A city block lighting intelligent scheduling system with double-path closed-loop adaptive optimization

Abstract

The present application relates to the technical field of smart city lighting control, and proposes a city block lighting scheduling system with double-loop closed-loop adaptive optimization. The system integrates lighting equipment operation data, environmental perception data and user demand data to construct multi-source fusion operation signals under a unified time scale. After Gammatone filtering, time-frequency analysis, singular spectrum entropy complexity modeling and morphological mutation positioning, the adaptive enhanced classifier outputs classification results containing abnormal levels to drive differentiated scheduling strategy execution. The core of the system lies in double-loop parallel optimization: the first path continuously updates the classification model parameters through back propagation to optimize the abnormal judgment logic; the second path dynamically adjusts the fusion weight according to the contribution of each data source to optimize the sensor perception strategy. The two paths evolve cooperatively, enabling the system to automatically adapt to the drift of operation characteristics during long-term operation, and achieve a continuous optimization balance between lighting stability and energy consumption control.

Description

Technical Field

[0001] This invention relates to the field of smart city lighting control technology, specifically to a smart scheduling system for urban street lighting with dual-path closed-loop adaptive optimization. Background Technology

[0002] Urban street lighting systems are a crucial component of smart city infrastructure, and their operational status directly impacts urban energy consumption, public safety, and residents' quality of life. Currently, urban lighting dispatch systems generally suffer from the following technical deficiencies:

[0003] Traditional lighting control systems rely primarily on single equipment operating data, such as electrical parameters like power, current, and voltage, lacking the ability to perceive environmental factors and user needs. This limitation of a single data source prevents the system from fully understanding the lighting operation status, easily leading to insufficient lighting or energy waste in densely populated commercial districts or areas with frequent climate changes.

[0004] Existing signal processing techniques are insufficient in suppressing high-frequency noise and interference. Simple mean or median filtering is insufficient to effectively distinguish useful signals from unwanted interference; conventional filters such as Butterworth or Chebyshev filters use uniform frequency resolution, which cannot adapt to the characteristics of lighting operation signals where low-frequency energy is concentrated and high-frequency harmonics are dense; while variational mode decomposition (VMD) is highly dependent on manually set parameters, and improper parameter selection can lead to signal distortion or mode aliasing, affecting the accuracy of subsequent analysis. None of the above methods can adaptively allocate frequency resolution based on signal energy distribution over a wide bandwidth. For example, Shang Qiufeng et al. (Control and Decision, 2024) decomposed signals using the VMD algorithm, but the performance of this method is highly dependent on the manual setting of the number of decomposition layers and the penalty factor, and inappropriate parameter settings can easily lead to problems such as mode aliasing, over-envelope, or under-envelope.

[0005] Feature extraction methods suffer from problems such as excessive dimensionality and redundant information. Existing techniques often employ high-dimensional feature representations such as Mel spectra and cepstral coefficients, leading to high computational costs and poor real-time performance. Short-time Fourier transforms are constrained by the Heisenberg uncertainty principle, making it impossible to simultaneously achieve both time and frequency resolution, thus hindering the accurate localization of the anomaly occurrence time and frequency components of non-stationary signals. While the Wigner-Ville distribution exhibits good time-frequency clustering, the cross-term interference generated by multi-component signals severely impacts analysis accuracy. Furthermore, existing methods do not adequately consider the correlation and redundancy between features; high-dimensional redundant features not only increase computational burden but also introduce noise interference, reducing the accuracy of subsequent classification and recognition.

[0006] Anomaly detection models exhibit weak generalization ability. While deep learning-based detection methods perform well with sufficient training data, their accuracy drops significantly when faced with anomaly types not covered by the training set or in noisy environments. Single classifiers often suffer from overlapping discrimination boundaries for multiple anomalies (such as lamp aging, harmonic pollution, and overload impacts), leading to high false alarm rates. Furthermore, existing methods often only output binary alarm signals, failing to differentiate anomaly levels and resulting in a lack of targeted scheduling responses. In addition, these methods are insufficiently adaptable to dynamic factors such as voltage fluctuations and load changes, making it difficult to maintain stable detection accuracy in complex electrical disturbance environments like urban lighting.

[0007] Most systems employ an open-loop control structure, lacking feedback and optimization of execution results after the scheduling strategy is implemented. If initial parameters are improperly set or the operating environment changes, the system cannot automatically adjust, requiring manual intervention for reconfiguration, resulting in high maintenance costs and slow response times. Therefore, a smart urban lighting scheduling system is needed that can integrate multi-source data, effectively suppress noise interference, accurately extract features, accurately identify anomalies, and possess adaptive optimization capabilities. Summary of the Invention

[0008] Existing urban lighting dispatch systems have significant shortcomings in data fusion, signal processing, feature extraction, anomaly detection, and adaptive optimization, making it difficult to meet the requirements of smart cities for efficient, stable, and intelligent operation of lighting systems. There is an urgent need to develop an intelligent urban street lighting dispatch system that integrates multi-source data, employs advanced signal processing technology, possesses high-precision anomaly detection capabilities, and has a closed-loop adaptive optimization mechanism.

[0009] To address this, a smart urban street lighting scheduling system with dual-path closed-loop adaptive optimization is proposed, comprising:

[0010] The data acquisition module is used to collect operating data, environmental perception data, and user demand data of urban street lighting equipment, form a time-series operating signal including power, current, and voltage, as well as a synchronous sampling sequence of environment and demand, and output the various types of sampled data to the data fusion module;

[0011] The data acquisition module includes three functional sub-units: an electrical parameter acquisition unit, an environmental perception acquisition unit, and a demand information acquisition unit.

[0012] The electrical parameter acquisition unit collects power data of the lighting equipment in real time through current transformers and voltage sensors installed in each street light controller. Current data and voltage data The sampling frequency is not less than 100 Hz;

[0013] The environmental sensing and acquisition unit collects light intensity data through illuminance sensors, temperature sensors, and video / infrared people counting sensors distributed throughout the street. Ambient temperature data and crowd density data The sampling period of each sensor is synchronized with the electrical parameter acquisition unit;

[0014] The demand information collection unit obtains user lighting brightness demand data through the urban lighting management platform interface. Lighting demand data The data is then aligned with a unified timestamp and output to the data fusion module.

[0015] The data fusion module is connected to the data acquisition module and performs fusion processing on the operational data, environmental perception data and user demand data to construct a multi-source fusion signal.

[0016] The data fusion module performs fusion processing on multi-source data using a weighted fusion method; the weighted fusion method assigns fusion weights based on the reliability and importance of different data sources; the multi-source fused signal reflects the comprehensive characteristics of the lighting equipment's operating status, environmental changes, and user needs.

[0017] Operational data includes power, current, and voltage data of lighting equipment; environmental perception data includes light intensity, temperature, and pedestrian density data; user demand data includes lighting brightness requirements and lighting time period requirements.

[0018] The time-domain filtering module receives the time-series running signal, suppresses high-frequency fluctuation components through a time-domain filter bank, distinguishes between useful running signals and useless interference signals, and outputs the filtered running signal to the time-frequency analysis module.

[0019] The time-frequency analysis module receives the filtered operating signal, constructs a time-frequency distribution feature reflecting the operating state in a two-dimensional time-frequency space using a time-frequency transformation method, and outputs the time-frequency distribution feature to the feature aggregation module.

[0020] The feature aggregation module receives the time-frequency distribution features, enhances the energy aggregation of the time-frequency features through synchronous compression processing, performs dimensionality reduction processing on the aggregated time-frequency features through a feature selection algorithm, extracts intrinsic modal features that characterize the essential characteristics of the running state, and outputs the intrinsic modal features to the complexity calculation module.

[0021] The complexity calculation module receives the intrinsic mode features, obtains the singular value spectrum through singular value decomposition, calculates the singular spectral entropy index reflecting the complexity of the energy distribution of the running signal based on the singular value spectrum, and outputs the singular spectral entropy index to the mutation identification module.

[0022] The mutation identification module receives the singular spectral entropy index, locates the mutation point in the singular spectral entropy index through the edge detection operator, identifies the time interval in which the operating state changes abnormally, and outputs the feature data of the abnormal change time interval to the classification identification module.

[0023] The classification and identification module receives the feature data of the abnormal change time interval, uses an adaptive neural network to classify and identify the operating status, outputs the abnormal operating status category, and outputs the abnormal operating status category to the scheduling response module.

[0024] The scheduling response module receives the abnormal operating status category, triggers the corresponding lighting scheduling strategy according to the preset mapping relationship between the abnormal category and the scheduling strategy, generates a scheduling instruction containing scheduling time, scheduling object, and scheduling parameters, and sends the scheduling instruction to the lighting equipment for execution. At the same time, the scheduling execution information is output to the feedback update module.

[0025] The feedback update module receives the scheduling execution information, collects the running data after scheduling execution, calculates the error between the prediction result of the classification and recognition module and the actual running state, updates the connection weights and bias parameters of the adaptive neural network through the backpropagation algorithm, and feeds back the updated network parameters to the classification and recognition module, forming a closed-loop adaptive optimization structure of recognition-scheduling-feedback.

