A smart park dynamic analysis and early warning method and system

Abstract

This invention belongs to the field of data processing technology, specifically relating to a dynamic analysis and early warning method and system for smart parks. The method includes: collecting equipment health status and operating condition data; training a support vector regression model to establish a dynamic baseline response surface and obtaining residual statistical feature vectors; establishing a directed spatiotemporal causal graph based on topological relationships, mapping the residual features to a decay energy factor, using Kalman filtering to estimate the link decay propagation state, and obtaining a path cumulative deviation index; extracting standard basis vectors based on historical normal cumulative deviation indices to construct a normal change subspace; orthogonally projecting the real-time situation vector to obtain a deviation vector; calculating the deviation vector norm and relative entropy; triggering an early warning when both simultaneously exceed an adaptive threshold for a preset duration, and tracing the upstream node with the largest contribution. This invention can reduce operating condition and periodic interference, reduce false alarms and missed alarms, and achieve early and reliable early warning.

Description

Technical Field

[0001] This invention relates to the field of data processing technology. More specifically, this invention relates to a method and system for dynamic analysis and early warning in smart parks. Background Technology

[0002] In smart parks, critical equipment such as refrigeration compressors and circulating water pumps need to be continuously monitored to ensure operational safety. Anomaly detection is usually achieved by using fixed baselines or absolute thresholds, which compares real-time parameters such as vibration and temperature with historical averages or set limits. Once these limits are exceeded, an alarm is triggered.

[0003] However, equipment status readings are deeply coupled with operating conditions such as load rate and ambient temperature. Normal changes in operating conditions can cause significant parameter drift. Fixed baselines cannot distinguish between such drift and equipment degradation, resulting in a large number of false alarms and drowning out early, weak fault characteristics such as bearing wear.

[0004] In addition, the park's equipment forms a strongly coupled directed network through refrigerant pipelines, power buses, etc. The initial degradation of a single piece of equipment will spread to the upstream and downstream along the physical link in the form of speed fluctuations, harmonic drift, etc. Traditional equal-weighted accumulation or mean analysis methods do not consider propagation time delay, energy attenuation and sensor noise, and cannot restore the true degradation spread trend from the mixed signals.

[0005] Meanwhile, the park's load is affected by circadian rhythms, production schedules, and seasons. Even with a healthy system, its monitoring indicators show regular fluctuations. Fixed threshold alarms frequently generate false alarms during peak load periods, while slight abnormal growth is missed during troughs. Summary of the Invention

[0006] To address the common problems in existing technologies that limit the accuracy and reliability of early warning of latent equipment faults, this invention provides solutions in the following aspects.

[0007] In a first aspect, the present invention provides a method for dynamic analysis and early warning of smart parks, employing the following technical solution: A method for dynamic analysis and early warning of smart parks includes:

[0008] Collect operating condition and status data of equipment during its health period, train a support vector regression model to construct a dynamic baseline response surface, obtain residual vectors, and obtain residual feature vectors through sliding window statistics;

[0009] A directed spatiotemporal causal graph is established based on the equipment topology, and the residual feature vectors of each node are mapped to the fading energy factor. The link fading propagation state is estimated by Kalman filtering and the path cumulative deviation index is calculated.

[0010] Based on the historical normal path cumulative deviation index, standard basis vectors are extracted to form a normal subspace and the normal residual distribution is determined. The real-time situation vector is orthogonally projected to obtain the deviation vector. The Euclidean norm of the deviation vector and the relative entropy between the deviation vector and the normal residual distribution are calculated. When both exceed the adaptive threshold and continue for a preset duration, an early warning is triggered and the upstream node is traced to complete the analysis and early warning.

[0011] Furthermore, the acquisition of the operating conditions and status data of the equipment during its health period includes: continuously acquiring a multi-element synchronous sequence for no less than a preset duration within a window when the equipment has undergone major repair and is confirmed to be in a healthy state. The sequence includes the effective value of vibration acceleration, shell temperature, winding temperature, operating current and voltage constituting the state vector, as well as the load factor, ambient temperature and frequency feedback of upstream associated equipment constituting the operating condition vector, and preprocessing the sequence.

