Prediction method for monitoring residual life of integrated solar street lamp

Abstract

The invention discloses a prediction method for monitoring the residual life of an integrated solar street lamp, and relates to the technical field of data processing and artificial intelligence, and the method comprises the steps: S1, collecting state parameters in real time through a sensor network installed on each part of the solar street lamp; s2, extracting an environment periodic component of the illumination parameter by using Fourier series, and performing decoupling normalization processing on the environment periodic component to obtain a normalized parameter; s3, calculating local curvature values of relevant parameters of the battery by using a discrete Laplacian operator to form an enhanced feature vector; s4, in combination with a time domain convolutional network with a gating update mechanism and a dynamic expansion factor, training a prediction model of the residual life; and S5, obtaining the residual life of the solar street lamp based on the residual life prediction model. The method effectively improves the prediction precision of the residual life of the solar street lamp.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence-based remaining life prediction technology, and in particular to a prediction method for monitoring the remaining life of integrated solar streetlights. Background Technology

[0002] With the rapid development of smart cities and new energy technologies, solar streetlights are widely used in urban and rural road lighting due to their advantages such as being environmentally friendly, flexible in deployment, and low in operating costs. Integrated solar streetlights combine photovoltaic modules, energy storage batteries, LED light sources, and intelligent controllers, resulting in a compact system structure and large deployment scale. However, affected by multiple factors such as environmental changes, component aging, and fluctuations in operating load, their performance gradually degrades over long-term operation, eventually leading to reduced lighting capacity or system failure, posing potential threats to urban management and public safety. Currently, common maintenance methods rely mainly on periodic manual inspections and passive fault response, lacking an efficient and accurate lifespan assessment mechanism. This increases maintenance costs and makes it difficult to detect potential faults in a timely manner.

[0003] Furthermore, traditional lifetime prediction methods often rely on simple capacity decay models of single components (such as batteries), neglecting the complex collaborative degradation behavior among components in solar streetlights. At the same time, existing algorithms generally adopt a uniform normalization strategy, failing to distinguish between the environmental periodic fluctuations of illumination parameters and the performance degradation of the equipment itself, resulting in model learning bias and limited prediction accuracy. Summary of the Invention

[0004] The purpose of this application is to provide a prediction method for monitoring the remaining lifespan of integrated solar streetlights, which solves the problems of traditional models ignoring diurnal / seasonal fluctuations in illumination parameters, poor sensitivity near capacity mutation points, limitations on detailed modeling of the aging process, and weak model interpretability.

[0005] To achieve the above objectives, this application provides the following solution:

[0006] In a first aspect, this application provides a prediction method for monitoring the remaining lifespan of an integrated solar street light, the method comprising:

[0007] S1. Status parameters are collected in real time through a sensor network installed on various components of the solar street light;

[0008] S2. Use Fourier series to extract the environmental periodic components of the illumination parameters, and decouple and normalize the environmental periodic components to obtain normalized parameters;

[0009] S3. Use the discrete Laplacian operator to calculate the local curvature values ​​of battery-related parameters, then perform a sign-preserving exponential transformation on the curvature values ​​to generate enhanced feature terms, and concatenate them with the normalized feature vector to form an enhanced feature vector;

[0010] S4. Combine a temporal convolutional network with a gated update mechanism and a dynamic dilation factor to adaptively fuse multi-rate sensor data and train a prediction model for remaining lifetime.

[0011] S5. Obtain the remaining lifespan of solar streetlights based on the prediction model of remaining lifespan.

[0012] Optionally, the normalization parameters obtained in S2 specifically include:

[0013] S201. Use a third-order Fourier series to extract the mean of the environmental periodic components of illumination parameters in historical data;

[0014] S202. Calculate the deperiodic standard deviation based on the mean of the environmental periodic components; perform normalization on the original illumination-related parameters based on the deperiodic standard deviation, and perform standard Z-score normalization on the non-illumination-related parameters to obtain the normalized parameters.

[0015] Optionally, in step S3, concatenating the normalized feature vector with the enhanced feature vector to form an enhanced feature vector further includes:

[0016] The calculated curvature values ​​are subjected to a sign-preserving exponential transform to generate enhanced feature terms. These enhanced feature terms are then concatenated with the normalized feature vector to form an enhanced feature vector, expressed as:

[0017] ;

[0018] In the formula, Represents the enhanced feature vector; Represents the normalized eigenvector; It is a natural exponential function; This represents a vector concatenation operation; The curvature sensitivity coefficient controls the exponential decay rate. This represents the absolute value of curvature calculated at the current time point; This represents the curvature value calculated at the current point in time. For a symbolic function, for Item, when Output when >0 ,when Output when <0 .

[0019] Optionally, training the prediction model for remaining lifespan in S4 further includes:

[0020] S401. A primary prediction model for remaining lifetime is constructed by combining a gated temporal convolutional network and the battery health index value;

[0021] S402. Combine feature optimization, loss improvement, sample expansion, multi-task prediction, and parameter update to train the primary prediction model and generate a prediction model for remaining lifetime.

[0022] Optionally, the preliminary prediction model for remaining lifetime constructed in S401 further includes:

[0023] S4011. Based on the enhanced feature vector at the current time and the hidden state vector at the previous time, a gating vector is calculated through causal convolution. Then, the gating vector is used to control the generation of candidate state vectors and the update ratio of the hidden state vector at the current time, so as to perform gating transmission of temporal information.

[0024] S4012. Dynamically adjust the dilation coefficient of the convolution kernel in the gated temporal convolutional network based on the value of the battery health index.

[0025] Optionally, the prediction model for the remaining lifetime generated in S402 further includes:

[0026] S4021. Based on the defined gated temporal convolutional network, parallel processing of enhanced feature vectors is used to form a fused residual feature matrix;

[0027] S4022. Generate dynamic weights corresponding to the actual remaining lifetime ratio based on the constructed exponential dynamic weight function, and calculate the weighted mean square error loss based on the adjusted dynamic weights;

[0028] S4023. Generate samples that conform to the physical degradation law from the feature vector of real fault samples;

[0029] S4024. After fusing the residual feature matrix, two fully connected layers are connected in parallel as the prediction head to form the total loss function;

[0030] S4025. Perform backpropagation based on the total loss function calculation results until the preset number of iterations or the loss convergence threshold is reached.

