Charging pile line fire-fighting early warning method based on big data and storage medium

Abstract

The invention provides a charging pile line fire-fighting early warning method based on big data and a storage medium, and the method comprises the steps: collecting a line operation data set of a target charging pile cluster, the line operation data set comprises multi-source time sequence monitoring data, carrying out the cross-modal feature alignment of the multi-source time sequence monitoring data, and generating a time-space correlation feature matrix; inputting the space-time correlation feature matrix into a pre-trained fire risk prediction model, generating a line abnormal risk probability distribution diagram, generating a layered early warning signal set according to risk levels corresponding to space nodes in the line abnormal risk probability distribution diagram, and triggering a dynamic protection mechanism based on the layered early warning signal set. The dynamic protection mechanism includes performing a current cut-off operation on a high-risk line segment, and performing a power attenuation operation on an adjacent line segment. According to the invention, the reliability and the intelligent level of the charging pile line fire-fighting early warning system can be comprehensively improved.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a charging pile line fire warning method and storage medium based on big data. Background Art

[0002] With the rapid development of new energy infrastructure, charging pile line fire warning technology has become a core link to ensure the safety of electric vehicle charging. The current mainstream warning system mainly relies on single parameter threshold monitoring or simple multi-source data superposition analysis, and sets current over-limit values ​​or temperature warning lines for passive alarm. However, such methods are difficult to capture the dynamic coupling relationship between current fluctuations and temperature changes, and ignore the continuous impact of line aging on electrical and thermal parameters, resulting in delayed identification of hidden fire risks and high misjudgment rate. The processing method based on fixed weight fusion of multi-source data in the existing technology cannot adapt to the evolution law of nonlinear characteristics of charging pile lines under complex working conditions, and the protection measures mostly adopt the extensive treatment of global power outage, which seriously restricts the operational efficiency of charging pile clusters. More importantly, the traditional method lacks the ability to predict the risk transmission path, and it is difficult to block the chain reaction caused by local faults in time. There is an irreconcilable contradiction between ensuring system safety and maintaining service continuity. These technical defects make it difficult for the existing warning system to meet the urgent needs for accurate warning and intelligent protection in high-density charging scenarios. Summary of the invention

[0003] In view of this, the embodiment of the present application provides a charging pile line fire warning method and storage medium based on big data. The technical solution of the embodiment of the present application is implemented as follows: On the one hand, the present invention provides a charging pile line fire warning method based on big data, the method comprising: collecting a line operation data set of a target charging pile cluster, the line operation data set comprising multi-source time series monitoring data; performing cross-modal feature alignment on the multi-source time series monitoring data to generate a spatiotemporal correlation feature matrix, wherein the cross-modal feature alignment comprises dynamically adjusting the weight allocation of temperature change features based on the current fluctuation amplitude, and correcting the correlation between current and temperature features based on line aging indicators; inputting the spatiotemporal correlation feature matrix into a pre-trained fire risk prediction model to generate a line abnormal risk probability distribution map, the fire risk prediction model being trained by a mapping relationship between a historical fire event data set and multi-source monitoring data; generating a hierarchical warning signal set according to the risk level corresponding to each spatial node in the line abnormal risk probability distribution map, the hierarchical warning signal set comprising differentiated control instructions for different charging pile line segments; triggering a dynamic protection mechanism based on the hierarchical warning signal set, the dynamic protection mechanism comprising performing a current cutoff operation on a high-risk line segment, and performing a power attenuation operation on an adjacent line segment.

[0004] On the other hand, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, and the computer program implements the steps in the above method when executed by a processor.

[0005] The charging pile line fire warning method based on big data provided by the present invention models the time-space correlation of multi-source time series monitoring data such as current fluctuation, temperature change and line aging through a dynamic cross-modal feature alignment mechanism, generates a time-space correlation feature matrix integrating physical degradation characteristics, and accurately identifies the spatial distribution characteristics of line abnormal risks based on the pre-trained fire risk prediction model, and triggers the dynamic protection mechanism in combination with the layered warning signal set, which significantly improves the warning accuracy and response efficiency of the fire hazard of the charging pile line. The method realizes the active blocking of the risk diffusion path through the coordinated operation of current interruption and power attenuation, effectively overcomes the false alarm and missed alarm problems caused by the static feature association in the traditional method, and dynamically optimizes the protection strategy by using the feedback mechanism, and maximizes the operation stability of the charging pile system on the basis of ensuring the timeliness of risk disposal. In addition, the risk prediction model is trained based on the mapping relationship between historical fire events and multi-source monitoring data, which can adaptively learn the evolution law of implicit risk factors, break through the limitations of artificial experience rules, and make the warning decision process conform to the physical law of electrical and thermal coupling and have data-driven forward-looking judgment capabilities, thereby comprehensively improving the reliability and intelligence level of the charging pile line fire warning system. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1 A schematic diagram of the implementation process of a charging pile line fire warning method based on big data provided in an embodiment of the present application.

