on equipment operating characteristics or environmental noise distribution. Specifically, it is defined as a fan-shaped area extending 15° to 30° to both sides of the main trajectory direction (i.e., the positive horizontal axis direction). or Within this angular range, due to sensor sensitivity or geometric projection effects, minute positional deviations can easily be amplified into significant slope or density anomalies. Therefore, the system adopts stricter judgment logic for data points within this region; while outside this range, it is considered a relatively stable observation area.

[0057] First accurate region: The closed region containing the origin is defined as the first accurate region, with the first radius boundary as its outer edge. All shape points falling into this region whose Mahalanobis distance is less than or equal to the first movement radius are considered to be in a completely normal operating state.

[0058] The first inaccurate region: The annular region between the second radius boundary and the first radius boundary is defined as the first inaccurate region. This region encompasses all morphological points whose Mahalanobis distance is greater than the first moving radius and less than or equal to the second moving radius, regardless of their azimuth angle; these are considered to be in a state of mild anomaly or require attention.

[0059] The second inaccurate region: The area enclosed by the third radius boundary and the third radius boundary is classified as the second inaccurate region. A point is classified into this region only if its Mahalanobis distance is greater than the second movement radius and less than or equal to the third movement radius, and its azimuth angle does not belong to the preset sensitive angle range; if the point is within this distance range but is located within the sensitive angle range, it is not included in this category, but enters the next level of judgment.

[0060] Regarding the definition of the third inaccurate area (i.e., the highest risk warning zone), this area consists of two specific high-risk points: the first type is morphological points whose Mahalanobis distance is between the second and third moving radii (the Mahalanobis distance of the morphological point is greater than the second moving radius and less than or equal to the third moving radius), and whose orientation falls within a preset sensitive angle range; the second type is all morphological points whose Mahalanobis distance exceeds the third moving radius (the Mahalanobis distance of the morphological point is greater than the second moving radius), i.e., the entire area located outside the boundary of the third radius. These two types together define the system's warning red line, meaning that once a moderate deviation occurs in the sensitive direction, or a serious deviation occurs in any direction, the highest level of risk assessment will be immediately triggered.

[0061] To address the dynamic characteristics of equipment at different operating stages, this solution completely abandons the static judgment mode that relies on fixed thresholds. Instead, it constructs an adaptive modeling mechanism based on real-time standard fluctuation components. During startup, given the large data dispersion and significant noise caused by thermodynamic effects or mechanical inertia, the algorithm automatically expands the judgment radius by calculating the standard deviation in real time, thereby effectively filtering out normal startup fluctuations and preventing false alarms. After entering the steady-state period, as the data distribution tends to converge and the standard deviation decreases, the judgment radius automatically shrinks, thus significantly improving the sensitivity to capturing minute abnormal signals. During the wear and tear period, in the face of data mean drift and changes in fluctuation characteristics, the system uses a linkage algorithm of moving average and relative fluctuation ratio for real-time tracking, ensuring that the center and shape of the judgment area always closely follow the actual data distribution evolution, achieving accurate adaptation to the fault characteristics throughout the entire life cycle.

[0062] This solution innovatively introduces a direction-weighted risk perception approach in its composite strategy that integrates sensitive angle ranges and multi-radius regions. Traditional single-radius determination methods often rely solely on the Euclidean distance of data points relative to the origin, neglecting the significant differences between different directions in the physical scene. This solution, however, pre-defines specific sensitive angle ranges for key orientations (such as the sensor's sensitive axis or vulnerable areas in geometric projection) and applies more stringent discrimination logic to samples falling within these ranges. Specifically, once a data point falls within the sensitive angle range, even if its spatial distance has not yet reached the conventional threshold, it will be warned in advance and classified as a high-risk inaccurate region, thus achieving proactive detection of potential hazards. Conversely, in non-critical directions, the system exhibits a wider tolerance for deviations. Based on a three-layer progressive radius boundary architecture of accurate region, first / second inaccurate region, and third inaccurate region, a warning network with significant anisotropic characteristics is constructed: this network can both keenly lock onto subtle anomalies in key orientations (thus significantly improving the detection rate) and effectively shield against drastic fluctuations caused by environmental interference in non-sensitive directions (significantly reducing the false alarm rate). This dual-judgment mechanism, which deeply integrates distance thresholds and directional constraints, successfully achieves precise separation of danger signals from environmental noise. While ensuring high sensitivity, it significantly optimizes the accuracy of judgment and its practical engineering value.