[0031] Optionally, the calculation of the weighted mean square error loss based on the adjusted dynamic weights in step S4022 further includes:

[0032] Construct an exponential dynamic weighting function based on the proportion of actual remaining lifespan, and set a lifespan threshold based on the proportion of actual remaining lifespan.

[0033] The dynamic weighting function is applied to the calculation of the weighted mean squared error loss, and the sensitive loss function is obtained by using the L2 regularization term to constrain all trainable parameters of the model.

[0034] Optionally, generating samples that conform to the laws of physical degradation in step S4023 further includes:

[0035] The gradient direction of the loss function is calculated at the feature vector of the real fault sample to generate the perturbation vector;

[0036] The perturbation vector is generated using the calculated gradient direction. Then, a random binary mask vector is applied to constrain the range of the perturbation vector. Finally, the constrained perturbation vector is superimposed on the enhanced feature vector to generate a critical state sample that conforms to the physical degradation law.

[0037] Optionally, forming the total loss function in S4024 further includes:

[0038] After fusing the residual feature matrix, two fully connected layers are connected in parallel as prediction heads. One prediction head is the main task, outputting the predicted value of the remaining lifetime ratio of the entire system, and the other prediction head is the auxiliary task, outputting the health status probability vector of each component.

[0039] Pseudo-labels for component health status are generated based on predefined physical threshold rules. Then, the cross-entropy loss of the auxiliary task is calculated, and the cross-entropy loss is added to the sensitivity loss function of the main task according to the weight coefficient to form the total loss function.

[0040] Optionally, obtaining the remaining lifespan of the solar street light using a prediction model based on the remaining lifespan in step S5 further includes:

[0041] S501. Perform state parameter normalization operation on the original state parameters of the current detection time node to obtain the normalized parameters;

[0042] S502. Perform data augmentation on the normalized parameters to obtain the augmented feature vector;

[0043] S503. Input the enhanced feature vector into the trained gated temporal convolutional network to output the overall remaining lifespan ratio prediction of the system and the health status probability of each component of the solar street light.

[0044] S504. When the remaining lifespan ratio is lower than a preset threshold, an early warning is triggered, and potential faulty components are located based on the component health probability.

[0045] According to the specific embodiments provided in this application, this application provides a prediction method for monitoring the remaining lifespan of integrated solar streetlights, which has the following technical effects:

[0046] 1. This application utilizes a third-order Fourier series to extract the periodic components of illumination-related parameters and performs deperiodic normalization, thereby solving the problem of interference from diurnal / seasonal illumination fluctuations on the training of lifetime prediction models and improving the accuracy of identifying equipment aging trends.

[0047] 2. This application uses the discrete Laplace operator to calculate local curvature and combines it with sign-preserving exponential transformation to generate enhanced features, effectively strengthening the model's ability to perceive nonlinear inflection points (such as the battery's accelerated decay stage).

[0048] 3. This application employs a temporal convolutional network with a gated update mechanism and a dynamic dilation factor to achieve adaptive fusion of multi-rate sensor data and enhance the model's ability to model long-term dependency patterns in the later stages of decay.

[0049] 4. This application predicts system lifetime through the main task and component health status through the auxiliary task, and enhances the model's ability to identify and interpret critical failure states by using critical state samples generated based on gradient perturbation and mask constraints. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a schematic flowchart of a prediction method for monitoring the remaining lifespan of an integrated solar street light according to an embodiment of this application.

[0052] Figure 2 The minimum and maximum normalized color scale heatmap provided in the embodiments of this application;

[0053] Figure 3 Z-score normalized colorimetric heatmap provided in the embodiments of this application;

[0054] Figure 4 This application provides a decoupled normalized colorimetric heatmap for embodiments of the present application.

[0055] Figure 5 This is a schematic diagram of the battery capacity decay process provided in an embodiment of this application;

[0056] Figure 6 This is a diagram showing the curvature analysis of the attenuation path provided in an embodiment of this application.

[0057] Figure 7 Curvature enhancement feature extraction map provided in the embodiments of this application;

[0058] Figure 8 This is a schematic diagram illustrating the effect of the dilation coefficient on prediction accuracy in a gated temporal convolutional network provided in an embodiment of this application. Detailed Implementation

[0059] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0060] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0061] like Figure 1 As shown in the embodiment of this application, a prediction method for monitoring the remaining lifespan of an integrated solar street light is provided, the method comprising:

[0062] S1. Status parameters are collected in real time through a sensor network installed on various components of the solar street light;

[0063] Specifically, status parameters are collected in real time through a sensor network installed on various components of the solar street light. Data sources include the photovoltaic panel monitoring module, battery management system, environmental sensing unit, and controller logs. The collected parameters are divided into two categories:

[0064] Sunlight-related parameters: variables reflecting the intensity of solar radiation, including at least the output voltage of the photovoltaic panel (unit: volts), charging current (unit: amperes), irradiance of the photovoltaic panel surface (unit: watts per square meter), and duration of sunlight (unit: hours). These parameters are obtained through the built-in voltage / current sensor of the photovoltaic panel and an external light intensity meter, and are significantly affected by the day-night and seasonal cycles.

[0065] In addition to the above, the light-related parameters may also include the open-circuit voltage (unit: volts) measured by the voltage sensor under no-load conditions of the photovoltaic panel, the short-circuit current (unit: amperes) recorded by the short-circuit current sensor of the photovoltaic panel, and the array operating temperature (unit: degrees Celsius) collected by the back panel temperature probe of the photovoltaic panel.

[0066] Non-light-related parameters: Variables related to the environment or the device's own state, including battery temperature (degrees Celsius), ambient humidity (percentage), ambient temperature (degrees Celsius), battery internal resistance (milliohms), LED drive current (milliamperes), and controller operating voltage (volts). These parameters are collected by the battery temperature probe, ambient temperature and humidity sensor, and the controller circuit board's ADC module.

[0067] In addition to the above, non-light-related parameters may also include the number of battery charge-discharge cycles (unit: times) obtained from the historical logs of the battery management system (BMS), the ambient wind speed (unit: meters per second) collected by the wind speed sensor at the top of the light pole, the LED junction temperature (unit: degrees Celsius) indirectly calculated by the LED driver thermocouple, and the controller charge-discharge status flag (unit: Boolean value) collected by the controller's internal status register.