[0012] Furthermore, the training of the support vector regression model to construct a dynamic baseline response surface includes: taking the working condition vector as input and the state vector as output, independently constructing a sub-support vector regression model for each dimension of the state vector, wherein the sub-support vector regression model adopts a scale-adaptive Gaussian radial basis kernel function, the scale parameter of which is adaptively set according to the median distance between the working condition vectors, and obtaining the dynamic baseline response surface of each dimension of the state by solving an optimization problem with a complexity penalty term and an insensitive interval.

[0013] Furthermore, obtaining the residual vector and obtaining the residual feature vector through sliding window statistics includes: real-time acquisition of operating condition and state data; calling the dynamic reference response surface to obtain the dynamic estimated reference state vector; subtracting the actual state vector from the dynamic estimated reference state vector dimension by dimension to obtain the instantaneous multidimensional residual vector to be analyzed; and within a preset residual statistical window width, calculating the sliding standard deviation and mean drift dimension by dimension of the instantaneous multidimensional residual vector to be analyzed to form a multidimensional residual statistical feature vector.

[0014] Furthermore, the decay energy factor includes: establishing a directed spatiotemporal causal graph based on the physical connection and interlocking logic between devices, where nodes in the graph correspond to devices and directed edges represent dependencies; for each node, compressing the multidimensional residual statistical feature vector into a scalar through weighted norm mapping to obtain an individual decay energy factor, wherein the weight coefficients of the weighted norm are set according to the sensitivity of each dimension to fault characterization.

[0015] Furthermore, the step of estimating the link fading propagation state using Kalman filtering includes: establishing a state-space model for the selected link, wherein the state-space model includes a state equation defining the hidden fading propagation state and an observation equation using the individual fading energy factor as an observation quantity, wherein the state equation includes a state transition factor matrix, a node self-injection driving factor, and process noise, and the observation equation includes observation noise; performing forward Kalman filtering and fixed hysteresis smoothing to obtain the optimal smoothed sequence of the hidden fading propagation state of each node in the link.

[0016] Furthermore, the calculation of the path cumulative deviation index includes: based on the optimal smoothing sequence, extracting the principal component of the decay amplitude, and performing a weighted integral on the principal components of the decay amplitude of the upstream node and the link end node within the integral backtracking time window to form the path cumulative deviation index, wherein the node propagation contribution weight is distributed proportionally according to the topological reciprocal of the distance of each node from the link end.

[0017] Furthermore, the step of extracting standard basis vectors to construct a normal subspace and determining the normal residual distribution includes: periodically obtaining time series segments of the path cumulative deviation index of each key link during the normal operation period of the equipment; obtaining a set of standard deviation basis feature vectors after amplitude normalization of the segments; extracting a set of orthogonal basis vectors from the set through principal component analysis to span the normal variation subspace; and statistically analyzing the distribution of historical normal residuals.

[0018] Furthermore, the completion of the analysis and early warning includes: in the early warning stage, obtaining the latest path cumulative deviation index sequence within a sliding observation window of a preset length to form a real-time situation observation vector; orthogonally projecting the real-time situation observation vector onto the normal change subspace, and using the projection residual as the deviation vector; calculating the Euclidean norm of the deviation vector as the non-orthogonal decomposition deviation degree; calculating the relative entropy between the distribution of the deviation vector and the historical normal residual distribution as the cross-entropy gain factor; comparing the non-orthogonal decomposition deviation degree and the cross-entropy gain factor with the corresponding adaptive thresholds respectively, and when both exceed the corresponding adaptive thresholds and remain at the preset holding time, triggering an early warning and tracing the upstream node that contributed the most.

[0019] Secondly, the present invention provides a smart park dynamic analysis and early warning system, employing the following technical solution: A smart park dynamic analysis and early warning system includes a processor and a memory, the memory storing computer program instructions, which, when executed by the processor, implement the aforementioned smart park dynamic analysis and early warning method. This system correspondingly implements the smart park dynamic analysis and early warning method as defined in any one of claims 1 to 9.