[0063] Step 54: Based on the accurate region, the first inaccurate region, the second inaccurate region, and the third inaccurate region, determine the target region where the morphological point is located, determine the quantitative index of the morphological point within the target region, and determine the fault severity level based on the quantitative index.

[0064] The accurate region / first inaccurate region / second inaccurate region / third inaccurate region where the shape point is located is taken as the target region.

[0065] Step 541: Determine the Euclidean distance between the morphological point and the determination origin; obtain the radius of the target region, and obtain a quantitative index based on the Euclidean distance and the radius.

[0066] The connection distance between the shape point and the determination origin is calculated using the Euclidean distance formula. The radius of the target area where the shape point is located is used as the radius of this embodiment. A quantitative index is obtained based on the ratio of the Euclidean distance to the radius.

[0067] Step 542: Obtain the comprehensive severity based on the quantitative indicators and Mahalanobis distance.

[0068] Normalization is performed based on Mahalanobis distance and quantification index to obtain normalized distance and normalized index. The comprehensive severity is obtained by weighted summation of normalized distance and normalized index, where the weight factors of normalized distance and normalized index are between 0 and 1.

[0069] Step 543: If the overall severity is less than a first threshold, the severity is determined to be the first severity level; if the overall severity is greater than or equal to the first threshold and less than a second threshold, the severity is determined to be the second severity level; if the overall severity is greater than or equal to the second threshold, the severity is determined to be the third severity level. The third severity level is the highest, the first severity level is the lowest, and the second severity level is moderate.

[0070] In practice, two key thresholds are first set: a first threshold and a second threshold. The first threshold must be less than the second threshold, and both the first and second thresholds must be greater than 0. When the calculated overall severity score is less than the first threshold, it is classified as the first severity level (low risk / normal state), indicating that the equipment is operating within an ideal or permissible range of safety fluctuations, and no intervention is required. When the overall severity score is greater than or equal to the first threshold and less than the second threshold, it is classified as the second severity level (medium risk / early warning state), indicating that the equipment has experienced a slight abnormality or trend deviation. The system automatically triggers a high-level alarm and recommends increased inspections. When the overall severity score is greater than or equal to the second threshold, it is directly classified as the third severity level (high risk / failure state), indicating that the equipment has suffered substantial damage or serious deviation, and the system immediately initiates an emergency shutdown or maintenance procedure.

[0071] By locating morphological points to specific target areas and calculating their quantitative indicators within those areas, a leap from absolute distance judgment to relative position assessment is achieved. This means that regardless of the equipment's load or environmental noise level, the severity level of a fault can be standardized based on its relative deviation within the current safety boundary, eliminating interference from differences in dimensions and benchmarks. Finally, the fault severity level determined based on this quantitative indicator not only provides a simple binary or ternary judgment of normal, warning, and fault, but also gives maintenance personnel a clear basis for decision-making: it can accurately identify which are minor deviations on the edge of safety that do not require intervention, which are early signs approaching the danger boundary, and which are emergency situations that have already caused substantial damage. This guides the priority allocation of resources to high-risk areas, significantly reducing unplanned downtime and improving the management efficiency and safety of the equipment throughout its entire lifecycle.

[0072] Step 6: Adjust the internal adjustable parameters based on the severity, generate new parameters, and repeat Step 1 until the residual sequence reaches a convergent state or the maximum number of iterations is reached to obtain the final predicted value and end the process.

[0073] Step 61: When the severity is the first severity, update the bias vector to obtain the updated bias vector. Repeat step 1 based on the updated bias vector until the residual sequence reaches the convergence state or the maximum number of iterations is reached to obtain the final predicted value. The process of updating the bias vector to obtain an updated bias vector includes: determining a unit direction vector based on the determination origin and shape point; and obtaining an updated bias vector based on the quantization index, the unit direction vector, and the bias vector.