[0068] Furthermore, data annotation is performed. The actual remaining lifespan ratio is calibrated through laboratory accelerated aging tests and on-site historical data. Specifically, the actual battery capacity decay rate is used as the core indicator, combined with the LED luminous flux decay rate and controller fault records, to calculate the overall remaining lifespan ratio of the system.

[0069] The specific labeling method is to periodically disassemble street light samples from the same batch for capacity testing, and record the ratio (unit: percentage) of the current maximum discharge capacity to the initial rated capacity, which is used as the label value for supervised training.

[0070] S2. Use Fourier series to extract the environmental periodic components of the illumination parameters, and decouple and normalize the environmental periodic components to obtain normalized parameters;

[0071] S3. Use the discrete Laplacian operator to calculate the local curvature values ​​of battery-related parameters, then perform a sign-preserving exponential transformation on the curvature values ​​to generate enhanced feature terms, and concatenate them with the normalized feature vector to form an enhanced feature vector;

[0072] S4. Combine a temporal convolutional network with a gated update mechanism and a dynamic dilation factor to adaptively fuse multi-rate sensor data and train a prediction model for remaining lifetime.

[0073] S5. Obtain the remaining lifespan of solar streetlights based on the prediction model of remaining lifespan.

[0074] Optionally, the normalization parameter is obtained in S2;

[0075] Specifically, the state parameters of solar streetlights include light-related variables such as photovoltaic panel output voltage and charging current, as well as non-light-related variables such as battery temperature. Conventional Min-Max normalization methods can confuse periodic light fluctuations with equipment degradation trends, causing the model to be unable to distinguish between natural light fluctuations and real aging signals, masking key degradation characteristics, failing to consider the periodic changes in light variables, and being unable to separate fluctuations caused by environmental factors from the equipment's own aging signals.

[0076] This invention utilizes Fourier series to extract the environmental periodic components of illumination parameters, performs deperiodic normalization on illumination-related parameters to eliminate the influence of periodic fluctuations, and uses standard Z-score normalization for non-illumination parameters, thereby achieving decoupling of illumination and non-illumination parameters and enhancing the model's ability to capture real aging signals of equipment.

[0077] Among them, S201. Use the third-order Fourier series to extract the mean of the environmental periodic components of the illumination parameters in the historical data;

[0078] Based on historical data, the mean of the environmental periodic components of the illumination parameters is extracted using a third-order Fourier series, separating the periodic fluctuations in the illumination parameters caused by environmental factors such as the day-night cycle, as follows:

[0079] ;

[0080] In the formula, Indicates the first The mean of the environmental periodic components of the dimension parameter; For parameter dimension index, superscript Identify specific parameter dimensions; It serves as a time point index and simultaneously controls the phase of the periodic function; The order of the Fourier series is denoted by 1, 2, and 3, which correspond to the fundamental and harmonic components. These are the Fourier cosine coefficients, corresponding to the order of... The cosine basis function components characterize the similarity to the cosine function in the periodic fluctuations of illumination parameters. For fundamental frequency, Harmonics; These are the Fourier sine coefficients, corresponding to the order. The sinusoidal basis function components characterize the similarity to the sinusoidal function in the periodic fluctuations of illumination parameters. For fundamental frequency, Harmonics; Pi; It is a cosine function; It is a sine function.

[0081] It should be noted that the time point index Corresponding to the Samples at each time point, indexed chronologically, such as... , This represents the total number of time steps. Each sample contains multi-parameter data at that time point, and there is a one-to-one correspondence between time points and samples. For all T samples, each sample has a dimension of D, where D represents the total number of parameters collected.

[0082] It should also be noted that, The diurnal / seasonal fluctuations are explicitly modeled using a third-order Fourier series. Item and The fundamental and harmonic frequencies of the term directly correspond to the actual periodic characteristics of the illumination parameters, rather than mechanically using a fixed time window for averaging.

[0083] It should also be noted that the Fourier cosine coefficients and Fourier sine coefficients It is automatically determined by fitting historical illumination data. For each illumination parameter, the historical time series data, such as hourly sampling values ​​over multiple consecutive days, is used to solve the third-order Fourier series, including the fundamental frequency, using the least squares method. Harmonic waves This minimizes the error between the series curve and the original data, thereby obtaining the control over the intensity of the cosine component. and control of the intensity of the sinusoidal component .

[0084] S202. Calculate the deperiodic standard deviation based on the mean of the environmental periodic components; perform normalization on the original illumination-related parameters based on the deperiodic standard deviation, and perform standard Z-score normalization on the non-illumination-related parameters to obtain the normalized parameters.

[0085] Specifically, the deperiodic standard deviation is calculated using the mean of the extracted environmental periodic components. Then, normalization is performed on the original illumination-related parameters based on the deperiodic standard deviation to eliminate the influence of periodic illumination fluctuations. Simultaneously, standard Z-score normalization is performed on non-illumination-related parameters, as shown below:

[0086] ;

[0087] ;

[0088] In the formula, Indicates the first The periodic standard deviation of the dimensional parameter; For the first The original value of the dimension parameter refers to the current sample; For the first Dimensional parameters in the first dimension The original values ​​at each time point; To distinguish from Time point index; For the first Dimensional parameters in the first dimension Mean values ​​of environmental cycle components at each time point; For the first Mean value of the environmental periodic component of the dimension parameter; This represents the total number of historical data samples. For the first The normalized value of the dimension parameter refers to the current sample.

[0089] It should be noted that when For non-light-related parameters, the standard Z-score normalization method is directly used for normalization. Light-related parameters include at least the photovoltaic panel output voltage. Charging current Non-light-related parameters include at least battery temperature. Ambient humidity wait.