[0020] The present invention has at least the following beneficial effects:

[0021] 1. By utilizing the multivariate operating conditions and state sequences collected during the equipment health window, a support vector regression model with a scale-adaptive kernel function is trained to construct a dynamic benchmark response surface. This allows for the real-time acquisition of multidimensional residual vectors that only reflect the degradation components of the equipment itself, thus avoiding a large number of false anomalies.

[0022] 2. By establishing a directed spatiotemporal causal graph based on device topology, the node residual characteristics are mapped to the decay energy factor. Combined with Kalman filtering and fixed hysteresis smoothing, the link decay propagation state is optimally estimated. This decouples the time delay, attenuation and measurement noise in the transmission process. The resulting path cumulative deviation index reflects the degree of cooperative deterioration, providing a reliable measurement basis for early fault detection and propagation path tracing.

[0023] 3. By extracting standard basis vectors from the historical normal path cumulative deviation index to form a normal change subspace, the deviation vector is obtained by orthogonally projecting the real-time situation vector. The Euclidean norm of the deviation vector and the relative entropy with the historical normal residual are used as early warning criteria. Adaptive threshold and dual-condition judgment are adopted to reduce false alarms and false alarms. While maintaining high sensitivity, it can achieve a reliable early warning of hidden faults and accurately trace the upstream node that contributes the most. Attached Figure Description

[0024] Figure 1 The schematic diagram illustrates the steps of a smart park dynamic analysis and early warning method and system according to the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0027] S1: Collect operating condition and status data of the equipment during its health period, train a support vector regression model to construct a dynamic baseline response surface, obtain residual vectors, and obtain residual feature vectors through sliding window statistics.

[0028] Personnel with relevant work experience should select a window period during which the equipment has undergone major repairs and is confirmed to be in good working order, and continuously collect multivariate synchronous sequences for no less than 72 hours. The sequences include effective values ​​of vibration acceleration. (Units are meters per second squared) Shell temperature (Unit: degrees Celsius) Winding temperature (Unit: degrees Celsius), Operating current (Unit: Ampere) and voltage The state vector (unit: volts) and by load factor (Dimensionless) Ambient temperature (Unit: degrees Celsius) and frequency feedback from upstream related equipment The working condition vector (unit: Hertz) consists of... The 72-hour duration is an empirical value and can be adjusted by the implementer according to the specific implementation scenario.

[0029] And apply the window width to all original sequences. The midpoint filtering and outlier removal preprocessing of each sampling point in a sliding window yield a cleaned working condition-state synchronization dataset.

[0030] To clean up the working condition vectors in the dataset Input and state vector For the output, a support vector regression model is independently constructed for each dimension of the state vector, and the model kernel function adopts a scale-adaptive Gaussian radial basis function. ,in, The kernel scale parameter is adaptively set based on the median Euclidean distance between the working condition vectors, and the training complexity penalty coefficient is used. We take an empirical value of 1.0 and consider it to be a dimensionless hyperparameter.

[0031] Furthermore, the dynamic baseline response surface for each state dimension can be obtained by solving the quadratic programming problem of standard support vector regression. This response surface can output the corresponding estimated baseline state value by accepting the operating condition vector.

[0032] During the monitoring phase, each frame of real-time operating condition vector is obtained. With synchronization state vector Immediately The trained dynamic reference response surface is fed into the dynamic prediction reference state vector under the current operating condition. Press again Subtracting the relationships dimension by dimension yields the instantaneous multidimensional residual vector to be analyzed. .

[0033] In the formula, For real-time sampling time, Each component represents the pure abnormal deviation value remaining after the normal drift of the stripping working condition in the corresponding dimension, and the unit is consistent with the corresponding state quantity.

[0034] Preset residual statistical window width Within a time interval of minutes, the instantaneous multidimensional residual vector to be analyzed... The moving standard deviation and moving mean drift of the sequence are calculated dimension by dimension, and the moving standard deviation and moving mean drift obtained from all dimensions are combined into a vector to form the device's value at the current time. Multidimensional residual statistical eigenvectors .

[0035] This vector suppresses instantaneous spike disturbances while compressing the time-series dynamic information of the residuals. The residual statistical window width is an empirical value that can be adjusted by the implementer according to the specific implementation scenario.

[0036] S2: Establish a directed spatiotemporal causal graph based on the equipment topology, map the residual feature vectors of each node to the fading energy factor, estimate the link fading propagation state through Kalman filtering, and calculate the path cumulative deviation index.