[0074] First, using the current origin as a reference, a unit direction vector is calculated between the two, combined with the current position of the morphological point. This vector accurately depicts the deviation trend and geometric direction of the data point in the feature space. Then, the quantitative index (reflecting the degree of deviation) obtained from the previous calculation is multiplied by the parameters of each dimension of the unit direction vector to generate a correction amount for the current bias vector, thus obtaining the updated bias vector. This process essentially fine-tunes the static boundary according to the dynamic trend of real-time data, allowing the origin to adaptively move with the drift of data distribution, thereby eliminating systematic deviations caused by changes in operating conditions. Next, the updated bias vector replaces the original bias vector, and step 1 is re-executed to obtain a new residual sequence. The process of determining direction, calculating correction, updating bias, and recalculating residuals is continuously looped. The changes in the residual sequence are monitored in real time. Once the residual sequence reaches the preset convergence state (i.e., the residual fluctuation is less than a certain tolerance, indicating that the model has been fitted to the optimal state) or reaches the set maximum number of iterations (to prevent infinite loops), the iteration is terminated and the final predicted value is output. Even when the equipment is in the first severity stage with slight fluctuations, the prediction model can be continuously calibrated to ensure that the predicted values ​​always closely follow the actual operating trajectory, which not only improves the accuracy of short-term predictions.

[0075] Step 62: When the severity is the second severity, update the weight vector to obtain the updated weight vector. Repeat step 1 based on the updated weight vector until the residual sequence reaches the convergence state or the maximum number of iterations is reached to obtain the final predicted value. The process of updating the weight vector to obtain the updated weight vector includes: constructing a decay coefficient based on the quantization index, obtaining a baseline weight vector, and obtaining the updated weight vector based on the baseline weight vector, the weight vector, and the decay coefficient.

[0076] In this embodiment, based on the current quantization index, the attenuation coefficient is directly constructed by substituting it into the formula 1 / (1+quantization index). When the quantization index value is small (indicating a slight deviation), the denominator is slightly greater than 1, and the attenuation coefficient is close to 1 but less than 1, allowing the weight vector to be moderately adjusted to track small changes. When the quantization index value is large (indicating a serious deviation or drastic fluctuations), the denominator increases significantly, causing the attenuation coefficient to rapidly approach 0, thereby greatly suppressing the impact of current abnormal data on the weight vector and preventing the model from oscillating or diverging due to overfitting sudden noise. Subsequently, a preset benchmark weight vector (representing the ideal parameter distribution of the device under standard conditions) is obtained. The weight vector of the current iteration is fused with the benchmark weight vector, and the calculated attenuation coefficient is used as the adjustment factor of the weight vector. 1-attenuation coefficient is used as the adjustment factor of the benchmark weight vector. The attenuation coefficient is multiplied by each term of the weight vector to obtain the first vector. 1-attenuation coefficient is multiplied by each term of the benchmark weight vector to obtain the second vector. The elements at the same position of the first and second vectors are added together and averaged to obtain the updated weight vector. Next, the updated weight vector is replaced with the new weight vector, and step 1 is repeated to calculate the new residual sequence. The process of building the attenuation coefficient, fusing the baseline, updating the weights, and recalculating the residual is continuously looped. The changes in the residual sequence are monitored in real time. Once the residual reaches the convergence state (i.e., the fluctuation is less than a certain threshold) or the maximum number of iterations is reached, the iteration is terminated and the final prediction value is output.

[0077] Step 63: When the severity level is the third most severe, execute an emergency security decision.

[0078] Once the monitoring data characteristics are identified and categorized by the algorithm as the third severity level (e.g., key parameters instantly exceed safety thresholds, residual sequence divergence, or signs of irreversible physical damage), the system immediately interrupts the current step 1 to convergence judgment loop, ceasing any weight vector updates or residual iteration calculations based on quantitative indicators, as even a slight computational delay at this point could lead to catastrophic consequences. Instead, it directly invokes the pre-set highest-priority safety policy library, automatically matching the corresponding emergency response actions based on the fault type (e.g., overheating, high pressure, severe vibration, etc.), including but not limited to: immediately executing emergency equipment shutdown commands, cutting off power source input, activating the backup cooling system, and isolating the fault area. It also sends the highest-level audible and visual alarm signals to the central control room and on-site personnel; simultaneously, it forcibly locks all non-core control channels to prevent human error or automated logic conflicts from interfering with the emergency response process, and writes the current fault waveform, environmental parameters, and decision logs into a read-only safety black box in real time for subsequent accident tracing and analysis; the core of this embodiment is to establish the principle of safety first and speed priority, sacrificing the precision of the predictive model in exchange for the immediacy and reliability of decision execution, ensuring that when facing extreme risks of the third severity level, the system can execute the most conservative and effective protective measures with a millisecond-level response speed, thereby maximizing the protection of personnel safety and the integrity of equipment assets.

[0079] This application uses a process of perception, analysis, decision-making, and execution as the core of intelligent control to directly control the hardware of the water dispenser.