[0090] In one embodiment, a comparison of feature separation using normalization methods is performed to verify the effectiveness of the decoupling normalization method proposed in this invention in feature separation, such as... Figure 2-4 As shown, by comparing with the traditional minimum-maximum normalization ( Figure 2 ), Standard Z-score normalization ( Figure 3 ) and the decoupling normalization method of the present invention ( Figure 4 The experiment focused on examining the correlation characteristics among six key parameters: photovoltaic voltage, charging current, surface irradiance (light-related parameters), and battery temperature, ambient humidity, and battery internal resistance (non-light-related parameters). The heatmap visually displays the correlation coefficients using color gradients, with red indicating high correlation and blue indicating low correlation. Experimental results show that traditional normalization methods exhibit a significant moderate correlation between light and non-light parameters, with correlation coefficients ranging from approximately 0.2 to 0.4. This indicates that these methods fail to effectively distinguish between environmental fluctuation signals and equipment aging signals. In contrast, the heatmap of the method presented by this invention exhibits a significantly different characteristic distribution. High correlation is maintained within the light parameter group (red area), and high correlation is also maintained within the non-light parameter group, but the correlation coefficient between the two groups significantly decreases to approximately 0.05-0.10, showing a distinct blue area. The feature separation effect demonstrates that the Fourier deperiodic normalization of this invention effectively decouples periodic environmental fluctuations from equipment aging trends, laying the foundation for subsequent models to accurately capture true degradation characteristics. In the figure, the horizontal and vertical axes represent parameter names and are dimensionless; the color scales represent correlation coefficients, ranging from 0 to 1.

[0091] Optionally, the calculation of the local curvature values ​​of battery-related parameters using the discrete Laplacian operator in step S3 further includes:

[0092] There are nonlinear inflection points in the process of battery capacity decay, such as the accelerated decay of the battery health index in the 80% to 70% stage. Conventional polynomial feature extension methods cannot accurately capture such abrupt changes, resulting in insufficient sensitivity of the model to key decay inflection points, inability to effectively identify curvature abrupt changes in the decay path, and weak response to nonlinear inflection points.

[0093] This invention uses the discrete Laplacian operator to calculate the local curvature values ​​of battery-related parameters, then performs a sign-preserving exponential transform on the curvature values ​​to generate enhanced feature terms, which are then concatenated with the normalized feature vector to form an enhanced feature vector. This strengthens the model's sensitivity to low-curvature smooth decay regions and curvature abrupt change points, and improves its ability to capture nonlinear inflection points. The specific steps are as follows:

[0094] 1) Battery parameter curvature calculation:

[0095] For the battery-related normalized parameters, the local curvature value at each time point is calculated using the discrete Laplacian operator to characterize the concavity and convexity of the battery degradation path within the local time window, expressed as:

[0096] ;

[0097] In the formula, This represents the curvature value calculated at the current time point; a positive value indicates convex decay, and a negative value indicates concave decay. It is the discrete Laplace operator, used to approximate the calculation of the second derivative; Indicates the first Battery-related normalized parameters at various time points; superscript Identify battery-related parameters, including at least the battery voltage. Battery internal resistance Battery temperature wait; Indicates the first The battery-related normalized parameters at each time point represent the parameters of the previous time point at the current time point. Indicates the first The battery-related normalized parameters at each time point represent the parameters at the next time point after the current time point.

[0098] Optionally, in step S3, concatenating the normalized feature vector with the enhanced feature vector to form an enhanced feature vector further includes:

[0099] The calculated curvature values ​​are subjected to a sign-preserving exponential transform to generate enhanced feature terms. These enhanced feature terms are then concatenated with the normalized feature vector to form an enhanced feature vector, which strengthens the model's sensitivity to low-curvature smooth decay regions and curvature abrupt change points, as shown below:

[0100] ;

[0101] In the formula, Represents the enhanced feature vector; Represents the normalized eigenvector; It is a natural exponential function; This represents a vector concatenation operation; The curvature sensitivity coefficient controls the exponential decay rate, such as... ; This represents the absolute value of curvature calculated at the current time point; For a symbolic function, for Item, when Output when >0 ,when Output when <0 .

[0102] It should be noted that, The term represents an exponential decay transformation on the absolute value of curvature. The output approaches 0 in high curvature regions, as high curvature may be caused by noise. Attenuation reduces its interference with the model. Furthermore, it combines... The item retains the concave and convex directions, that is or This allows the model to focus on smooth decay and abrupt changes in low curvature regions, such as The sign inversion enhances the sensitivity to nonlinear inflection points.

[0103] In one embodiment, curvature enhancement feature analysis is performed to reveal the nonlinear feature extraction mechanism of the battery degradation path, wherein, for example... Figure 5 As shown, the battery capacity degradation process begins with an initial linear decrease (0-100 days), followed by accelerated degradation in the middle stage (orange area, 100-150 days), and then severe degradation in the later stage (red area, after 200 days). Figure 6 The curvature curve shown exhibits a significant positive peak in the accelerated decay region (the green curve is convex), consistent with the mathematical characteristics of "convex decay." It maintains a consistently high positive value in the later decay region, reflecting a continuously accelerating decay rate. Figure 7 The enhancement features shown, through sign-preserving exponential transformation, output high absolute values ​​close to ±1 (marked by red triangles) in the steady decay region and low values ​​approaching zero in the accelerated decay region (the curve approaches the zero axis after 200 days). Experimental results indicate that the high absolute value enhancement model is sensitive to steady decay, with the zero-value highlighting the curvature abrupt change point, especially in the 180-220 day range, despite drastic fluctuations in observation capacity (e.g., Figure 5 The blue curve shown indicates that the region is in a state of accelerated decay, as the enhancement feature remains stable. Figure 7 The data shown (which is consistently close to zero) verifies the design concept of "exponential transformation to enhance sensitivity in low curvature regions" in this invention. The coordinate axes are labeled with battery capacity (percentage), curvature (dimensionless), and enhanced feature value (dimensionless), with the time unit being days.

[0104] Optionally, the specific steps for training the prediction model for remaining lifespan in S4 are as follows:

[0105] S401. A primary prediction model for remaining lifetime is constructed by combining a gated temporal convolutional network and the battery health index value;

[0106] Specifically, traditional LSTM and TCN models are insufficient in modeling long-term dependencies during battery aging and do not consider the differences in update frequency of different sensor parameters. For example, the temperature sampling rate is lower than the current sampling rate, resulting in poor multi-rate sensor data fusion. They also lack the ability to capture features in the critical decay stage in the later stages of battery health index decline, cannot adaptively adjust the time receptive field, and have limited ability to model multi-rate data and long-term degradation patterns.

[0107] This invention employs a gated temporal convolutional network, which uses causal convolution to calculate a gate vector to control the update ratio of candidate and hidden states, thereby achieving gated transmission of temporal information. Simultaneously, it dynamically adjusts the convolution kernel expansion coefficient based on the battery health index value to increase or decrease the receptive field, enabling multi-rate sensor data fusion and long-term dependency modeling, and enhancing the model's ability to capture features at different decay stages.