[0026] The time-domain filtering module uses a Gammatone filter bank for time-domain filtering.

[0027] The Gammatone filter bank includes multiple Gammatone filters with different center frequencies, and the impulse response function of each filter is determined according to the filter order, bandwidth parameter and center frequency parameter.

[0028] The center frequency parameter is obtained by dividing the target frequency range into equal-distance segments on an equivalent rectangular bandwidth scale.

[0029] The core reason for choosing the Gammatone filter bank instead of a conventional uniform bandwidth filter bank is that the harmonic interference frequency structure (fundamental frequency and its harmonic components) of urban lighting power grid signals has a natural correspondence with the non-uniform frequency division on the ERB scale. Traditional uniform bandwidth FIR filter banks oversample in the low-frequency band and have insufficient resolution in the high-frequency band, failing to adaptively match the characteristics of concentrated low-frequency energy and dense high-frequency harmonics in lighting operation signals. The Gammatone filter bank improves the frequency resolution of low-frequency harmonic components (1-3 Hz segment) on the ERB scale by about 2-3 times compared to the uniform bandwidth filter bank, effectively distinguishing the fundamental frequency component and harmonic interference components of the urban power grid. It avoids the inherent defect of uniform division methods in handling the aliasing of adjacent harmonics in the low-frequency band, which is something other linear filter banks cannot achieve with the same computational cost.

[0030] The time-domain filtering module also includes a gain control unit, which is used to normalize the amplitude of the filtered signal so that the amplitude of the signal of different frequency components is unified to a preset reference range.

[0031] The gain control unit adaptively adjusts the gain factor to determine the optimal gain factor value based on the signal-to-noise ratio of the filtered signal.

[0032] The gain control unit adaptively adjusts the gain factor and determines the optimal gain factor value based on the signal-to-noise ratio of the filtered signal. The specific algorithm is as follows:

[0033] Let the first The output signal of each filter channel is In length of Calculate signal power and noise power within the estimation window:

[0034] ;

[0035] Noise power is calculated using the signal in the high-frequency subband ( Approximate estimation of the output power of the channel:

[0036] ;

[0037] The current signal-to-noise ratio estimate is Optimal gain factor Determined by the following rules:

[0038] ;

[0039] in This is the upper limit of the gain (to prevent saturation due to excessive gain; the default value is 1.0). The gain normalization coefficient (set according to the target dynamic range, default is 0). , (The system reference signal-to-noise ratio is 30 dB by default); the normalized discrete impulse response unifies the signal amplitude of different frequency channels to the preset reference range, while adaptively compensating for the signal attenuation of low signal-to-noise ratio channels.

[0040] The time-frequency analysis module uses a combination of the Cui-Williams distribution analysis method and the multiple synchronous compression transform analysis method to construct time-frequency distribution characteristics;

[0041] The described Twee-Williams distribution analysis method is based on the kernel function to perform time-frequency transformation on the running signal, and the kernel function bandwidth is adjusted by a periodic controllable factor to suppress cross-term interference in the time-frequency distribution;

[0042] The aforementioned multi-synchronous compression transformation analysis method segments the running signal using short-time Fourier transform combined with a window function and performs synchronous compression in the frequency direction.

[0043] CWD suppresses cross-term interference through kernel functions, addressing the problem of "spurious frequency components" in multi-component signals. MSST enhances energy concentration through synchronous compression, addressing the problem of insufficient time-frequency resolution leading to blurred frequency components. CWD alone can suppress cross-terms but has limited frequency resolution; MSST alone can improve concentration but cannot eliminate cross-terms. Only the serial combination of the two can overcome both of these shortcomings simultaneously, achieving a clear, spurious-free time-frequency representation in the analysis of non-stationary urban lighting signals.

[0044] When performing synchronous compression processing, the feature aggregation module divides the running signal into multiple time windows using a window function, performs a short-time Fourier transform on the signal in each window, and obtains local time-frequency features.

[0045] The synchronous compression process compresses the short-time Fourier transform results in the frequency direction, concentrating the dispersed frequency components towards the instantaneous frequency position and enhancing the concentration of time-frequency energy.

[0046] The feature aggregation module uses the Fisher Score algorithm for feature dimensionality reduction;

[0047] The Fisher Score algorithm evaluates the contribution of each time-frequency feature to the classification of abnormal states by calculating the ratio of inter-class dispersion to intra-class dispersion. Features with a contribution below a preset threshold are removed based on their contribution.

[0048] The extraction of the intrinsic mode features is achieved through a local mean decomposition algorithm, which obtains the intrinsic mode components that characterize the inherent oscillation mode of the operating state by iteratively separating the local mean component and the pure frequency modulation component in the operating signal.

[0049] The key reason for choosing LMD over VMD or EMD is that LMD's iteration termination condition is automatically determined by the distribution of extreme points in the signal itself, eliminating the need for manually preset decomposition levels or penalty factors. This provides an adaptive advantage for urban lighting anomaly signals (diverse anomaly types and unknown signal morphologies). Background technology has pointed out that VMD is highly dependent on manual parameter settings, and inappropriate parameters can lead to modal aliasing, while EMD suffers from endpoint effects and modal aliasing. LMD overcomes these shortcomings through an adaptive extreme point detection mechanism, enabling the system to automatically determine a reasonable number of decomposition levels when facing different types of anomaly signals, such as aging lamps, harmonic pollution, and overload impacts, ensuring the effectiveness of intrinsic modal characteristics.

[0050] When calculating the singular spectral entropy, the complexity calculation module first performs singular value decomposition on the feature matrix composed of intrinsic mode features to obtain a sequence of singular values.

[0051] Then, the singular value sequence is permuted and combined, and the probability of different permutation patterns is calculated.

[0052] Based on the probability, the singular spectrum entropy value is calculated using the permutation entropy formula. The singular spectrum entropy value reflects the distribution complexity of the running signal energy among different modal components.

[0053] The mutation identification module uses a combination of dilatational edge detection operators and erosional edge detection operators to locate mutation points;

[0054] The dilatational edge detection operator is based on the signal opening operation morphological spectrum. It performs an opening operation on the singular spectral entropy index through structuring elements to extract rising edge abrupt features.

[0055] The erosion-type edge detection operator is based on the signal closing operation morphological spectrum. It performs a closing operation on the singular spectral entropy index through structuring elements to extract the falling edge abrupt feature.

[0056] By combining the rising edge abrupt change characteristics and the falling edge abrupt change characteristics, the location and direction of the abrupt change point in the operating state are determined.

[0057] The classification and recognition module uses an adaptive augmentation classifier based on a multilayer perceptron to classify abnormal operating states;

[0058] The adaptive enhancement classifier trains multiple multilayer perceptron weak classifiers iteratively. In each iteration, it adjusts the weights of the training samples based on the classification error of the previous classifier, increases the weights of misclassified samples, and builds a strong classifier.

[0059] The multilayer perceptron includes an input layer, at least one hidden layer, and an output layer. The input layer receives dimensionality-reduced intrinsic mode features. The hidden layer performs feature transformation through a nonlinear activation function. The output layer obtains the final classification result by weighted voting on the outputs of multiple weak classifiers.

[0060] The dual-path parallel update mechanism of the feedback update module is as follows:

[0061] First path: Construct a cross-entropy loss function by calculating the error between the predicted category of the classification and recognition module and the actual running state after scheduling execution; use the backpropagation algorithm to calculate the gradient of the loss function with respect to the connection weights and bias parameters of the adaptive augmentation classifier, and update the connection weights and bias parameters according to the preset learning rate to update the system's judgment decision logic for anomaly types;

[0062] The second approach is to update the fusion weights of the data fusion module by gradient descent based on the Fisher Score contribution of each data source feature in the current abnormal sample, thereby increasing the perception weights of high-contribution data sources and updating the system's perception emphasis strategy for different sensor information; and to normalize the updated fusion weights so that the sum of the fusion weights is always equal to 1.

[0063] The coupling relationship between the two paths is as follows: the current anomaly type determines the contribution distribution of each data source, which in turn drives the fusion weight to tilt towards the data source with high contribution, so that the system can automatically focus on the sensor with the most information in different operating environments, and realize the collaborative adaptive evolution of the identification strategy and the perception strategy.

[0064] The specific derivation of the method for calculating the fused weight gradient in the second path is as follows:

[0065] Let the first The Fisher Score for the current anomalous sample based on the data source features is: Normalize it to obtain the target weight distribution: ;

[0066] Define the loss function for fusion weight optimization as the mean squared error between the current weight distribution and the target weight distribution: ;

[0067] right The gradient is obtained as follows: ;

[0068] The gradient descent update rule is:

[0069] ;

[0070] After the update is completed, normalization is performed on all weights to make them normal. Hengchengli: ;

[0071] The physical meaning of the above mechanism is that when the Fisher Score contribution of a data source is higher than its current weight, the gradient is negative and the weight is adjusted upward; when the contribution is lower than the current weight, the gradient is positive and the weight is suppressed downward; after multiple iterations, the fused weight distribution gradually converges to the optimal distribution that matches the actual information contribution of each data source under the current anomaly type.