[0080] This scheme first captures the complete signal details. Then, it calculates the residuals between the model's predicted values ​​and the actual measured values, mapping the current and vibration residuals into a two-dimensional error trajectory within a sliding time window. This step not only preserves the magnitude information of the error but, more importantly, captures the dynamic correlation pattern (i.e., the trajectory's shape) between the two physical quantities over time. Next, principal component analysis is performed on this trajectory to extract the slope of the principal direction (reflecting the dominant trend of error change) and the trajectory distribution density index (reflecting the degree of error dispersion), constructing a novel, highly condensed morphological point. This morphological point, as a comprehensive feature vector, has a discriminative power far exceeding that of the original current or vibration values ​​themselves.

[0081] Ultimately, this solution places this morphological point within an adaptive diagnostic space dynamically generated from historical data, encompassing multiple movement radii, for evaluation. This space automatically adjusts the judgment threshold based on the current operating status of the equipment (such as ambient noise levels and component aging), and assigns different risk weights to anomalies in different directions by incorporating azimuth information. This refined hierarchical diagnostic mechanism enables the system to keenly identify early, subtle fault symptoms while effectively suppressing interference from occasional noise.

[0082] Therefore, although the input signals are only current and vibration, through the synergistic effect of the above-mentioned high-dimensional feature construction and dynamic adaptive criteria, this application is fully capable of accurately sensing the complex working conditions of water pumps, reliably predicting faults, and intelligently controlling model parameters, thereby achieving the expected technical effects and possessing good engineering feasibility.

[0083] Experimental data show that in the second severity (medium risk) scenario, the prediction accuracy of this embodiment is improved by an average of about 5% to 8% compared with the prior art. This improvement effectively eliminates the minor false alarms caused by overfitting noise, while significantly reducing the risk of missed alarms, making the system more robust and reliable under complex operating conditions.

[0084] like Figure 2 As shown, this embodiment also discloses a solar-powered self-charging public smart water dispenser control system, including the following modules: Acquisition module: used to acquire real-time current time-series data and real-time vibration time-series data of the water pump at the current operating moment, and to preprocess the data to obtain real-time current time-series preprocessed data and real-time vibration time-series preprocessed data; Prediction module: Based on several consecutive sampling points of the current running time and the previous time, it inputs the corresponding real-time current time series preprocessed data and real-time vibration time series preprocessed data into the fault prediction model to obtain the corresponding predicted current value and predicted vibration value. Data fitting module: used to calculate the first difference between the real-time current time series preprocessed data and the predicted current value at each continuous sampling point, and to calculate the second difference between the real-time vibration time series preprocessed data and the predicted vibration value at each continuous sampling point; all the first and second differences constitute a residual sequence; within a preset sliding time window, the residual sequence is mapped to a set of trajectory points on a two-dimensional plane to construct a two-dimensional error trajectory; Data processing module: used to perform principal component analysis on the two-dimensional error trajectory to obtain the slope of the principal direction of the trajectory, determine the geometric center based on the two-dimensional error trajectory, calculate the average Euclidean distance from all trajectory points in the two-dimensional error trajectory to the geometric center, denoted as the trajectory distribution density index, construct a two-dimensional space based on the slope of the principal direction of the trajectory and the trajectory distribution density index, construct data pairs based on the slope of the principal direction of the trajectory and the trajectory distribution density index, and use the data pairs in the two-dimensional space as morphological points; Data analysis module: used to determine Mahalanobis distance and azimuth angle based on historical baseline distribution and the morphological point; to determine a first threshold, a second threshold, and a third threshold based on historical baseline distribution and to determine inaccurate regions in combination with the azimuth angle; and to generate the corresponding severity level according to the inaccurate region to which the morphological point belongs. Intelligent control module: used to adjust the internal adjustable parameters based on the severity, generate new parameters and repeat step 1 until the residual sequence reaches the convergence state or the maximum number of iterations is reached to obtain the final predicted value and end.

[0085] To implement the above embodiments, this application also proposes an electronic device. Please see [link to relevant documentation]. Figure 3 , Figure 3 This is a schematic diagram of the structure of the electronic device provided in an embodiment of this application. For example... Figure 3 As shown, the electronic device 500 includes: a processor 501 and a memory 502 communicatively connected to the processor 501; the memory 502 stores computer-executable instructions; the processor 501 executes the computer-executable instructions stored in the memory to implement the method provided in the foregoing embodiments.

[0086] To implement the above embodiments, this application also proposes a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the methods provided in the foregoing embodiments.