[0108] S402. Combine feature optimization, loss improvement, sample expansion, multi-task prediction, and parameter update to train the primary prediction model and generate a prediction model for remaining lifetime.

[0109] Optionally, the construction of the primary prediction model for remaining lifetime further includes:

[0110] S4011. Based on the enhanced feature vector at the current time step and the hidden state vector at the previous time step, a gating vector is calculated through causal convolution. Then, the gating vector is used to control the generation ratio of candidate state vectors and the update ratio of the hidden state vector at the current time step, thereby realizing the gating transfer of temporal information, as expressed as:

[0111] ;

[0112] ;

[0113] ;

[0114] In the formula, Let represent the gate vector at time point t, which controls the state update ratio, and the sampling interval of the convolution kernel is d; This represents the candidate state vector at time point t. This represents the hidden state vector at time point t. This represents the hidden state vector at time point t-1. This represents the enhanced feature vector at time point t; This indicates a causal convolution operation that ensures the output at time t depends only on the input at time t and previous time points; The weights of the gated convolution kernel are trainable parameters; The convolution kernel weights for candidate states are trainable parameters; This is the gated bias term, which is a trainable parameter; Use the Sigmoid activation function; It is the hyperbolic tangent activation function; Represents the Hadamard product; This indicates a concatenation operation along the feature dimension.

[0115] It should be noted that, in In the item, It is a scalar, but it is extended to AND through the broadcast mechanism. Vectors of the same dimension This represents the proportion vector that retains the old state, for example, if ,but .

[0116] It should also be noted that causal convolution is a type of temporal convolution that ensures that the output time depends only on the input time and previous times, and does not depend on future data. This is different from ordinary convolution operations, which may use future inputs, such as symmetric kernels, thus destroying temporal causality. Causal convolution, on the other hand, achieves this through left padding or kernel offset.

[0117] S4012. Dynamically adjust the dilation coefficient of the convolution kernel in the gated temporal convolutional network based on the value of the battery health index.

[0118] Specifically, the decay-aware dynamic adjustment also includes dynamically adjusting the dilation coefficient of the convolutional kernels in the gated temporal convolutional network based on the battery health index value. When the battery health index is below a threshold, a larger dilation coefficient is used to increase the receptive field and capture long-term dependencies; when the battery health index is above the threshold, a smaller dilation coefficient is used to focus on short-term features, expressed as:

[0119] ;

[0120] In the formula, This represents the convolution kernel expansion coefficient, which controls the sampling interval of the time series. This represents the battery health index, with a value range of [0%, 100%]. It is calculated as the ratio of the current maximum battery capacity to the rated capacity.

[0121] It should be noted that when At that time, the convolution kernel covers the short-term time window, when At this time, the convolution kernel covers a long time window, enhancing the capture of slow degradation features in the later stages of decay. In specific implementations, The expansion coefficient is automatically switched in time, and the problem of "delayed dependence on extension" is solved by increasing the time receptive field.

[0122] In one embodiment, the dynamic adjustment effect of the gating network is analyzed using a dual Y-axis graph, such as... Figure 8 As shown, the effectiveness of the dynamic adjustment strategy for the dilation coefficient in a gated temporal convolutional network is evaluated. Experiments analyze the impact of different convolutional kernel dilation coefficient settings on prediction error as the battery health status decreases from 100% to 0%. The left vertical axis represents the absolute prediction error (percentage), and the right vertical axis represents the dilation coefficient value; the horizontal axis represents the battery health status (percentage). The experiment compares three configurations: a fixed dilation coefficient of 2 (short-term receptive field), a fixed dilation coefficient of 8 (long-term receptive field), and the dynamic adjustment method of this invention. Experimental results show that when the battery health status is above 85%, this invention uses a dilation coefficient of 2 (short-term mode); when it is below 85%, it automatically adjusts the dilation coefficient. Switching to an inflation factor of 8 (long-term mode), as shown by the purple dashed line, in the higher health stage (85%-100%), the model with a fixed inflation factor of 8 experiences an error increase of approximately 0.5% due to an excessively large receptive field. However, this invention maintains a lower error by using a short-term mode. When the health state drops below 85%, the model with a fixed inflation factor of 2 experiences a sharp increase in error to a peak of approximately 3.5% because it cannot capture long-term dependencies. In contrast, the error of this invention stabilizes at approximately 1.5% after switching to the long-term mode. The dynamic adjustment strategy keeps the prediction error at the lowest level throughout the entire lifecycle, with the most significant advantage in the critical decay stage when the health state is below 85%. The vertical dashed line in the figure marks the inflation factor switching threshold.

[0123] Optionally, the prediction model for the remaining lifetime generated in S402 further includes:

[0124] S4021. Based on the defined gated temporal convolutional network, parallel processing of enhanced feature vectors is used to form a fused residual feature matrix;

[0125] Specifically, forward propagation and feature processing also include: solar street light systems contain components with different aging rates, such as batteries, LEDs, and controllers. For example, the aging cycle of batteries is longer than that of LED light decay cycles. Feature extraction methods on a single time scale are difficult to capture both fast and slow degradation modes simultaneously, resulting in inaccurate assessment of the collaborative aging state of multiple components and low accuracy in modeling the collaborative aging of multiple components.

[0126] This invention utilizes a defined gated temporal convolutional network to process enhanced feature vectors in parallel. One branch captures short-window features for faster-aging components, while another branch captures long-window features for slower-aging components. These features are then weighted and fused using a component-adaptive attention weight matrix, and short-circuit connections are added to form a fused residual feature matrix. This process highlights the features of key components, achieving accurate fusion of multiple component degradation modes. The specific steps are as follows:

[0127] 1) Gated temporal convolutional networks for dual-scale feature extraction:

[0128] A gated temporal convolutional network is used to process the enhanced feature vectors in parallel. One branch uses a smaller kernel size to capture short-term features to identify degradation patterns in rapidly aging components, while the other branch uses a larger kernel size to capture long-term features to identify degradation patterns in slowly aging components. This is represented as follows:

[0129] kern ;

[0130] kern ;

[0131] In the formula, Represents the short-time window feature matrix; Represents the feature matrix of the long time window; This represents a gated temporal convolutional network; kern is the kernel size. Corresponding to the short-window convolution kernel, kernel Corresponding to long-time window convolution kernel.