[0072] Compared with the prior art, the beneficial effects of the present invention are:

[0073] I. This invention constructs a unified analysis framework for the overall operational status by using multi-source fused operating signals as a unified representation of the operational status of urban street lighting systems. Compared with existing technologies that rely solely on single electrical parameters such as power and current for judgment, this invention weighted and fused operational data, environmental perception data, and user demand data at a unified time scale. The constructed multi-source fused signal can simultaneously reflect the comprehensive characteristics of lighting equipment operation status, external environmental changes, and user demands, fundamentally eliminating the problems of misjudgment, missed judgment, and response delay caused by single-parameter judgment. Compared with fixed-weight fusion schemes, this invention adaptively optimizes the fusion weights through a feedback update module, enabling the system to dynamically adjust the weights based on the actual contribution of each data source under different anomaly types, thus improving the sensitivity and accuracy of anomaly detection.

[0074] II. This invention employs a time-frequency analysis method combining the Cui-Williams distribution and multiple synchronous compression transform, and directly incorporates the constructed time-frequency complexity index into the anomaly detection process. Compared to the short-time Fourier transform, the Cui-Williams distribution, through adaptive adjustment of the kernel function bandwidth by a periodic controllable factor, can maintain both high time and frequency resolution, overcoming the inherent limitations of the Heisenberg uncertainty principle on the short-time Fourier transform. Compared to the Wigner-Ville distribution, the Cui-Williams distribution effectively suppresses cross-term interference in multi-component signals. The multiple synchronous compression transform further enhances the time-frequency energy concentration through iterative compression, making the frequency abrupt change characteristics in non-stationary signals more prominent. Based on this, this invention proposes a singular spectral entropy composite index. Its core innovation lies in achieving an effective compressed mapping from a high-dimensional time-frequency feature matrix to a "single quantifiable scalar": first, the feature matrix is ​​decomposed into a sequence of singular values ​​using SVD, preserving energy distribution information; then, the permutation entropy algorithm (extracting structural complexity information) is applied to the singular value sequence. This composite index carries both energy amplitude information and modal arrangement complexity information, thus creating a clear quantitative boundary between normal operation (singular spectral entropy 0.4-0.8) and structural anomalies (singular spectral entropy > 0.8).

[0075] Using singular values ​​alone cannot distinguish between so-called "orderly large fluctuations" and "complex small anomalies"; using permutation entropy alone loses energy amplitude information and also fails to establish robust anomaly judgment boundaries. The composite mapping of these two methods is a necessary condition for achieving the aforementioned distinguishing ability and is the core innovation that distinguishes this invention from existing feature extraction methods. Subsequently, a morphological edge detection operator precisely locates abrupt change points, realizing the shift from "anomaly judgment" to "anomaly interval location," significantly improving the temporal accuracy and level resolution of anomaly identification.

[0076] Third, this invention employs an adaptive boosting classifier based on a multilayer perceptron to achieve fine-grained classification of anomaly types and levels, and directly drives the generation of scheduling strategies based on the classification results. Compared with a single deep learning classifier, AdaBoost integrates multiple weak classifiers and significantly improves the ability to identify rare anomaly types through weighted iteration of misclassified samples, maintaining strong generalization performance even with limited training samples. Compared with existing solutions that only output binary alarm signals, this invention outputs classification results with information on normal, mild, moderate, and severe anomalies, and combines them with a preset scheduling strategy mapping relationship to generate differentiated scheduling response schemes for specific anomaly levels. This achieves a fundamental shift from "single-point passive adjustment" to "fine-grained control based on anomaly levels," effectively avoiding energy waste or decreased lighting stability caused by over-scheduling or under-scheduling.

[0077] IV. This invention constructs a closed-loop optimization mechanism of identification-scheduling-feedback, overcoming the fundamental defect of existing open-loop control structures that cannot adaptively adjust. Compared with existing open-loop scheduling schemes, this invention incrementally updates the connection weights and bias parameters of the classification model through a feedback update module after scheduling execution. Simultaneously, it optimizes the fusion weights using gradient descent based on the Fisher Score contribution of each data source for the current anomaly sample, transforming the fusion weights from fixed parameters to dynamically evolving, adaptive parameters. This dual-layer closed-loop mechanism—parameter updates at the classification model level and weight optimization at the data fusion level—works synergistically, enabling the system to continuously improve the identification accuracy of frequently occurring anomaly types during long-term operation and gradually enhance the perception weight of data sources with high contribution (such as power and current). This fundamentally solves the performance degradation problem caused by the drift of operating characteristics in fixed-parameter identification models during long-term deployment, significantly reducing the manual maintenance cost throughout the entire lifecycle.

[0078] It should be noted that the fundamental difference between the closed-loop mechanism of this invention and existing closed-loop control systems lies in the implementation of dual-path parallel updates: Path A involves scheduling error → backpropagation → updating MLP connection weights and bias parameters, updating the decision logic of "how to judge anomalies"; Path B involves Fisher Score dynamically evaluating the feature contribution of each data source → updating the fusion weights of the data fusion module, updating the perception strategy of "which sensors to focus on". The coupling relationship between the two paths is as follows: the current anomaly type determines which data sources have high contribution, thereby driving the fusion weights to tilt towards high-contribution data sources, enabling the system to automatically "focus" on the most informative sensors under different operating environments. Existing closed-loop systems in urban lighting scheduling literature typically only update one path of the classification model weights or only update one path of the fusion weights. This invention simultaneously achieves the coordinated update of both paths, which is the key reason why it is superior to existing solutions in terms of long-term operational stability.

[0079] V. The time-domain filtering module addresses the pollution problem caused by high-frequency interference from the power grid on subsequent time-frequency analysis, serving as a prerequisite for ensuring the quality of feature extraction. The time-frequency analysis module addresses the non-stationary characteristics of operating signals, overcoming the fundamental deficiency of traditional frequency-domain methods in being unable to pinpoint the exact time of anomalies. The feature aggregation and dimensionality reduction module addresses the high-dimensional redundancy problem of the time-frequency matrix, a necessary engineering means to ensure system real-time performance. The complexity calculation module compresses high-dimensional features into quantifiable scalar indicators, substantially simplifying the computational burden of subsequent anomaly detection. The mutation identification module addresses the shortcomings of simple threshold methods, such as high false alarm rates and the inability to pinpoint the start and end times of anomaly intervals. The classification and identification module addresses the practical needs of different scheduling strategies corresponding to multiple types of anomalies. The feedback update module addresses the model degradation problem during long-term operation, reducing the manual maintenance costs throughout the system's lifecycle. These modules have strict data flow dependencies, collectively forming a complete technical solution for the complex operating environment of urban street lighting systems.

[0080] Through the above technical solutions, this invention organically combines multi-source data fusion, time-frequency complexity modeling, adaptive anomaly identification, and intelligent scheduling control to construct a complete operational status analysis and scheduling optimization system. It maintains stable identification under normal operating conditions and achieves rapid response and fine-tuning under abnormal operating conditions, thereby improving the overall operational reliability and intelligence level of urban street lighting systems. Attached Figure Description

[0081] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort. In the drawings:

[0082] Figure 1 This is an architecture diagram of a smart urban street lighting dispatching system with dual-path closed-loop adaptive optimization. Detailed Implementation

[0083] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0084] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0085] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0086] Example 1

[0087] This embodiment illustrates the complete workflow of the intelligent urban street lighting dispatching system with dual-path closed-loop adaptive optimization described in this invention when the urban street lighting system is in normal operation, and verifies the system's identification accuracy, dispatching stability, and closed-loop operation capability under stable operating conditions. By modeling and extracting features from the normal operating state, a baseline range for system operating complexity can be established, providing a reference standard for subsequent judgment of abnormal operating states.

[0088] In real-world urban street lighting scenarios, normal operation typically involves stable fluctuations in electrical parameters, slow changes in environmental perception data, and stable user demand data within a given time period. Under these operating conditions, the system of this invention constructs a unified operating signal through multi-source data fusion, and then sequentially undergoes processing steps including time-domain filtering, time-frequency analysis, feature aggregation and dimensionality reduction, complexity calculation, mutation identification, classification identification, and scheduling response, forming a complete identification-scheduling-feedback closed-loop structure.

[0089] This data acquisition module simultaneously collects operational data, environmental perception data, and user demand data from 100 streetlights in a commercial district of a city, with a sampling frequency of [missing information]. The time window is Each record contains Data points.

[0090] (1) Operational data: including the first The power of the lamp Current and voltage ,in For example, at time Collected the first The data for each lamp is: , , .

[0091] (2) Environmental perception data: including light intensity Ambient temperature and population density For example, a certain area in The data for that time are: , , 15 / minute.

[0092] (3) User demand data: including target illuminance demand and lighting time requirements For example, the demand in commercial areas during the evening hours is: , .

[0093] The data fusion module uses a weighted fusion method to fuse three types of data. The fusion weights are determined based on the reliability and importance of the data sources.

[0094] For the A lamp constructs a multi-source fusion signal :

[0095] ;

[0096] in, To integrate the weights, the following conditions must be met:

[0097] ;

[0098] The initial weights are set as follows: , , , , , , .