[0087] To implement the above embodiments, this application also proposes a computer program product, including a computer program that, when executed by a processor, implements the methods provided in the foregoing embodiments.

[0088] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A method for intelligent control of a solar-powered self-charging public smart water dispenser, characterized in that, Includes the following steps: Step 1: Collect real-time current time-series data and real-time vibration time-series data at the current operating moment of the water pump, and perform preprocessing to obtain real-time current time-series preprocessed data and real-time vibration time-series preprocessed data; Step 2: Based on several continuous sampling points of the current running time and the previous time, the corresponding real-time current time series preprocessed data and real-time vibration time series preprocessed data are input into the fault prediction model to obtain the corresponding predicted current value and predicted vibration value; wherein, the fault prediction model includes a weight matrix and a bias vector as internal adjustable parameters. Step 3: Calculate the first difference between the real-time current time series preprocessed data and the predicted current value at each continuous sampling point, and calculate the second difference between the real-time vibration time series preprocessed data and the predicted vibration value at each continuous sampling point; all the first and second differences constitute a residual sequence; within a preset sliding time window, map the residual sequence to a set of trajectory points on a two-dimensional plane to construct a two-dimensional error trajectory; Step 4: Perform principal component analysis on the two-dimensional error trajectory to obtain the slope of the principal direction of the trajectory. Determine the geometric center based on the two-dimensional error trajectory. Calculate the average Euclidean distance from all trajectory points in the two-dimensional error trajectory to the geometric center, denoted as the trajectory distribution density index. Construct a two-dimensional space based on the horizontal axis and the vertical axis, using the slope of the principal direction of the trajectory as the horizontal axis and the trajectory distribution density index as the vertical axis. Construct data pairs based on the slope of the principal direction of the trajectory and the trajectory distribution density index, and use the data pairs in the two-dimensional space as morphological points. Step 5: Determine the Mahalanobis distance and azimuth angle based on the historical baseline distribution and the morphological point; determine the first threshold, second threshold, and third threshold based on the historical baseline distribution and combine them with the azimuth angle to determine the inaccurate region; generate the corresponding severity level according to the inaccurate region to which the morphological point belongs; Step 6: Adjust the internal adjustable parameters based on the severity, generate new parameters, and repeat Step 1 until the residual sequence reaches a convergent state or the maximum number of iterations is reached to obtain the final predicted value and end the process.

2. The intelligent control method for a solar-powered self-charging public intelligent water dispenser according to claim 1, characterized in that, The Mahalanobis distance and azimuth are determined based on the historical baseline distribution and the morphological points. The first threshold, the second threshold and the third threshold are determined based on the historical baseline distribution and the inaccurate area is determined in combination with the azimuth. The severity level is generated based on the inaccurate region to which the morphological point belongs, including the following steps: Step 51: Obtain historical residual data, and calculate the historical baseline distribution mean vector and covariance matrix based on the historical residual data; wherein, the historical baseline distribution mean is composed of the arithmetic mean of the slope of the main direction of the trajectory of the historical residual data and the arithmetic mean of the trajectory distribution density index, and the position of the arithmetic mean of the slope of the main direction of the trajectory of the historical residual data and the arithmetic mean of the trajectory distribution density index in two-dimensional space is used as the determination origin; Step 52: Determine the Mahalanobis distance between the shape point and the determination origin; based on the shape point, the arithmetic mean of the slope of the main direction of the trajectory of the historical residual data, and the arithmetic mean of the trajectory distribution density index, determine the azimuth angle of the shape point relative to the determination origin; Step 53: Determine the weighted fluctuation radius based on the covariance matrix, and generate a first moving radius, a second moving radius, and a third moving radius based on the weighted fluctuation radius; using the determination origin as the starting point and the first moving radius, the second moving radius, and the third moving radius as the radius boundaries, divide the two-dimensional space into four concentric determination regions: accurate region, first inaccurate region, second inaccurate region, and third inaccurate region; Step 54: Based on the accurate region, the first inaccurate region, the second inaccurate region, and the third inaccurate region, determine the target region where the morphological point is located, determine the quantitative index of the morphological point within the target region, and determine the fault severity level based on the quantitative index.