[0132] It should be noted that the kernel size kern and the dilation coefficient d jointly determine the receptive field of the convolutional layer. kern represents the temporal width of the convolutional kernel, i.e., the number of time steps it covers, directly controlling the range of local feature capture. For example, kern=3 means the convolutional kernel covers 3 consecutive time steps. d represents the dilation rate of the convolutional kernel, i.e., the sampling interval, controlling the jump distance across time steps and used to expand the receptive field. The total receptive field of the convolutional layer characterizes the temporal range covered by the convolutional operation, and is jointly determined by the kernel size kern and the dilation coefficient d. For example, when kern=3 and d=8, the receptive field It covers 17 time steps, skipping intermediate points.

[0133] 2) Attention-weighted residual fusion:

[0134] An adaptive attention weight matrix is ​​generated for each component. This matrix is ​​then used to weight and fuse the short-window and long-window feature matrices. Simultaneously, short-circuit connections are used to enhance the feature vectors, forming a fused residual feature matrix. This ensures that the features corresponding to key components are highlighted during the fusion process, as shown below:

[0135] ;

[0136] ;

[0137] In the formula, Represents the fused residual feature matrix; This is the attention weight matrix; Let be the attention weight matrix, which consists of trainable parameters; For 1×1 convolutional short-circuit connections, it is represented as ; is the convolution kernel weight matrix, which consists of trainable parameters; The bias term is a trainable parameter.

[0138] It should be noted that battery parameters such as internal resistance are automatically obtained from high... Value, such as 2. LED parameters such as luminous flux are automatically obtained. Value, such as ,at the same time, The term is used as a residual short-circuit term in the attention weight matrix. Collaborative feature-level load balancing is achieved by combining short-time-window feature matrices. With long time window feature matrix This causes LED light decay to correlate. Battery aging related .

[0139] It should also be noted that, in In the item, It is a scalar, but is expanded to be analogous through broadcasting. Matrices of the same dimension The attention weight matrix represents the features of a long time window, for example, if ,but .

[0140] S4022. Generate dynamic weights corresponding to the actual remaining lifetime ratio based on the constructed exponential dynamic weight function, and calculate the weighted mean square error loss based on the adjusted dynamic weights;

[0141] Specifically, end-of-life-time sensitive loss functions also include: the traditional mean squared error loss function treats prediction errors equally across all life stages, but in practical applications, the cost of end-of-life prediction errors is higher. For example, misjudging short-term remaining lifespan is more serious than long-term misjudgment. Conventional loss functions cannot reflect this stage sensitivity, which can easily lead to insufficient prediction accuracy of the model at the critical end of lifespan.

[0142] This invention constructs an exponential dynamic weighting function. When the proportion of actual remaining life expectancy is lower than a threshold, the weight is increased to enhance the penalty for end-of-life prediction errors. This weight is then applied to the calculation of weighted mean squared error loss, forcing the model to prioritize optimizing end-of-life prediction accuracy and preventing overfitting. The specific steps are as follows:

[0143] 1) Construction of dynamic weight function:

[0144] An exponential dynamic weighting function based on the proportion of actual remaining lifespan is constructed. When the proportion of actual remaining lifespan is lower than a set end-of-life threshold, the value of this weighting function increases sharply, thereby increasing the penalty for end-of-life prediction errors. This is expressed as:

[0145] ;

[0146] In the formula, Indicates the proportion of actual remaining lifespan The corresponding dynamic weights; The value represents the proportion of the actual remaining lifespan, and the range of values ​​is [not specified]. ; For example, set the end-of-life threshold. ,when It is determined to be the end of life; This is the upper limit coefficient for weights, controlling the maximum weight multiple, such as... ; The decay rate coefficient controls the steepness of the weight change, for example, .

[0147] It should be noted that when hour, ,when hour, Follow Decrease and increase, for example, hour, ; hour, , When the dynamic weight limit value is .

[0148] 2) Ensemble of regularized loss functions:

[0149] A dynamic weighting function is applied to the weighted mean squared error loss calculation, allowing the prediction error of end-of-life samples to account for a larger proportion of the total loss. Simultaneously, an L2 regularization term is used to constrain all trainable parameters of the model to prevent overfitting, expressed as:

[0150] ;

[0151] In the formula, Represents the sensitive loss function; This refers to the batch sample size. For sample index; For the first The true remaining lifespan proportion of each sample; For the first The proportion of predicted remaining lifespan for each sample, with an initial value set at 50%; Indicates the first The dynamic weights corresponding to the proportion of the true remaining lifespan of each sample; For example, the regularization coefficient, ; It is an L2 norm; It represents all trainable parameters of the model, in set form.

[0152] S4023. Generate samples that conform to the physical degradation law from the feature vector of real fault samples;

[0153] Specifically, the generation of critical state failure samples also includes: actual failure samples of solar street light systems are scarce, accounting for only a small proportion of the training set, resulting in a serious imbalance in model training data. Samples generated by conventional oversampling methods may not conform to the physical degradation law and cannot accurately simulate the critical failure state of the equipment. Since the actual degradation mechanism of the equipment is ignored, the generated samples lack physical rationality.

[0154] This invention calculates the gradient direction of the loss function at the feature vector of a real fault sample to generate a perturbation vector, applies a random binary mask to constrain the perturbation range, and superimposes the constrained perturbation onto the feature vector to generate critical state samples. In each iteration, samples conforming to the physical degradation law are generated, effectively expanding the pre-fault state data, as shown below:

[0155] 1) Fault sample gradient calculation:

[0156] At the augmented feature vector of the real fault sample, the gradient direction of the model loss function with respect to the augmented feature vector is calculated. This gradient indicates the degradation path where the model's prediction error increases the fastest, and is expressed as:

[0157] :

[0158] In the formula, Represents the perturbation vector; Represents the enhanced feature vector; The feature vector representing a real fault sample is the real enhanced feature vector extracted from historical fault samples, i.e. The value to be taken when the equipment fails; The disturbance intensity coefficient controls the sample offset amplitude. ; The sign for partial derivatives; Indicating in real fault samples Calculate the gradient at the location;

[0159] It should be noted that, The gradient direction indicates the degradation path that the model finds hardest to identify. The gradient points in the direction of the fastest increase in loss, which is the path with the largest prediction error of the model. This path often corresponds to the key failure mode in the actual degradation mechanism, such as a sharp increase in battery internal resistance. Because the model has weak generalization ability in these areas, additional samples are needed for reinforcement learning, which is consistent with the actual failure mechanism of the device.