[0099] After normalizing the collected raw data, a multi-source fused signal is constructed. , forming a collection Time series running signal of data points .

[0100] The time-domain filtering module uses a Gammatone filter bank to filter the multi-source fused signal and suppress high-frequency noise interference.

[0101] To avoid symbolic ambiguity, this embodiment uses Indicates the filter number ( ),use This indicates the street light number.

[0102] (1) Gammatone filter design:

[0103] Set the filter order Set the target frequency band as . No. The impulse response function of a Gammatone filter is:

[0104] ;

[0105] in, For operation parameters, For time, Let the initial phase be (set as) ), For the first The center frequency of each filter For bandwidth.

[0106] (2) Bandwidth calculation:

[0107] ;

[0108] (3) Equivalent Rectangular Bandwidth (ERB) Division: The target frequency band is divided at equal intervals on the ERB scale. The ERB is defined as:

[0109] By dividing into equal intervals, determine One center frequency:

[0110] (4) Discretization: The impulse response function is discretized, and the sampling frequency is... The filter length is set to ,but:

[0111] ;

[0112] (5) Gain control: The gain control unit normalizes the discrete impulse response of the filter:

[0113] ;

[0114] And by adaptively adjusting the gain factor (initially set to 0.2), the signal-to-noise ratio of the filtered signal is optimized, making... The useful signal characteristics are prominent in the Hz range.

[0115] The filtered output is defined as a discrete convolution of the fused signal:

[0116] ;

[0117] The filtered operating signal can be obtained by combining the channels of the filter bank. .

[0118] The time-frequency analysis module uses a combination of the Cui-Williams distribution (CWD) and multiple synchronous compression transform (MSST) to construct time-frequency distribution characteristics.

[0119] (1) CWD analysis: for the filtered running signal Corresponding continuous representation Perform CWD transformation, with the kernel function as follows:

[0120] ;

[0121] in, As a periodic and controllable factor, it is set to 1.5 in this embodiment.

[0122] (2) MSST analysis:

[0123] First of all Perform a short-time Fourier transform (STFT) using the Hanning window function. Window length The STFT expression is:

[0124] ;

[0125] Then, synchronous compression is performed in the frequency direction. Through multiple iterations of compression (set to 3 times), the dispersed frequency components are concentrated towards the instantaneous frequency position, resulting in a time-frequency distribution feature matrix with enhanced energy concentration. Its size is (Time point × Frequency point).

[0126] The feature aggregation module performs compression aggregation and dimensionality reduction on time-frequency distribution features.

[0127] (1) Calculation of energy concentration:

[0128] For the time-frequency distribution characteristic matrix For each time window, calculate the root mean square magnitude:

[0129] ;

[0130] in, The window length is 0.1 seconds. This represents the mean value within the window. The calculation results show that the energy concentration in the three time windows is as follows: , , .

[0131] (2) Fisher Score Feature Selection:

[0132] For the time-frequency distribution characteristic matrix Features ( Fisher Score (FSS) assessment was performed on the [number]th [item]. Each feature dimension is defined as follows:

[0133] ;

[0134] in For inter-class dispersion, This represents the within-class dispersion. A threshold is set. Remove The low contribution features reduce the feature space from 5000 dimensions to 800 dimensions.

[0135] (3) Local Mean Decomposition (LMD):

[0136] The reduced-dimensional features are processed using LMD (Low Mode Decomposition) to extract intrinsic mode features. Adjacent extreme points... The local mean is:

[0137] ;

[0138] The standard iterative process of the LMD algorithm is as follows:

[0139] Step 1, Constructing the Local Mean Function: For the signal Find all adjacent maxima. and minimum point Calculate the local mean of adjacent extreme point pairs. and local amplitude :

[0140] ;

[0141] right and The sequences are smoothed by moving average interpolation to obtain continuous local mean functions. and envelope estimation function .

[0142] The second step is pure frequency modulation component separation: Subtract the local mean function from the original signal and divide by the envelope function to obtain the normalized frequency modulation component.

[0143] ;

[0144] The third step is to determine the termination of the iteration: For Repeat the above process until the obtained component satisfies (i.e., pure frequency modulation condition), then the final normalized frequency modulation component is denoted as .

[0145] Step 4, Extraction of intrinsic mode components (product function): Multiply the envelope function by the pure frequency-modulated component to obtain the first intrinsic mode component (product function):

[0146] ;

[0147] Step 5, Residual Signal Iteration: Let ,right Repeat the above steps to extract sequentially. The iteration continues until the residual signal becomes a monotonic function or the number of extreme points is insufficient for the next decomposition. The termination condition is automatically determined by the distribution of extreme points of the signal itself, without the need for manual preset of the number of decomposition layers or penalty factor parameters.

[0148] The complexity calculation module performs singular value decomposition on the intrinsic modal features and calculates the singular spectral entropy.

[0149] (1) Singular Value Decomposition (SVD): The eigenvalue matrix composed of 800-dimensional eigenmode eigenvectors is decomposed into singular values. (size Perform SVD decomposition:

[0150] ;

[0151] in, The diagonal elements are singular values. The calculation yielded: , , , .

[0152] (2) Permutation entropy calculation: Perform permutation entropy calculation on the singular value sequence, with the embedding dimension set to 1. The number of symbol sequence types is The permutation entropy is:

[0153] ;

[0154] The singular spectral entropy was calculated. This value reflects that the complexity of the energy distribution of the operating signal is at a normal level (an empirical threshold range of 0.4 to 0.8 indicates normal operation).

[0155] This value reflects that the complexity of the energy distribution of the operating signal is at a normal level. The above-mentioned normal operating range [0.4, 0.8] is determined as follows: Historical operating data of 300 streetlights over 30 consecutive days in a commercial district of a city are selected as the basic dataset, which includes 2160 sets of normally operating samples, 486 sets of slightly abnormal samples, 312 sets of moderately abnormal samples, and 142 sets of severely abnormal samples, all manually verified and labeled. The time-domain filtering, time-frequency analysis, feature aggregation, and singular spectral entropy calculation process described in this invention are sequentially executed on all samples, and the distribution characteristics of singular spectral entropy for each category are statistically analyzed: the mean singular spectral entropy of the normally operating samples is 0.61, the standard deviation is 0.09, and the 99% confidence interval is [0.38, 0.8]. [0.84]; The mean value of mildly abnormal samples is 0.89, and the overlap rate with the distribution of the normal range is less than 3.2%; According to Fisher's linear discriminant analysis, the optimal decision boundary between normal and abnormal is 0.80. Samples below this value are judged to be operating normally, and samples above this value trigger the abnormal identification process. The accuracy of normal state identification on the above verification dataset is 97.3%, the false negative rate is 1.8%, and the false positive rate is 2.7%, which is statistically significant.

[0156] The mutation identification module uses dilatational and erosion edge detection operators to locate mutation points.

[0157] (1) Morphological feature extraction:

[0158] Singular spectral entropy time series (Contains 100 time points) Morphological analysis was performed using the value, with the scale parameter set to... structural elements Take a rectangular structuring element with a length of 3. The opening and closing morphological spectra are defined as follows:

[0159] ;

[0160] ;

[0161] in, For operation parameters, It is a non-negative function. This indicates the opening operation. This represents the closing operation. The scale is represented as . structural elements.

[0162] (2) Location of mutation points:

[0163] The rising edge abruptness feature was extracted using the dilatational edge detection operator, and the detection was performed. There is a slight increase; however, by extracting abrupt descent edge features using an erosion-type edge detection operator, no significant decrease was detected. Combining these two types of features, it was determined that... There is a slight mutation, but the magnitude of the mutation is small. Less than the preset threshold The fluctuation is determined to be normal and does not trigger the anomaly identification process.

[0164] Since no significant mutation was detected, the classification and recognition module categorized the operating state as "normal operating state". If a mutation is detected, an adaptive boosting (AdaBoost) classifier based on a multilayer perceptron (MLP) is used for classification.

[0165] (1) MLP network structure: Input layer: receives 800-dimensional intrinsic modality features; Hidden layer 1: 256 neurons, activation function is ReLU; Hidden layer 2: 128 neurons, activation function is ReLU; Output layer: 4 neurons, corresponding to 4 operating state categories (normal, mild abnormality, moderate abnormality, severe abnormality), activation function is Softmax.

[0166] (2) AdaBoost Classification: Iterative Training A weak classifier is used, and each iteration is based on the classification error. Adjusting the weights of the weak classifier:

[0167] ;

[0168] In cases with multiple categories, weighted voting is used to determine the final category:

[0169] ;

[0170] In this embodiment, the classification result is "normal operating status" with a confidence level of 0.96.

[0171] The scheduling response module receives the classification result "normal operation status," queries the preset anomaly category-scheduling strategy mapping table, and determines the "maintenance strategy" to execute. The scheduling instructions include: the scheduling time is the current moment. The scheduling target is 100 streetlights in the block, and the scheduling parameter is to maintain the current power. Brightness maintained The dispatch instructions are sent to each street light controller via the ZigBee wireless communication protocol, and the execution time is less than [time missing]. s.

[0172] The feedback update module collects runtime data after scheduling execution. (Data from the last minute) to calculate the prediction error.