3. The intelligent control method for a solar-powered self-charging public intelligent water dispenser according to claim 2, characterized in that, The weighted fluctuation radius is determined based on the covariance matrix, and a first moving radius, a second moving radius, and a third moving radius are generated based on the weighted fluctuation radius. Taking the origin as the starting point and the first, second, and third moving radii as the radius boundaries, the two-dimensional space is divided into four concentric determination regions: an accurate region, a first inaccurate region, a second inaccurate region, and a third inaccurate region. This includes the following steps: Step 531: Perform eigenvalue decomposition based on the covariance matrix to obtain the first eigenvalue and the second eigenvalue; construct the anisotropy ratio based on the first eigenvalue and the second eigenvalue. Step 532: Preset reference standard deviation, and generate standard fluctuation components based on the reference standard deviation, the first eigenvalue, and the second eigenvalue; Step 533: Obtain the Mahalanobis distance of the most recent m historical pattern points, and determine the mean of the Mahalanobis distances of the m historical pattern points as the current moving average; generate the relative fluctuation ratio based on the current moving average and the standard fluctuation component; Step 534: Based on the standard fluctuation component, the current moving average, and the relative fluctuation ratio, generate a first moving radius, a second moving radius, and a third moving radius; the first moving radius < the second moving radius < the third moving radius; Step 535: Using the determination origin as the center, and based on the position of the first moving radius along the radial direction of the center as the first radius boundary, the area formed by the first radius boundary with the determination origin is defined as the accurate region. Step 536: Using the determination origin as the center, and based on the position of the second moving radius moved radially along the center as the second radius boundary, the area enclosed by the second radius boundary and the first radius boundary is defined as the first inaccurate area; Step 537: Using the determination origin as the center, and based on the position of the third moving radius along the radial direction of the center as the third radius boundary, the area enclosed by the third radius boundary and the second radius boundary is defined as the second inaccurate area; Step 538: Define the region formed outside the third radius boundary as the third inaccurate region.

4. The intelligent control method for a solar-powered self-charging public intelligent water dispenser according to claim 3, characterized in that, Based on the accurate region, the first inaccurate region, the second inaccurate region, and the third inaccurate region, the target region where the morphological point is located is determined, the quantitative index of the morphological point within the target region is determined, and the fault severity level is determined based on the quantitative index, including the following steps: Step 541: Determine the Euclidean distance between the shape point and the determination origin; obtain the radius of the target region, and obtain a quantitative index based on the Euclidean distance and the radius; Step 542: Obtain the comprehensive severity based on the quantitative indicators and Mahalanobis distance; Step 543: If the overall severity is less than the first threshold, the severity is determined to be the first severity; if the overall severity is greater than or equal to the first threshold and less than the second threshold, the severity is determined to be the second severity; if the overall severity is greater than or equal to the second threshold, the severity is determined to be the third severity.

5. The intelligent control method for a solar-powered self-charging public intelligent water dispenser according to claim 4, characterized in that, The third level of severity is the highest, the first level is the lowest, and the second level is moderate.

6. The intelligent control method for a solar-powered self-charging public intelligent water dispenser according to claim 5, characterized in that, Based on the severity, the internal adjustable parameters are adjusted, new parameters are generated, and step 1 is repeated until the residual sequence reaches a convergent state or the maximum number of iterations is reached to obtain the final predicted value and end the process. This includes the following steps: Step 61: When the severity is the first severity, update the bias vector to obtain the updated bias vector. Repeat step 1 based on the updated bias vector until the residual sequence reaches the convergence state or the maximum number of iterations is reached to obtain the final predicted value. Step 62: When the severity is the second severity, update the weight vector to obtain the updated weight vector. Repeat step 1 based on the updated weight vector until the residual sequence reaches the convergence state or the maximum number of iterations is reached to obtain the final predicted value. Step 63: When the severity level is the third most severe, execute an emergency security decision.

7. The intelligent control method for a solar-powered self-charging public intelligent water dispenser according to claim 6, characterized in that, The process of updating the bias vector to obtain an updated bias vector includes: determining a unit direction vector based on the determination origin and shape point; and obtaining an updated bias vector based on the quantization index, the unit direction vector, and the bias vector.

8. The intelligent control method for a solar-powered self-charging public intelligent water dispenser according to claim 6, characterized in that, The process of updating the weight vector to obtain the updated weight vector includes: constructing a decay coefficient based on the quantization index, obtaining a baseline weight vector, and obtaining the updated weight vector based on the baseline weight vector, the weight vector, and the decay coefficient.

9. The intelligent control method for a solar-powered self-charging public intelligent water dispenser according to claim 3, characterized in that, First moving radius < Second moving radius < Third moving radius.

Images