[0160] 2) Mask-constrained sample generation:

[0161] A perturbation vector is generated using the calculated gradient direction. Then, a random binary mask vector is applied to constrain the range of this perturbation vector. The constrained perturbation vector is then superimposed onto the enhanced feature vector to generate a critical state sample that conforms to the physical degradation law, represented as:

[0162] ;

[0163] In the formula, This represents a sample in a critical state. It is a random binary mask vector.

[0164] It should be noted that the random binary mask vector Randomly mask some feature disturbances, such as modifying only battery parameters while keeping LED parameters unchanged, to avoid generating samples that violate the device linkage mechanism.

[0165] Furthermore, the generated critical state samples tags Set as , This represents the remaining lifespan percentage of a real failure sample, for example, if the failure occurs... ,but If the critical state ,but .

[0166] Furthermore, in the next iteration of model training, By inputting real data, the problem of scarce fault samples can be solved, critical state data can be expanded, the model's ability to identify pre-fault states can be improved, and prediction bias can be reduced.

[0167] S4024. After fusing the residual feature matrix, two fully connected layers are connected in parallel as prediction heads, and pseudo labels are generated based on the physical threshold rule. The cross-entropy loss of the auxiliary task is added to the loss of the main task according to the weight to form the total loss function.

[0168] Specifically, directly predicting the overall remaining lifespan of the system ignores the independent degradation information of key components, resulting in poor interpretability of model features and an inability to use component-level failure patterns to constrain the learning process.

[0169] This invention uses two fully connected layers in parallel as prediction heads after fusing the residual feature matrix. One outputs the proportion of remaining system lifetime, and the other outputs the probability vector of component health status. Pseudo-labels are generated based on physical threshold rules. The cross-entropy loss of the auxiliary task is weighted and added to the loss of the main task to form the total loss function, improving the model's feature discriminativeness and interpretability. This is expressed as:

[0170] 1) Dual-task prediction head construction:

[0171] After fusing the residual feature matrix, two fully connected layers are connected in parallel as prediction heads. One prediction head serves as the main task, outputting the predicted proportion of remaining lifetime of the entire system. The other prediction head serves as an auxiliary task, outputting the health status probability vector of each component, as follows:

[0172] ;

[0173] ;

[0174] In the formula, This represents the predicted remaining lifetime percentage, which is the predicted remaining lifetime percentage output by the main task. This represents the component health probability vector, which is the component health status probability vector output by the auxiliary task; The main task weight matrix consists of trainable parameters. The bias term for the main task is a trainable parameter; The auxiliary task weight matrix consists of trainable parameters. The bias vector for the auxiliary task is a trainable parameter; This is the Softmax normalized exponential function.

[0175] In one embodiment, the component health probability vector Battery health probability LED health probability Controller health probability and the health probability of photovoltaic panels composition.

[0176] 2) Expert rule label generation and loss fusion:

[0177] Pseudo-labels for component health status are generated based on predefined physical threshold rules. Then, the cross-entropy loss of the auxiliary task is calculated, and the cross-entropy loss is added to the sensitivity loss function of the main task according to the weight coefficients to form the final total loss function used for model optimization, expressed as:

[0178] In the formula, Represents the total loss function; Indicates loss of the main task; For auxiliary task weighting coefficients, such as, ; Let cross-entropy be the loss function. The calculation method is expressed as ; component health probability vector The ke-th element; ke is the component index, ke=1 corresponds to the battery health probability. ke=2 corresponds to the healthy probability of an LED. ke=3 corresponds to the controller health probability ke=4 corresponds to the health probability of the photovoltaic panel. .

[0179] It should be noted that the design of the loss function couples the physical rule pseudo-labels with multi-task prediction, and the component health probability vector Generated by threshold rules, it eliminates the need for component-level annotation, enabling unsupervised knowledge injection. Furthermore, in auxiliary tasks, it utilizes... Item forced fusion feature By decoupling the component-sensitive subspace, the model automatically learns feature combinations that conform to physical laws.

[0180] S4025. Perform backpropagation based on the total loss function calculation results until the preset number of iterations or the loss convergence threshold is reached.

[0181] Specifically, backpropagation and parameter updates also include:

[0182] Backpropagation is performed based on the total loss function calculation result. Specifically, firstly, the partial derivatives of the total loss with respect to each trainable parameter of the model are calculated, and the gradient is propagated backward along the computation graph. Then, an adaptive moment estimation optimization algorithm is used to dynamically adjust the learning rate by combining the historical gradient squared terms with the current gradient direction in order to balance the convergence speed and stability.

[0183] During gradient update, critical state sample data are integrated synchronously to ensure that pre-fault state features participate in parameter optimization.

[0184] After the parameters are updated, the current batch computation graph is cleared, and the next batch of normalized training data (including real samples and generated samples) is loaded. The forward propagation process is repeated until the preset number of iterations or the loss convergence threshold is reached.

[0185] Optionally, the remaining lifespan of the solar streetlights is obtained from the prediction model based on the remaining lifespan in S5. After the model is trained, the remaining lifespan is predicted based on the real-time monitoring data of the new streetlights. The specific steps are as follows:

[0186] 1. Perform the state parameter normalization operation defined in S2 on the newly collected raw state parameters to obtain the normalized parameters;

[0187] 2. Perform the data augmentation operation defined in S3 on the normalized parameters to obtain the augmented feature vector;

[0188] 3. The enhanced feature vector is input into the trained gated temporal convolutional network. After dual-scale feature extraction (short / long time window branches) and component adaptive attention fusion, the dual-task prediction head synchronously outputs the overall remaining lifespan ratio prediction value of the system (main task) and the health status probability of the battery / LED / controller / photovoltaic panel (auxiliary task).

[0189] 4. When the remaining lifespan percentage is lower than a preset threshold (e.g., 15%), an early warning is triggered, and potential faulty components are located based on the component health probability.