[0173] (1) Error calculation: Prediction category "Normal Operation Status", Actual Category "Normal operating status" is correctly categorized; error is acceptable. .

[0174] (2) Weight update: Since the classification is accurate, there is no need to update the connection weights and bias parameters of the MLP network.

[0175] (3) Optimization of fusion weights: The feedback update module feeds back the error information to the data fusion module and updates the fusion weights according to gradient descent:

[0176] ;

[0177] in, The learning rate is set to 0.01. Here is the loss function (cross-entropy loss). Since the error is 0, the fusion weights remain unchanged: , , ..., And satisfy .

[0178] This completes a full identification-scheduling-feedback closed-loop cycle. The system continuously monitors its operational status and executes the above process every 10 seconds.

[0179] Example 2

[0180] This embodiment illustrates how, when an urban street lighting system experiences abnormal operation, the system of this invention achieves a complete closed-loop process of anomaly identification and intelligent scheduling through multi-source data fusion, time-domain filtering, time-frequency analysis, feature aggregation and dimensionality reduction, complexity index calculation, mutation identification, adaptive neural network classification, and scheduling response and feedback updates. Unlike the normal operation scenario in Embodiment 1, this embodiment introduces a typical scenario of "voltage fluctuations causing sudden changes in electrical parameters." This anomaly is characterized by its suddenness, short duration, and relatively stable environment and user demands. If only a single operating quantity (such as power) or only threshold alarms are used, misjudgments, missed judgments, or inaccurate identification of the anomaly range and level are likely to occur, thus affecting the timeliness and effectiveness of the scheduling strategy.

[0181] At 22:00, a regional voltage fluctuation occurred in a commercial district of a city, causing abnormal power fluctuations in some streetlights. The data acquisition module simultaneously collected data from 100 streetlights in the district 30 seconds before and after the anomaly occurred. to The operational data, environmental perception data, and user demand data of the system are sampled at corresponding frequencies. Therefore, each fusion sequence during this abnormal period contains Data points. Taking street light number 50 as an example: Before the anomaly occurred ( )for , , When an abnormality occurs ( ) leap to , , After the anomaly occurred ( (returned to) , , The power is in The instantaneous surge was approximately 47%, with current increasing by about 37% and voltage rising by about 7.6%. Meanwhile, the environmental sensing data remained stable. , , No significant changes were observed; user demand data remained unchanged: target illuminance Lighting demand during certain periods The above comparison shows that the anomaly mainly originated from electrical disturbances, rather than environmental or demand changes, which provides a basis for subsequent feature contribution analysis and adaptive updating of fusion weights.

[0182] The system first performs multi-source data fusion. The fusion module uses the same weighted fusion method as in Example 1, mapping operational data, environmental perception data, and user demand data into a fused signal. For the first... A lamp constructs a fused signal:

[0183] ;

[0184] The initial fusion weights are set to , , , , , , ,satisfy For abnormal periods ( to The fused sequence is obtained after normalizing the original data. .because and exist A significant abrupt change occurred nearby, indicating signal fusion. exist The site exhibits obvious mutation characteristics, which allows the anomaly to be enhanced and presented at the fusion level.

[0185] The process then proceeds to time-domain filtering. The time-domain filtering module uses the same Gammatone filter bank (order) as in Example 1. Number of filters The center frequency is from arrive Target frequency band The discrete impulse response is then normalized. The filtered output is expressed in convolutional form as follows:

[0186] ;

[0187] in This is the filter length. To accommodate the high amplitude characteristics of anomalous signals, the gain control unit adaptively adjusts the gain factor, automatically changing it from 0.2 to 0.3 to enhance the retention of anomalous change features and improve the resolution of subsequent time-frequency analysis. The filtered operating signal is... The abrupt change structure is preserved nearby, while high-frequency noise is effectively suppressed, and the signal-to-noise ratio is improved from 45 dB of the original signal to 68 dB.

[0188] In the time-frequency analysis phase, the system employs a combination of the Cui-Williams distribution (CWD) and multiple synchronous compression transform (MSST) to construct the time-frequency distribution characteristics. The kernel function of CWD in the fuzzy function domain is:

[0189] ;

[0190] in For Doppler frequency variables, These are time delay variables, all defined in the fuzzy function domain, and are related to the signal time variable. They are in different domains and there is no symbolic ambiguity. The periodic controllable factor (set to 1.5 in this embodiment) is adjusted... The size of the kernel function controls the strength of its suppression of cross terms. The larger the value, the stronger the suppression of cross terms, while the time-frequency resolution of the autocorrelation term decreases slightly. This can be adjusted by weighing the number of components and frequency intervals of the actual signal. The corresponding CWD time-frequency distribution is obtained by multiplying the kernel function and the signal ambiguity function and then performing a two-dimensional Fourier transform:

[0191] ;

[0192] in For signal The fuzzy function, where * denotes complex conjugate. , These are the time and frequency coordinates on the time-frequency plane.

[0193] CWD analysis results show Nearby, the frequency components were previously concentrated in approximately The narrow band area spread to The range indicates a significant change in energy distribution. The STFT window length remains constant in MSST analysis. The synchronous compression iterations were kept to a maximum of 3 times, resulting in the time-frequency distribution feature matrix. Size is (Time point × Frequency point), its position A significant energy transition occurs at this point, with the time-frequency energy jumping from 0.3 to 0.8, further characterizing the energy redistribution caused by the anomalous disturbance.

[0194] Next, feature aggregation and dimensionality reduction are performed. The system calculates the energy aggregation degree (root mean square magnitude) over 30 time windows with a 1-second window. : Window before the exception ( ; )for Window in case of an error ( to )for (Growth of 65%), after the anomaly (window) to )for (A 34% increase), the significant change in energy concentration indicates an abnormal operating state. Then, regarding... Fisher Score is applied to each time-frequency feature, with the threshold maintained. Because the anomalous features are more significant and the number of high-contribution features increases, 1200 dimensions of features are retained after dimensionality reduction (800 dimensions in Example 1). In the intrinsic mode feature extraction stage, the system performs LMD processing on the dimensionality-reduced features to obtain 6 intrinsic mode components. Each component contains 200 eigenvalues, forming a 1200-dimensional intrinsic mode feature vector; where and exist A significant amplitude jump occurred nearby, and to The relative stability indicates that the anomalies are mainly concentrated in the low-order dominant modes.

[0195] During the complexity calculation stage, the system calculates the feature matrix formed by the 1200-dimensional intrinsic mode eigenvectors. (size Perform singular value decomposition:

[0196] ;

[0197] Obtain the singular value sequence , , , , , Compared with Example 1 ( , , , Compared to the previous method, all singular values ​​are significantly increased, reflecting a more complex energy distribution under anomalous conditions. Subsequently, permutation entropy is used to calculate the singular spectral entropy:

[0198] ;

[0199] Preserving Embedded Dimensions The six singular values ​​are slide-expanded with an embedding dimension of 3, forming a total of Given a subsequence of length 3, the maximum number of permutation patterns when the embedding dimension is 3 is: The number of arrangement patterns actually observed in this embodiment is: (Under abnormal conditions, the energy distribution of each mode is discrete, and the number of arrangement modes approaches the upper limit), and the singular spectral entropy is calculated. This value is less than The theoretical upper limit is physically achievable and is significantly higher than the normal threshold upper limit of 0.8, indicating that the energy distribution complexity of the running signal is abnormal, triggering the anomaly identification process.

[0200] During the mutation identification phase, the system analyzes the singular spectral entropy time series. Morphological analysis was performed. Unlike Example 1, the entropy sequence in this example contains 30 time points. Value, scale parameter The structuring element is a rectangular structuring element of length 3. The opening and closing morphological spectra are defined as follows:

[0201] ;

[0202] The calculation results show that: Place A significant positive peak appears, with a peak amplitude ;exist Place A secondary positive peak appears, with an amplitude ;exist Place A negative peak appears, with an amplitude Furthermore, the rising edge abrupt change point is located using an dilatational edge detection operator to obtain the main abrupt change point. mutation amplitude (Exceeding the threshold) ); to obtain the secondary mutation point mutation amplitude (Exceeding the threshold). The erosion-type edge detection operator is used to locate the abrupt change point of the falling edge, and the recovery point is obtained. mutation amplitude (Absolute value exceeds threshold). The overall abnormal change range is determined to be... The anomaly type was determined to be "power mutation anomaly".

[0203] In the classification and recognition stage, the system uses an AdaBoost classifier based on MLP to classify abnormal variation intervals. The input is 1200-dimensional intrinsic modality features (800-dimensional in Example 1). The MLP structure is as follows: Hidden layer 1 has 512 neurons, hidden layer 2 has 256 neurons, the activation function is ReLU, and the output layer has 4 neurons corresponding to "normal, mild abnormality, moderate abnormality, and severe abnormality", with the output using Softmax. The network parameters are loaded from the model trained in Example 1 (pre-training complete). Forward propagation is as follows:

[0204] ;

[0205] Obtain the category probability distribution: , , , AdaBoost iteratively trains 10 MLP weak classifiers and performs weighted voting. The weighted voting result is... The final classification was "moderately abnormal" with a confidence level of 0.75.