[0007] Figure 2 A schematic diagram of the composition structure of a charging pile line fire warning device based on big data provided in an embodiment of the present application. DETAILED DESCRIPTION

[0008] In order to make the purpose, technical solutions and advantages of the present application clearer, the technical solutions of the present application are further elaborated in detail below in conjunction with the drawings and embodiments. The described embodiments should not be regarded as limiting the present application. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0009] The embodiment of the present application provides a charging pile line fire warning method based on big data. The method can be executed by a processor in a computer system at the edge side of the charging station, or by a cloud server remotely connected to the charging station. Figure 1 A schematic diagram of the implementation process of a charging pile line fire warning method based on big data provided in an embodiment of the present application is shown in FIG. Figure 1 As shown, the method comprises the following steps: Step S100: collecting a line operation data set of a target charging pile cluster, wherein the line operation data set includes multi-source time series monitoring data, and the multi-source time series monitoring data includes a current fluctuation feature sequence, a temperature change feature sequence, and a line aging index sequence.

[0010] In an embodiment of the present invention, the target charging pile cluster is a set of charging piles that need to be monitored for fire warning. These charging piles are related in physical location or function, such as being located in the same charging station or in the same area. The line operation data set is a set formed after collecting data on the line operation status of the target charging pile cluster. Multi-source time series monitoring data refers to data collected from multiple different sources in chronological order, which includes current fluctuation feature sequences, temperature change feature sequences, and line aging index sequences.

[0011] The current fluctuation characteristic sequence is a data sequence obtained by continuously monitoring the current fluctuation in the charging pile line within a set time range. For example, a high-precision current sensor can be used to sample the current at a frequency of once per minute (only for example, high-frequency sampling, such as kHz level, can also be used), and the current values ​​obtained by each sampling are arranged in chronological order to form a current fluctuation characteristic sequence. This sequence can reflect the change in the magnitude of the current at different times, such as a sudden increase or decrease in the current. The temperature change characteristic sequence is a data sequence obtained by monitoring the change in the temperature of the charging pile line over time. A temperature sensor can be used to measure the line temperature at a frequency of once per minute, and the measured temperature values ​​are arranged in chronological order to form a temperature change characteristic sequence. The line aging index sequence is a data sequence used to measure the degree of aging of the charging pile line. Line aging may be caused by a variety of factors, such as long-term use, environmental factors, etc. The line aging index can be obtained by testing the conductivity, insulation performance, etc. of the line. For example, the resistance value and insulation resistance value of the line are regularly tested, and these test results are arranged in chronological order to form a line aging index sequence. This sequence can reflect the degree and trend of line aging.

[0012] In the actual data collection process, corresponding sensors such as current sensors, temperature sensors, etc. can be installed on each key line node of the target charging pile cluster. These sensors transmit the real-time collected data to the data acquisition device, which performs preliminary processing and storage on the data to form a line operation data set.

[0013] Step S200: performing cross-modal feature alignment on multi-source time series monitoring data to generate a spatiotemporal correlation feature matrix, wherein the cross-modal feature alignment includes dynamically adjusting the weight allocation of temperature change features based on current fluctuation amplitude, and correcting the correlation between current and temperature features based on line aging indicators.

[0014] Cross-modal feature alignment refers to processing different types of feature data (such as current fluctuation features, temperature change features, and line aging indicators in the embodiment of the present invention) so that they are consistent and correlated in time and space dimensions. In the embodiment of the present invention, the specific method of cross-modal feature alignment includes dynamically adjusting the weight allocation of temperature change features based on the current fluctuation amplitude, and correcting the correlation between current and temperature features based on the line aging indicator. Dynamically adjusting the weight allocation of temperature change features based on the current fluctuation amplitude is to dynamically change the importance of temperature change features in subsequent processing according to the size of current fluctuations.

[0015] The correlation between current and temperature characteristics is corrected based on the line aging index because line aging will affect the normal relationship between current and temperature. For example, as the line ages, the resistance of the line may increase, resulting in a greater temperature increase at the same current. Therefore, it is necessary to correct the correlation between current and temperature characteristics based on the line aging index to more accurately reflect the actual operating conditions of the line.

[0016] Generating a spatiotemporal correlation feature matrix is ​​to integrate the multi-source time series monitoring data after cross-modal feature alignment processing to form a matrix containing time and space information. This matrix can more comprehensively reflect the operating status of the target charging pile cluster line. In actual operation, corresponding algorithms and models can be used to achieve cross-modal feature alignment and the generation of spatiotemporal correlation feature matrices. For example, the attention mechanism in deep learning can be used to dynamically adjust the weight allocation of temperature change characteristics based on the current fluctuation amplitude, and by establishing a regression model to correct the correlation between current and temperature characteristics based on line aging indicators, the processed data is finally arranged according to the time and space dimensions to form a spatiotemporal correlation feature matrix.

[0017] As an implementation mode, step S200 performs cross-modal feature alignment on multi-source time series monitoring data to generate a spatiotemporal correlation feature matrix, which may specifically include the following steps S210-S250: Step S210: construct a time dimension sliding window according to the amplitude change rate of the current fluctuation characteristic sequence, and extract the current fluctuation peak-to-valley difference and the duration time interval in the sliding window as the first dynamic characteristic vector.