[0037] Because the equipment in the park forms a strongly coupled directed network through refrigerant pipelines, cooling water circuits, power busbars and linkage control logic, the initial degradation of components such as bearings inside a single piece of equipment is transmitted to upstream and downstream equipment along the physical link in the form of small speed fluctuations and harmonic drift, and the sensor measurement noise of each node is superimposed during the transmission process.

[0038] If we directly perform equal-weighted summation or mean analysis on the residual statistical characteristics of each node, we will be unable to distinguish between the decay energy transmitted from upstream and the degradation energy of the node itself, nor will we be able to decouple the transmission time delay and attenuation effects, making it difficult to restore the true decay spread trend.

[0039] Therefore, a directed spatiotemporal causal graph is established based on the actual physical connections and interlocking logic between devices. ,node Corresponding devices and their monitoring units, directed edges The direction from upstream equipment to downstream equipment indicates a physical transport or control dependency.

[0040] Obtain the multidimensional residual statistical feature vectors of each node at the current time. It is compressed into a scalar individual decline energy factor through weighted norm mapping. .

[0041] In the formula, This represents the element-wise multiplication of a vector with a diagonal weight matrix, where the diagonal weight matrix... The diagonal elements are the weight coefficients for each residual statistical dimension. Based on the sensitivity of each dimension to fault characterization, they are determined through Fisher discriminant analysis of historical fault samples. Typical empirical values ​​are taken as 0.5 for vibration dimension, 0.3 for temperature dimension, and 0.2 for current dimension. These are dimensionless proportional coefficients, which can also be adjusted by the implementer according to the specific implementation scenario. This energy factor characterizes the abnormal energy amplitude generated by the node due to the degradation of its own components, and is dimensionless.

[0042] For the physical links within the park that require key monitoring, the nodes are numbered in ascending topological order as follows: Define the hidden decay propagation state And establish a discrete state-space model.

[0043] The state transition equation is ,in, This is the state transition factor matrix, reflecting the energy transfer efficiency of decay between upstream and downstream nodes. Its elements are initialized based on physical parameters such as pipeline length and the inertia of the transfer medium, and can be adaptively corrected using historical operating data. For node self-injection of driving factors, For the first Real-time decay energy factor at nodes, process noise It follows a pattern with a mean of zero and a covariance matrix of... The normal distribution Determined by the variance statistic of the residuals during the historical quiet periods of this node; the observation equation is: Observation Take real-time decay energy factor Observation matrix Take the identity matrix and observe the noise. Follows a pattern with a mean of zero and a variance of . The normal distribution Similarly, the residual variance statistic for the historical quiet period at this node is determined.

[0044] At each monitoring time, standard Kalman filtering is performed sequentially from top to bottom on the link to obtain the forward filter estimate. Then, a fixed lag smoothing process is applied to the filtering results to eliminate the filtering delay and improve the smoothness of the state estimation, thus obtaining the optimal smoothed sequence of hidden decay propagation states for each node. .

[0045] Kalman filtering and fixed hysteresis smoothing are well-known signal processing techniques, and their recursive processes will not be elaborated here. The resulting smoothed sequence reflects the instantaneous spatial distribution of fading energy accumulation along the link.

[0046] Based on the optimal smoothed sequence, the principal component of the recession magnitude is extracted and the path cumulative deviation index is calculated. This indicates the degree of collaborative degradation along the physical link.

[0047] From the state vector Extract the principal component of the recession magnitude. Using the Euclidean norm of vectors, i.e. The dimensionless path cumulative deviation index is calculated as follows:

[0048]

[0049] In the formula, The integral backtracking time window is set to an empirical value of 6 hours, expressed in hours (h). This is a hyperparameter and can also be adjusted by the implementer according to the specific implementation scenario.

[0050] For the first The propagation contribution weight of each node is distributed proportionally according to the inverse of the topological distance of each node from the end of the link, so that the closer the node is to the end, the higher its upstream propagation influence weight, and it is dimensionless.