[0206] During the scheduling response phase, the scheduling response module triggers strategy selection and generates scheduling instructions within 0.5 seconds of detecting an anomaly. The system queries the anomaly category and scheduling strategy mapping table to determine that the "voltage limiting and power reduction strategy" will be executed. The scheduling objects are street light No. 50 and its adjacent street lights No. 48, 49, 51, and 52 (a total of 5 lights), and the scheduling parameter is the target power. Target voltage The scheduling method is PWM dimming, with the duty cycle reduced from 85% to 70%, and the scheduling duration lasting until the voltage returns to normal (estimated 5 seconds). The scheduling command is sent to the corresponding controller via the ZigBee wireless communication protocol, with an execution time of less than 0.3 seconds. The target is achieved and the voltage returns to normal. This indicates that the scheduling strategy effectively suppressed abnormal fluctuations and restored stable operation.

[0207] During the feedback update phase, the system collects operational data after scheduling execution. to ), and in (At the time of scheduling completion) the complete recognition process is re-executed. The new singular spectrum entropy at this point is... It has returned to normal. The new classification result is , , , The actual category is normal operating status. Because in The system determined the state to be moderately abnormal at that time and the scheduling was triggered in a timely manner. The system determines that it has returned to normal, but these two correspond to the actual states at different points in time. Therefore, this embodiment calculates the classification error based on the criterion that "the prediction is consistent with the actual state at that time". Because the classification is accurate, the weights and bias parameters of the MLP network remain unchanged. However, the system adds the abnormal samples (1200-dimensional feature vectors and the label "moderate anomaly") to the training set for subsequent incremental learning, thereby improving the ability to identify similar anomalies.

[0208] Meanwhile, the system adaptively optimizes the fusion weights. Through feature contribution analysis of this abnormal sample, it was found that power... With current The highest contribution was made to anomaly detection (Fisher Score of 0.85 and 0.78 respectively), followed by voltage, while temperature contributed the most. With crowd density The contributions were low (Fisher Scores of 0.05 and 0.08 respectively). The feedback update module uses gradient descent to update the fused weights.

[0209] ;

[0210] Learning rate The updated example is as follows: (Power weighting) is determined by Adjusted to , (Current weighting) is determined by Adjusted to , Adjusted to , Adjusted to , Adjusted to , Adjusted to , Adjusted to If the normalization constraint is not met, normalization processing is required: divide each weight by the sum of the weights. The normalized weights are obtained as follows: , , , , , , ,satisfy The updated fusion weights increase the weight of operational data (power, current, voltage) and decrease the weight of environmental data (temperature, population density), thereby improving the system's sensitivity to sudden power anomalies. Normal and anomaly comparison data are shown in Table 1.

[0211] Table 1. Comparison of Results: ;

[0212] To verify the closed-loop optimization effect, the updated fusion weights were applied to subsequent monitoring periods. to During this period, the system did not detect any new anomalies, and all classification results were in normal operating condition with an average confidence level of 0.92. The verification feedback optimization mechanism improved the system's stability and recognition consistency without introducing additional false alarms.

[0213] This completes a full anomaly identification-scheduling-feedback closed-loop cycle, demonstrating that the system of this invention can achieve anomaly interval location, anomaly level identification, rapid scheduling intervention, and adaptive optimization of fusion weights under abnormal operating conditions, thereby improving the robustness and intelligence level of urban street lighting systems.

[0214] Example 3

[0215] This embodiment illustrates the complete workflow of the system of the present invention when a city street lighting system exhibits progressive aging anomalies. Unlike the scenario in Embodiment 2 where "voltage fluctuations lead to sudden changes in electrical parameters," progressive aging anomalies are characterized by a continuous and slow decline in power, small fluctuation amplitude at a single moment, and a continuous increase in the complexity of the signal energy structure. This scenario is specifically used to verify the necessity of the following three core innovations: First, the ability of the singular spectral entropy composite index to distinguish between "complex small-amplitude anomalies" and "orderly large-amplitude fluctuations"—singular values ​​alone cannot identify this type of anomaly; Second, the collaborative triggering of the dual-path parallel update mechanism of the feedback update module when the classification error is non-zero—neither Embodiment 1 nor Embodiment 2 showed a situation of simultaneous dual-path updates; Third, the cross-scenario adaptive switching of fusion weights between different anomaly types—the data source contribution distribution in this embodiment is significantly different from that in Embodiment 2, and the system should adjust the bias of the fusion weights accordingly.

[0216] At 11:00 PM in a commercial district of a city, the data acquisition module detected a continuous and slow decline in the power of street light number 37. A total of 60 street lights in this area were included in the synchronous monitoring. Sampling frequency. The monitoring time window is 60 seconds, and each fusion sequence contains 6000 data points. The operational data is taken as an example of street light number 37: hour , , ;exist hour , , ;exist hour , , The power decreased by approximately 9.4% cumulatively over 60 seconds, the current decreased by approximately 9.2% proportionally, and the voltage remained stable throughout. Meanwhile, environmental sensing data and user demand data remained stable. , , , The above data indicates that this anomaly is a gradual degradation on the purely electrical side, with no significant disturbance on the voltage side. This is fundamentally different from the anomaly type in Example 2, which is mainly driven by voltage fluctuations. This provides a basis for subsequent feature contribution analysis and cross-scenario switching of fusion weights.

[0217] Data fusion phase. The data fusion module uses the updated fusion weights (rather than the initial weights) from Example 2 for fusion processing, reflecting the system's historical learning capabilities: , , , , , , ,satisfy To construct a fused signal for street light number 37:

[0218] ;

[0219] After normalizing the original data and inputting it, the continuous slow decreasing trend of power and current in the fused signal... The component exhibits a low-slope trend with a monotonically decreasing amplitude of approximately 8.7% within 60 seconds, unlike the component in Example 2. The abrupt change at the point of origin causes the anomaly to exhibit a "gradual" rather than a "mutation" characteristic at the fusion level.

[0220] Time-domain filtering stage. The time-domain filtering module uses the same Gammatone filter bank parameters (order) as in Examples 1 and 2. Number of filters The center frequency is from arrive Target frequency band Since the abnormal signal in this embodiment exhibits a low-frequency gradual change trend and relatively weak high-frequency noise components, the gain control unit adaptively adjusts the gain factor to 0.22. After filtering, the signal-to-noise ratio is increased from 45 dB of the original signal to 66 dB, effectively preserving the low-frequency trend component of the fused signal while suppressing measurement noise interference.

[0221] Time-frequency analysis phase. The time-frequency analysis module uses a combination of CWD and MSST to construct time-frequency distribution characteristics, maintaining a controllable periodicity factor. STFT window length The synchronous compression iteration count is maintained at 3 times. This is similar to Example 2. Unlike the energy jump (time-frequency energy suddenly increases from 0.3 to 0.8) observed at a certain point, the CWD analysis results in this embodiment show that the energy distribution exhibits a continuous and slow migration characteristic within 60 seconds, and the frequency components change from... Time concentrated on approximately In the narrow band region, Gradually spread to The range indicates that aging of the luminaires leads to a progressive increase in the complexity of the operating modal structure. The time-frequency distribution feature matrix after MSST enhancement. Size is (Time point × Frequency point) Its energy center of gravity exhibits a monotonous, gradual shift rather than a single-point jump within the entire window. This is a hallmark feature of the time-frequency domain that distinguishes progressive aging from abrupt changes.

[0222] Feature aggregation and dimensionality reduction stage. Using a 1-second time window, the energy aggregation degree (root mean square amplitude) was calculated for 60 time windows. The energy aggregation degree showed a continuously monotonically decreasing trend, indicating a gradual degradation of the operating state. Fisher Score is applied to each time-frequency feature, with thresholds maintained. After dimensionality reduction, 1000-dimensional features are retained. In the intrinsic mode feature extraction stage, the system performs LMD processing on the dimensionality-reduced features. The LMD iteration termination condition is automatically determined by the distribution of extreme points in the signal itself, eliminating the need for manual pre-setting of the decomposition layer number; this yields 5 intrinsic mode components. Each component contains 200 eigenvalues, forming a 1000-dimensional intrinsic mode feature vector. and The component exhibits a trend of monotonically decreasing amplitude throughout the entire 60-second window. to The multi-frequency modal fluctuation structure that emerges during the aging process was captured, reflecting the essential difference in modal structure between the aging signal and the normal signal.

[0223] Complexity calculation stage and verification of the non-substitutability of singular spectral entropy. The system's feature matrix is ​​formed by 1000-dimensional intrinsic mode eigenvectors. (size Perform singular value decomposition: ;

[0224] Obtain the singular value sequence , , , , .

[0225] Here lies the most crucial verification result of this embodiment: the maximum singular value. Smaller than the normal state in Example 1 If the system only determines anomalies based on the largest singular value, then The comparison results will lead the system to misjudge that this aging anomaly is "more normal" than the normal state, thus resulting in missed detection. This directly proves the assertion that "singular values ​​alone cannot distinguish between ordered large fluctuations and complex small anomalies" - the power reduction caused by lamp aging lowers the overall energy level, and the amplitude of singular values ​​is smaller than that of normal values, but the structural complexity of its modal energy distribution is significantly increased.