[0190] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0191] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. In particular, any reference to memory, database, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0192] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0193] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0194] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A predictive method for monitoring the remaining lifespan of an integrated solar street light, characterized in that, The method includes: S1. Status parameters are collected in real time through a sensor network installed on various components of the solar street light; S2. Use Fourier series to extract the environmental periodic components of the illumination parameters, and decouple and normalize the environmental periodic components to obtain normalized parameters; S3. Use the discrete Laplacian operator to calculate the local curvature values ​​of battery-related parameters, then perform a sign-preserving exponential transformation on the curvature values ​​to generate enhanced feature terms, and concatenate them with the normalized feature vector to form an enhanced feature vector; S4. Combine a temporal convolutional network with a gated update mechanism and a dynamic dilation factor to adaptively fuse multi-rate sensor data and train a prediction model for remaining lifetime. S5. Obtain the remaining lifespan of solar streetlights based on the prediction model of remaining lifespan.

2. The prediction method for monitoring the remaining lifespan of an integrated solar street light according to claim 1, characterized in that, The normalized parameters obtained in S2 specifically include: S201. Use a third-order Fourier series to extract the mean of the environmental periodic components of illumination parameters in historical data; S202. Calculate the deperiodic standard deviation based on the mean of the environmental periodic components; perform normalization on the original illumination-related parameters based on the deperiodic standard deviation, and perform standard Z-score normalization on the non-illumination-related parameters to obtain the normalized parameters.

3. The prediction method for monitoring the remaining lifespan of an integrated solar street light according to claim 1, characterized in that, The process of concatenating the normalized feature vector with the enhanced feature vector in S3 also includes: The calculated curvature values ​​are subjected to a sign-preserving exponential transform to generate enhanced feature terms. These enhanced feature terms are then concatenated with the normalized feature vector to form an enhanced feature vector, expressed as: ； In the formula, Represents the enhanced feature vector; Represents the normalized eigenvectors; It is a natural exponential function; This represents a vector concatenation operation; The curvature sensitivity coefficient controls the exponential decay rate. This represents the absolute value of curvature calculated at the current time point; This represents the curvature value calculated at the current point in time. For a symbolic function, for Item, when Output when >0 ,when Output when <0 .

4. The prediction method for monitoring the remaining lifespan of an integrated solar street light according to claim 1, characterized in that, The training of the remaining lifespan prediction model in S4 also includes: S401. A primary prediction model for remaining lifetime is constructed by combining a gated temporal convolutional network and the battery health index value; S402. Combine feature optimization, loss improvement, sample expansion, multi-task prediction, and parameter update to train the primary prediction model and generate a prediction model for remaining lifetime.

5. The prediction method for monitoring the remaining lifespan of an integrated solar street light according to claim 1, characterized in that, The preliminary prediction model for remaining lifetime constructed in S401 also includes: S4011. Based on the enhanced feature vector at the current time and the hidden state vector at the previous time, a gating vector is calculated through causal convolution. Then, the gating vector is used to control the generation of candidate state vectors and the update ratio of the hidden state vector at the current time, so as to perform gating transmission of temporal information. S4012. Dynamically adjust the dilation coefficient of the convolution kernel in the gated temporal convolutional network based on the value of the battery health index.

6. The prediction method for monitoring the remaining lifespan of an integrated solar street light according to claim 5, characterized in that, The prediction model for generating remaining lifetime in S402 also includes: S4021. Based on the defined gated temporal convolutional network, parallel processing of enhanced feature vectors is used to form a fused residual feature matrix; S4022. Generate dynamic weights corresponding to the actual remaining lifetime ratio based on the constructed exponential dynamic weight function, and calculate the weighted mean square error loss based on the adjusted dynamic weights; S4023. Generate samples that conform to the physical degradation law from the feature vector of real fault samples; S4024. After fusing the residual feature matrix, two fully connected layers are connected in parallel as the prediction head to form the total loss function; S4025. Perform backpropagation based on the total loss function calculation results until the preset number of iterations or the loss convergence threshold is reached.

7. The prediction method for monitoring the remaining lifespan of an integrated solar street light according to claim 6, characterized in that, The calculation of weighted mean square error loss based on the adjusted dynamic weights in S4022 also includes: Construct an exponential dynamic weighting function based on the proportion of actual remaining lifespan, and set a lifespan threshold based on the proportion of actual remaining lifespan. The dynamic weighting function is applied to the calculation of the weighted mean squared error loss, and the sensitive loss function is obtained by using the L2 regularization term to constrain all trainable parameters of the model.

8. The prediction method for monitoring the remaining lifespan of an integrated solar street light according to claim 6, characterized in that, The generation of samples conforming to the laws of physical degradation in S4023 also includes: The gradient direction of the loss function is calculated at the feature vector of the real fault sample to generate the perturbation vector; The perturbation vector is generated using the calculated gradient direction. Then, a random binary mask vector is applied to constrain the range of the perturbation vector. Finally, the constrained perturbation vector is superimposed on the enhanced feature vector to generate a critical state sample that conforms to the physical degradation law.

9. A prediction method for monitoring the remaining lifespan of an integrated solar street light according to claim 6, characterized in that, The process of forming the total loss function in S4024 also includes: After fusing the residual feature matrix, two fully connected layers are connected in parallel as prediction heads. One prediction head is the main task, outputting the predicted value of the remaining lifetime ratio of the entire system, and the other prediction head is the auxiliary task, outputting the health status probability vector of each component. Pseudo-labels for component health status are generated based on predefined physical threshold rules. Then, the cross-entropy loss of the auxiliary task is calculated, and the cross-entropy loss is added to the sensitivity loss function of the main task according to the weight coefficient to form the total loss function.

10. The prediction method for monitoring the remaining lifespan of an integrated solar street light according to claim 1, characterized in that, The method for obtaining the remaining lifespan of solar streetlights based on the prediction model of remaining lifespan in S5 also includes: S501. Perform state parameter normalization operation on the original state parameters of the current detection time node to obtain the normalized parameters; S502. Perform data augmentation on the normalized parameters to obtain the augmented feature vector; S503. Input the enhanced feature vector into the trained gated temporal convolutional network to output the overall remaining lifespan ratio prediction of the system and the health status probability of each component of the solar street light. S504. When the remaining lifespan ratio is lower than a preset threshold, an early warning is triggered, and potential faulty components are located based on the component health probability.

Images