[0051] The index measures the degree of co-correlation between the estimated degradation energy accumulated at the end of the link and the degradation energy of each upstream node in the form of a weighted projected integral. Its continuous increase indicates that the degradation is co-enhancing along this path.

[0052] S3: Extract standard basis vectors based on the historical normal path cumulative deviation index to construct a normal subspace and determine the normal residual distribution.

[0053] In the daily operation of the park facilities, even if the equipment itself is in good condition, its path cumulative deviation index will still exist. It will still exhibit regular fluctuations due to diurnal load cycles, production schedules, and seasonal changes in ambient temperature.

[0054] For example, during peak summer cooling periods or early morning start-up times, fluctuations in normal operating conditions can cause... The appearance of temporary spikes in peak values ​​is not caused by hidden equipment malfunctions.

[0055] If such normal fluctuations are not distinguished from true degradation trends and are judged directly, it will lead to frequent false alarms during periods of high load and underreporting of slight abnormal growth during transition seasons.

[0056] Therefore, it is necessary to analyze historical data during the long-term normal operation period of the equipment. A baseline pattern representing the range of normal operating conditions is extracted from the sequence, and an adaptive reference space capable of absorbing legitimate fluctuations is constructed.

[0057] Therefore, during the period when the equipment is confirmed to be operating stably without faults, the processing unit calculates the cumulative path deviation index for each critical physical link in 24-hour cycles. Slicing is performed to obtain time series segments for the day. The 24-hour period is an empirical value that can be adjusted by the implementer according to the specific implementation scenario.

[0058] Each segment is normalized by its maximum value, so that waveforms under different load levels can be used for shape feature comparison.

[0059] The normalized segments from multiple consecutive days are accumulated to form the standard deviation basis eigenvector set. In the formula, each It is a column vector whose elements correspond to the normalized samples taken at fixed time intervals within that day. Value, dimensionless.

[0060] This collection covers various typical fluctuation patterns resulting from normal operating cycles, seasonal adjustments, and normal inter-equipment linkages.

[0061] For sets Principal component analysis was performed to extract the top eigenvalues ​​from the eigenvalue distribution that explained at least 95% of the cumulative variance. The eigenvectors corresponding to each principal component are orthogonally transformed to obtain a set of orthonormal basis vectors. Of these, 95% is an empirical value and can be adjusted by the implementer according to the specific implementation scenario.

[0062] These basis vectors span a Normal change subspace The physical meaning of this subspace is: any historical event caused solely by conventional operational factors. Fluctuations can all be approximated as The linear combination of basis vectors, and the components that cannot be expressed by this combination, can be attributed to unusual latent anomalous states.

[0063] At the same time, the processing unit sends data to each historical segment. Projected residual vector Probability density estimation is performed, denoted as , used as a normal residual distribution, in the formula This represents the vector dot product operation.

[0064] Once the early warning phase begins, the processing unit acquires the current sliding situation observation window width in each analysis cycle. Latest within hours Sequence, where, For hyperparameters, an empirical value of 12 hours is used, in hours (h). It can also be adjusted by the implementer based on link inertia.

[0065] The sequence is processed using sampling interval and amplitude normalization methods consistent with the standard deviation basis eigenvector, and then concatenated to form a real-time situation observation vector. .

[0066] because It contains both normal operating condition fluctuations and potential abnormal degradation trends, and must be decomposed into two parts: components that can be explained by normal patterns and unexplained residues, in order to expose novel degradation trends.

[0067] Decomposition operation by Subspace of normal change This is achieved by orthogonal projection, and the projection residual is the spatial deviation eigenvector. The calculation relationship is as follows:

[0068]

[0069] In the formula, for In the Projected coordinates on an orthonormal basis, summation term represent exist The best approximation within, and It is the residual vector after approximation, and its physical meaning is the spatial pattern component in the current situation that cannot be linearly expressed by any historical normal cycle pattern. It is dimensionless.

[0070] get Then, it was quantitatively determined from two dimensions: amplitude and distribution structure.

[0071] The magnitude is calculated using the Euclidean norm of the vector, defined as the nonorthogonal decomposition deviation:

[0072]

[0073] In the formula, The Euclidean norm of a vector. It is a scalar quantity and is dimensionless.