[0226] Subsequently, the singular spectral entropy is calculated using permutation entropy: embedding dimension. For 5 singular values ​​with time delay Phase space reconstruction is performed to form a co-form Given a subsequence of length 3, the theoretical upper limit of permutations is 3. Actual observation The permutation entropy formula is as follows: (Aging causes the energy of each intrinsic mode to exhibit an irregular redistribution structure during the decomposition process, and the ratio of adjacent singular values ​​changes, resulting in diverse permutation patterns for different subsequences.)

[0227] ;

[0228] The singular spectral entropy was calculated. This value is higher than the normal threshold of 0.8, triggering the anomaly identification process; at the same time, it is lower than that in Example 2. This indicates that the severity of this anomaly is lower than that of voltage-sudden anomalies, demonstrating the ability of singular spectral entropy to quantify and differentiate anomaly levels. A comprehensive comparison is as follows: based solely on the maximum singular value... The system will miss this anomaly; based solely on permutation entropy (which does not carry energy amplitude information), it is impossible to determine whether the anomaly level requires triggering scheduling; the singular spectrum entropy, composed of both, is insufficient. It not only identifies anomalies (exceeding the threshold of 0.8) but also provides quantified level information (lower than 1.25 in Example 2), proving that composite mapping is a necessary condition for this type of anomaly detection.

[0229] Mutation identification phase. The system analyzes the 60-second singular spectral entropy time series. Perform morphological analysis, scale parameters The structural element is a rectangular structural element of length 3. Similar to Example 2... Unlike the single major mutation point, this embodiment detects multiple low-amplitude asymptotic mutation points: the dilatational edge detection operator in A rising edge abrupt change point was detected at the location, and the abrupt change amplitude was [not specified]. (Above the threshold of 0.15, it is considered normal fluctuation); Detected at the location ;exist Detected at the location (First time exceeding the threshold of 0.15); the erosion-type edge detection operator did not detect a significant falling edge within the 60 s window. Based on the cumulative rising edge characteristics, the system determines... The anomaly was identified as a "gradual power decay anomaly" within a range of continuous abnormal changes, which is clearly distinguishable from the "power mutation anomaly" in Example 2.

[0230] The classification and recognition phase and classification error triggering. The classification and recognition module uses the same MLP-AdaBoost classifier structure as in Examples 1 and 2 (hidden layer 1: 512 neurons, hidden layer 2: 256 neurons, output layer: 4 classes), with 1000-dimensional intrinsic modality features as input. The network parameters are loaded from the updated model in Example 2. Forward propagation yields the class probability distribution: , , , AdaBoost weighted voting results The final classification was normal, with a confidence level of 0.52.

[0231] However, in the feedback verification after scheduling execution (collection) (Data re-identification), new singular spectral entropy The value is still above the threshold of 0.8, and the actual category was confirmed as slightly abnormal after verification (specifically, it was determined to be a stage of slight degradation due to aging of the lamps). This resulted in a non-zero classification error. The predicted category "normal" is inconsistent with the actual category "slightly abnormal". A cross-entropy loss is constructed as follows:

[0232] ;

[0233] This is the first time in the three embodiments of the present invention that a non-zero classification error has occurred, thereby triggering the first path update of the feedback update module.

[0234] Dispatch Response Phase. Within 1.0 seconds of detecting an accumulated anomaly, the dispatch response module triggers a "preventive power reduction protection strategy" (different from the "voltage limiting power reduction strategy" in Implementation Example 2). The dispatch target is street light number 37, and the dispatch parameter is the target power. (Further reducing the current 77W by 9% to slow down the aging process), the scheduling method is PWM dimming, and the duty cycle is reduced from the current 90% to 82%. At the same time, the system generates a maintenance warning work order, requiring on-site inspection of street light No. 37 within 48 hours. The scheduling command is sent to the corresponding controller via the ZigBee wireless communication protocol, and the execution time is less than 1.0s.

[0235] Feedback Update Phase – The first complete triggering of dual-path parallel updates. This embodiment is the only one of the three embodiments that achieves simultaneous dual-path parallel updates, fully verifying the core design of the closed-loop mechanism described in claim 1.

[0236] First path (updating classification decision logic): Because the predicted category "normal" is inconsistent with the actual category "slightly abnormal", the loss function... The backpropagation algorithm calculates the gradient of the loss function with respect to the connection weights and bias parameters of each layer of the MLP, according to the learning rate. Execution parameter update:

[0237] ;

[0238] After the update, under subsequent similar progressive aging mode inputs, the network's output probability for "mild anomaly" increased from 0.31 to 0.58, and the classification confidence was significantly improved, indicating that the system updated its ability to identify progressive aging anomalies through the first path.

[0239] The second approach (updating the perception-focused strategy): The system performs Fisher Score analysis on the contribution of each data source feature to the anomaly sample, and the results are: power Contribution 0.79, current Contribution 0.71, voltage Contribution 0.18 (significantly lower than in Example 2, as there was no disturbance on the abnormal voltage side this time), ambient temperature Contribution 0.22 (significantly higher than 0.05 in Example 2, due to accelerated aging in high-temperature environments), light intensity Contribution 0.11, Crowd density Contribution score 0.06, user demand Contribution: 0.08.

[0240] Based on the above contribution distribution, the feedback update module performs gradient descent updates on the fusion weights, starting from the weights updated in Example 2: , , , , , , :

[0241] ;

[0242] The updated result is: The power weighting has been adjusted from 0.268 to 0.276. The current weighting has been adjusted from 0.216 to 0.223. (Voltage weight) was reduced from 0.150 to 0.144 (compared to Example 2). The upward direction is opposite, reflecting dynamic switching across scenarios. The value was adjusted from 0.142 to 0.141. The temperature weighting was increased from 0.044 to 0.048 (reflecting the correlation between temperature and aging). The value was adjusted from 0.088 to 0.085. The value was adjusted from 0.092 to 0.089. The sum of the weights after the update is... Normalization is performed by dividing each weight by 1.006 to obtain the normalized weights: , , , , , , ,satisfy .

[0243] Verification of the two-path coupling mechanism: The type of anomaly (progressive aging) determines the voltage. Low contribution and temperature The distribution of contributions with relatively high contribution levels drives The voltage weighting decreased from 0.150 at the end of Example 2 to 0.143, while in Example 2... The weight increased from an initial 0.15 to 0.156 (due to the high contribution of voltage fluctuations). This reverse adjustment of weights across embodiments directly demonstrates that the system can "focus" on the most informative sensor based on different anomaly types, and the fusion weights evolve from fixed parameters to dynamic parameters that reflect the characteristics of the current anomaly environment.

[0244] A comprehensive comparison of the three embodiments is shown in Table 2:

[0245] Table 2: Comprehensive Comparison of Examples: ;

[0246] Thus, the three embodiments together constitute a complete innovative verification system: Embodiment 1 establishes a normal benchmark, Embodiment 2 verifies the rapid response to mutation anomalies, and Embodiment 3 specifically verifies the irreplaceability of the singular spectral entropy composite index and the complete triggering mechanism of dual-path parallel updates. The three embodiments cover all the key scenarios of the core innovation of this invention.

[0247] To verify the technical superiority of the dual-path parallel closed-loop mechanism over the single-path update scheme, a systematic comparative experiment was conducted on the same urban street lighting dataset (same as the aforementioned threshold calibration dataset, with a total of 3100 samples, divided into training and test sets in a 7:3 ratio):

[0248] Option A (first path only, only update classification model weights): Keep the fusion weights of the data fusion module fixed (using the initial values), and only perform backpropagation updates on the MLP-AdaBoost network parameters of the classification and recognition module.

[0249] Option B (Second path only, only update fusion weights): Keep the network parameters of the classification and recognition module unchanged (using pre-trained values), and only update the fusion weights of the data fusion module by gradient descent based on the Fisher Score contribution.

[0250] Option C (Dual-path parallel update of the present invention): Two paths run simultaneously and evolve collaboratively.

[0251] The experiment was conducted under a scenario simulating the drift of urban lighting operation characteristics. Specifically, an electrical parameter distribution shift was introduced in the first 300 cycles (simulating baseline drift caused by equipment aging), the system returned to normal in cycles 301 to 600, and a sensor type switch was introduced in cycles 601 to 900 (simulating the replacement of pedestrian flow sensors with millimeter-wave radar, resulting in changes in the measurement characteristic distribution). The experimental results are shown in Table 3.

[0252] Table 3. Comparison of experimental results for dual-path parallel update mechanisms: ;

[0253] The experimental results above show that: Scheme A suffers from low efficiency in utilizing new sensor information and a significant decrease in recognition accuracy during the sensor switching phase because the fusion weights cannot be adaptively adjusted; Scheme B suffers from insufficient generalization ability for new abnormal patterns during the feature drift phase because the classification model parameters cannot be updated; only the dual-path parallel update scheme C of this invention can maintain the highest recognition accuracy under both typical operating environment changes, verifying that the co-evolution of the two paths is a necessary condition for long-term stable operation, and that implementing only one path cannot achieve the same effect, proving that the dual-path parallel closed-loop mechanism of this invention has substantial technical progress.

Claims

1. A smart urban street lighting dispatching system with dual-path closed-loop adaptive optimization, characterized in that, include: The data acquisition module is used to collect operational data, environmental perception data, and user demand data of urban street lighting equipment; The data fusion module is connected to the data acquisition module and performs weighted fusion processing on the operation data, environmental perception data and user demand data to construct a multi-source fusion signal that reflects the comprehensive characteristics of the lighting equipment's operating status, external environment changes and user demand. The time-domain filtering module is connected to the data fusion module, receives the multi-source fused signal, performs time-domain filtering processing using a non-uniform frequency resolution filter bank to suppress high-frequency noise interference, and outputs the filtered running signal to the time-frequency analysis module. The time-frequency analysis module, connected to the time-domain filtering module, uses a time-frequency transformation method that combines suppressing cross-term interference with enhancing time-frequency energy concentration to perform time-frequency transformation on the filtered running signal and construct time-frequency distribution characteristics. The feature aggregation module is connected to the time-frequency analysis module. It performs compression and aggregation processing on the time-frequency distribution features, and performs dimensionality reduction processing on the aggregated time-frequency features by a feature selection algorithm based on the inter-class-intra-class dispersion ratio to extract intrinsic mode features. The complexity calculation module is connected to the feature aggregation module. It performs singular value decomposition on the feature matrix composed of the intrinsic modal features to obtain a singular value sequence, and calculates the singular spectral entropy index by applying the permutation entropy algorithm to the singular value sequence. The mutation identification module, connected to the complexity calculation module, uses a morphological method combining dilatational edge detection operators and erosion edge detection operators to locate mutation points in the singular spectral entropy index sequence and identify abnormal change intervals. The classification and recognition module is connected to the mutation recognition module. It uses an adaptive augmentation classifier based on a multilayer perceptron to classify the abnormal change range and output the abnormal operating state category. The scheduling response module is connected to the classification and identification module. It triggers the lighting scheduling strategy according to the abnormal operating state category, generates scheduling instructions, and outputs the scheduling execution information to the feedback update module. The feedback update module is connected to the scheduling response module and the classification and identification module to collect the running data after scheduling execution and realize dual-path parallel update.

2. The intelligent urban street lighting dispatching system with dual-path closed-loop adaptive optimization as described in claim 1, characterized in that, The non-uniform frequency resolution filter bank is a Gammatone filter bank. The Gammatone filter bank includes multiple Gammatone filters with different center frequencies, and the impulse response function of each filter is determined according to the filter order, bandwidth parameter and center frequency parameter. The center frequency parameter is obtained by dividing the target frequency range into equal intervals on an equivalent rectangular bandwidth scale, so that the frequency resolution of the frequency band where the low-frequency harmonic components are located is higher than that of the high-frequency band. Compared to uniform bandwidth filter banks, the Gammatone filter bank improves the frequency resolution of low-frequency harmonic components on an equivalent rectangular bandwidth scale.

3. The intelligent urban street lighting dispatching system with dual-path closed-loop adaptive optimization as described in claim 1, characterized in that, The time-domain filtering module also includes a gain control unit, which is used to normalize the amplitude of the filtered signal. The gain control unit adaptively adjusts the gain factor to determine the optimal gain factor value based on the signal-to-noise ratio of the filtered signal.

4. The intelligent urban street lighting dispatching system with dual-path closed-loop adaptive optimization as described in claim 1, characterized in that, The time-frequency transformation method that combines suppressing cross-term interference with enhancing time-frequency energy concentration is specifically a serial combination of the Cui-Williams distribution analysis method and the multiple synchronous compression transformation analysis method. The described Twee-Williams distribution analysis method performs time-frequency transformation on the running signal based on a kernel function. By adjusting the bandwidth of the kernel function through a periodic controllable factor, it addresses the interference problem of spurious frequency components in multi-component signals and suppresses cross terms in the time-frequency distribution. The multi-synchronous compression transform analysis method uses short-time Fourier transform combined with window function to segment the running signal and performs iterative synchronous compression in the frequency direction. This method addresses the problem of insufficient frequency resolution leading to blurred frequency components and enhances the concentration of time-frequency energy. The described Twee-Williams distribution and multiple synchronous compression transform address two independent defects in time-frequency analysis: interference from spurious frequency components and insufficient time-frequency energy concentration.

5. A smart urban street lighting dispatching system with dual-path closed-loop adaptive optimization as described in claim 1, characterized in that, The extraction of the intrinsic modal features is achieved through a local mean decomposition algorithm; The local mean decomposition algorithm obtains the intrinsic mode components by iteratively separating the local mean component and the pure frequency modulation component in the running signal. The iteration termination condition of the local mean decomposition algorithm is automatically determined by the distribution of extreme points of the signal itself, without the need for manual preset of the number of decomposition layers or penalty factor parameters. It has an adaptive advantage for the characteristics of diverse types of abnormal urban lighting signals and unknown signal morphology in advance. The feature selection algorithm based on the inter-class to intra-class dispersion ratio is the Fisher Score algorithm. It evaluates the contribution of each time-frequency feature to the classification of abnormal states by calculating the ratio of inter-class dispersion to intra-class dispersion, and removes features whose contribution is lower than a preset threshold.

6. The intelligent urban street lighting dispatching system with dual-path closed-loop adaptive optimization as described in claim 1, characterized in that, The process by which the complexity calculation module calculates the singular spectral entropy is as follows: Singular value decomposition is performed on the feature matrix composed of intrinsic mode features to obtain a singular value sequence. The amplitude of the singular value sequence carries the energy distribution information of each modal component of the running signal. The singular value sequence is expanded into a symbol sequence according to a set embedding dimension, the probability of occurrence of each permutation pattern is counted, and the entropy value is calculated according to the permutation entropy formula. The entropy value carries the complexity information of the operating signal mode structure. The singular spectral entropy, through the above steps, compresses and maps the high-dimensional time-frequency feature matrix into a single quantizable scalar that simultaneously contains energy distribution information and structural complexity information, thus establishing a clear quantization boundary between the normal operating state and the structurally abnormal state of the lighting system.

7. A smart urban street lighting dispatching system with dual-path closed-loop adaptive optimization as described in claim 1, characterized in that, The mutation point is located using a combination of dilatational edge detection operators and erosional edge detection operators; The dilatational edge detection operator extracts rising edge abrupt features based on the signal opening operation morphological spectrum. The erosion-type edge detection operator extracts falling edge abrupt features based on the signal closing operation morphological spectrum; The location, direction, and start and end times of the abnormal change range of the operating state change point are determined by the rising edge change feature and the falling edge change feature.

8. A smart urban street lighting dispatching system with dual-path closed-loop adaptive optimization as described in claim 1, characterized in that, The adaptive boosting classifier builds a strong classifier by iteratively training multiple multilayer perceptron weak classifiers and adjusting the weights of training samples based on the classification error in each iteration, increasing the weights of misclassified samples. The multilayer perceptron includes an input layer, at least one hidden layer, and an output layer. The output layer obtains a classification result by weighted voting on the outputs of multiple weak classifiers. The classification result includes level information of normal, mildly abnormal, moderately abnormal, and severely abnormal, so as to support the scheduling response module in generating differentiated scheduling schemes corresponding to the abnormality level.

9. A smart urban street lighting dispatching system with dual-path closed-loop adaptive optimization as described in claim 1, characterized in that, The dual-path parallel update mechanism of the feedback update module is as follows: First path: Construct a cross-entropy loss function by calculating the error between the predicted category of the classification and recognition module and the actual running state after scheduling execution; use the backpropagation algorithm to calculate the gradient of the loss function with respect to the connection weights and bias parameters of the adaptive augmentation classifier, and update the connection weights and bias parameters according to the preset learning rate to update the system's judgment decision logic for anomaly types; The second approach is to update the fusion weights of the data fusion module by gradient descent based on the Fisher Score contribution of each data source feature in the current abnormal sample, thereby increasing the perception weights of high-contribution data sources and updating the system's perception emphasis strategy for different sensor information; and to normalize the updated fusion weights so that the sum of the fusion weights is always equal to 1. The coupling relationship between the two paths is as follows: the current anomaly type determines the contribution distribution of each data source, which in turn drives the fusion weight to tilt towards the data source with high contribution, so that the system can automatically focus on the sensor with the most information in different operating environments, and realize the collaborative adaptive evolution of the identification strategy and the perception strategy.

10. A smart urban street lighting dispatching system with dual-path closed-loop adaptive optimization as described in claim 1, characterized in that, The operational data includes power data, current data, and voltage data of the lighting equipment; the environmental perception data includes light intensity data, temperature data, and pedestrian density data; the user demand data includes lighting brightness demand data and lighting time period demand data. The data fusion module maps the aforementioned multi-source data into a multi-source fusion signal at a unified time scale using a weighted fusion method. The fusion weights are initialized and allocated based on the reliability and importance of different data sources, and are dynamically updated by the feedback update module based on the contribution of each data source's characteristics.

Images