[0074] This deviation measure measures the overall magnitude of the deviation of the current situation from all normal modes. The larger the value, the more significant the component that the real-time observation vector cannot be absorbed by the normal subspace.

[0075] Differences in distribution structure are characterized by relative entropy.

[0076] The processing unit will currently The element values ​​are statistically analyzed to obtain the distribution. And calculate its difference from the pre-stored normal residual distribution. The relative entropy between them yields the cross-entropy gain factor:

[0077]

[0078] In the formula, the integral iterates through... The range of values ​​for each component It is the natural logarithm function. For scalars, the unit for entropy-free quantities is the nat.

[0079] This factor reflects the statistical structural difference between the probability distribution shape of the current deviation vector and the historical normal random fluctuation shape. The higher the value, the stronger the mutual exclusion between the two distributions. That is, the current situation not only deviates in magnitude, but also shows non-accidental abnormal characteristics in pattern structure.

[0080] The final warning trigger adopts a dual-condition parallel judgment to avoid misjudgment from a single dimension.

[0081] The processing unit will decompose the non-orthogonal deviation. With its adaptive threshold Comparison, this threshold Taken from all historical data during the healthy operation of the equipment The 95th quantile of the value distribution is dimensionless and can be adaptively updated as the equipment operates.

[0082] At the same time, the cross-entropy gain factor With discrimination threshold Compare, The empirical value is 0.25, in nats (kN), which is a hyperparameter and can be adjusted by the implementer based on the acceptable false alarm rate.

[0083] when and When both conditions are met, the timer starts and checks whether the state is maintained continuously for the preset duration. minutes, of which As an hyperparameter, an empirical value of 30 minutes is used, in minutes (min), but it can also be adjusted by the implementer.

[0084] If this dual condition is interrupted at any time, the timer will reset to zero.

[0085] when and Simultaneously continuously established and exceeding [a certain number] without interruption. Subsequently, the system determined that the current link had shown a hidden fault co-deterioration trend that could not be explained by the normalized change set, and then received an early warning signal and pushed it to the operation and maintenance terminal.

[0086] During the warning, the processing unit simultaneously... Principal components of the decline magnitude of each upstream node within each time period The correlation strength is traced to identify the upstream node that contributes the most to the deviation vector within the sliding trend observation window and its physical link, providing spatial positioning information for operation and maintenance personnel.

[0087] This invention also discloses a smart park dynamic analysis and early warning system, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the smart park dynamic analysis and early warning method of this invention. The system also includes other components well-known to those skilled in the art, such as a communication bus and communication interface; their configurations and functions are known in the art and will not be described further here.

[0088] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make several improvements and substitutions without departing from the technical principles of the present invention, and these improvements and substitutions should also be considered within the scope of protection of the present invention.

Claims

1. A method for dynamic analysis and early warning in smart parks, characterized in that, include: Collect operating condition and status data of equipment during its health period, train a support vector regression model to construct a dynamic baseline response surface, obtain residual vectors, and obtain residual feature vectors through sliding window statistics; A directed spatiotemporal causal graph is established based on the equipment topology, and the residual feature vectors of each node are mapped to the fading energy factor. The link fading propagation state is estimated by Kalman filtering and the path cumulative deviation index is calculated. Based on the historical normal path cumulative deviation index, standard basis vectors are extracted to form a normal subspace and the normal residual distribution is determined. The real-time situation vector is orthogonally projected to obtain the deviation vector. The Euclidean norm of the deviation vector and the relative entropy between the deviation vector and the normal residual distribution are calculated. When both exceed the adaptive threshold and continue for a preset duration, an early warning is triggered and the upstream node is traced to complete the analysis and early warning.

2. The method for dynamic analysis and early warning of a smart park according to claim 1, characterized in that, The data collected during the health period of the acquisition equipment includes: Within a window after the equipment has undergone major repairs and is confirmed to be in a healthy state, a multi-element synchronization sequence of no less than a preset duration is continuously collected. The sequence includes the effective value of vibration acceleration, shell temperature, winding temperature, operating current and voltage constituting the state vector, as well as the load factor, ambient temperature and frequency feedback of upstream associated equipment constituting the operating condition vector. The sequence is then preprocessed.

3. The method for dynamic analysis and early warning of a smart park according to claim 1, characterized in that, The training of the support vector regression model to construct a dynamic baseline response surface includes: taking the working condition vector as input and the state vector as output, independently constructing a sub-support vector regression model for each dimension of the state vector. The sub-support vector regression model adopts a scale-adaptive Gaussian radial basis kernel function, whose scale parameter is adaptively set according to the median distance between the working condition vectors, and obtaining the dynamic baseline response surface of each dimension of the state by solving an optimization problem with a complexity penalty term and an insensitive interval.

4. The method for dynamic analysis and early warning of a smart park according to claim 1, characterized in that, The process of obtaining the residual vector and then using sliding window statistics to obtain the residual feature vector includes: Real-time acquisition of operating conditions and status data; calling the dynamic reference response surface to obtain the dynamic predicted reference state vector; subtracting the actual state vector from the dynamic predicted reference state vector dimension by dimension to obtain the instantaneous multidimensional residual vector to be analyzed. Within a preset residual statistical window, the sliding standard deviation and mean drift are calculated dimension by dimension for the instantaneous multidimensional residual vector to be analyzed, forming a multidimensional residual statistical feature vector.

5. The method for dynamic analysis and early warning of a smart park according to claim 1, characterized in that, The decay energy factor includes: A directed spatiotemporal causal graph is established based on the physical connections and interlocking logic between devices. Nodes in the graph correspond to devices, and directed edges represent dependencies. For each node, the multidimensional residual statistical feature vector is compressed into a scalar through weighted norm mapping to obtain the individual decay energy factor, wherein the weight coefficients of the weighted norm are set according to the sensitivity of each dimension to the fault representation.

6. The method for dynamic analysis and early warning of a smart park according to claim 1, characterized in that, The estimation of link fading propagation state using Kalman filtering includes: A state-space model is established for the selected link. The state-space model includes a state equation that defines the hidden decay propagation state and an observation equation that uses the individual decay energy factor as an observation quantity. The state equation includes a state transition factor matrix, a node self-injection driving factor and process noise, and the observation equation includes observation noise. By performing forward Kalman filtering and fixed hysteresis smoothing, the optimal smoothing sequence of the hidden decay propagation state of each node in the link is obtained.

7. The method for dynamic analysis and early warning of a smart park according to claim 1, characterized in that, The calculation path cumulative deviation index includes: Based on the optimal smoothing sequence, the principal component of the decay amplitude is extracted. Within the integral backtracking time window, the principal components of the decay amplitude of the upstream node and the link end node are weighted and integrated to form the path cumulative deviation index. The node propagation contribution weight is distributed proportionally according to the topological reciprocal of the distance of each node from the link end.

8. The method for dynamic analysis and early warning of a smart park according to claim 1, characterized in that, The extraction of standard basis vectors to construct a normal subspace and determine the normal residual distribution includes: During normal equipment operation, time series segments of the path cumulative deviation index of each critical link are periodically obtained. After amplitude normalization of the segments, the standard deviation basis feature vector set is obtained. Principal component analysis is used to extract a set of orthogonal basis vectors from the set, which spans a normally changing subspace. And statistically analyze the distribution of historical normal residuals.

9. The method for dynamic analysis and early warning of a smart park according to claim 1, characterized in that, The completed analysis and early warning includes: During the early warning phase, the latest path cumulative deviation index sequence within a sliding observation window of a preset length is used to construct a real-time situation observation vector. The real-time situation observation vector is orthogonally projected onto the normal change subspace, and the projection residual is used as the deviation vector. Calculate the Euclidean norm of the deviation vector as the nonorthogonal decomposition deviation degree; The relative entropy between the distribution of the deviation vector and the historical normal residual distribution is calculated as the cross-entropy gain factor; The nonorthogonal decomposition deviation and cross-entropy gain factor are compared with their corresponding adaptive thresholds. When both exceed the corresponding adaptive thresholds and remain for a preset duration, an early warning is triggered and the upstream node that made the largest contribution is traced.

10. A smart park dynamic analysis and early warning system, characterized in that, The system includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement a smart park dynamic analysis and early warning method as described in any one of claims 1 to 9.

Images