[0018] The time dimension sliding window is a method for performing local analysis on time series data. It can slide on the data sequence according to a preset step size, and intercept a fixed length of data segment for analysis each time. In an embodiment of the present invention, the time dimension sliding window is constructed according to the amplitude change rate of the current fluctuation characteristic sequence. The amplitude change rate reflects the speed of the change of the current fluctuation amplitude over time. For example, the amplitude change rate is obtained by calculating the ratio of the difference between the current fluctuation amplitudes at two adjacent time points and the time interval. According to the size of the amplitude change rate, the length and step size of the sliding window can be determined.

[0019] The current fluctuation peak-to-valley difference and the duration time interval are extracted in the sliding window as the first dynamic feature vector. The current fluctuation peak-to-valley difference is the difference between the maximum and minimum values ​​of the current fluctuation in the sliding window, which reflects the magnitude of the current fluctuation. The duration time interval refers to the time from the peak value to the valley value or from the valley value to the peak value of the current fluctuation. These two features are combined to form the first dynamic feature vector.

[0020] Step S220: extracting the temperature gradient variation and extreme value distribution density of the temperature variation feature sequence in the sliding window as the second dynamic feature vector.

[0021] The temperature gradient change refers to the rate of change of temperature over time in the sliding window, which reflects the speed of temperature change. The temperature gradient change can be obtained by calculating the ratio of the difference between the temperature values ​​of two adjacent time points and the time interval.

[0022] The extreme value distribution density refers to the frequency of occurrence of temperature extremes (maximum and minimum values) in the sliding window. The extreme value distribution density can be obtained by counting the number of occurrences of the maximum and minimum temperature values ​​in the sliding window and dividing it by the total time of the sliding window. The temperature gradient change and the extreme value distribution density are combined to form the second dynamic feature vector.

[0023] Step S230: decomposing the line aging index sequence into two independent components, namely, the conductive material degradation coefficient and the insulation layer damage rate, and respectively calculating the cross-influence factors between the independent components and the first dynamic feature vector and the second dynamic feature vector.

[0024] The line aging index sequence reflects the degree of aging of the charging pile line over time. In an embodiment of the present invention, the line aging index sequence is decomposed into two independent components: the conductive material degradation coefficient and the insulation layer damage rate. The conductive material degradation coefficient is used to measure the degree of degradation of the conductive material performance in the line. For example, as the use time increases, the resistance of the conductive material may increase and the conductive performance will decrease. The conductive material degradation coefficient can be determined by measuring the resistance change of the line. The insulation layer damage rate is the degree of damage to the insulation layer of the line, which can be evaluated by detecting the insulation resistance and other methods.

[0025] The cross-influence factor between the independent component and the first dynamic eigenvector and the second dynamic eigenvector is calculated to analyze the influence of line aging on current fluctuation and temperature change. The cross-influence factor reflects the degree of influence of one variable on another variable. For example, by calculating the cross-influence factor between the conductive material degradation coefficient and the first dynamic eigenvector, the influence of the conductive material performance degradation on the current fluctuation can be analyzed; by calculating the cross-influence factor between the insulation layer breakage rate and the second dynamic eigenvector, the influence of the insulation layer breakage on the temperature change can be analyzed. In actual calculations, correlation analysis and other methods can be used to calculate the cross-influence factor. For example, by calculating the Pearson correlation coefficient between the conductive material degradation coefficient and the current fluctuation peak-to-valley difference in the first dynamic eigenvector, the cross-influence factor between the two is obtained. If the correlation coefficient is positive and the value is large, it means that the increase in the conductive material degradation coefficient will lead to an increase in the current fluctuation peak-to-valley difference; if the correlation coefficient is negative and the value is large, it means that the increase in the conductive material degradation coefficient will lead to a decrease in the current fluctuation peak-to-valley difference.

[0026] As an implementation mode, step S230 decomposes the line aging index sequence into two independent components, namely, the conductive material degradation coefficient and the insulation layer damage rate, and respectively calculates the cross-influence factors between the independent components and the first dynamic feature vector and the second dynamic feature vector. Specifically, the following steps S231 to S235 may be included: Step S231: performing frequency domain decomposition on the line aging index sequence, extracting the low-frequency fluctuation component as the conductive material degradation coefficient, and separating the high-frequency mutation component as the insulation layer damage rate.

[0027] Frequency domain decomposition is a method of converting time domain signals to frequency domain for analysis. Through frequency domain decomposition, the signal can be decomposed into components of different frequencies. In an embodiment of the present invention, the line aging indicator sequence is subjected to frequency domain decomposition, for example, the line aging indicator sequence is converted from the time domain to the frequency domain using a fast Fourier transform (FFT). Low-frequency fluctuation components usually reflect the slow change trend of the signal. In this scenario, the degradation of the conductive material is a relatively slow process, so the low-frequency fluctuation component is extracted as the conductive material degradation coefficient. The high-frequency mutation component reflects the rapid change of the signal. The damage of the insulating layer may cause a sudden change in the line aging indicator, so the high-frequency mutation component is separated as the insulation layer damage rate.

[0028] For example, assuming that the line aging index sequence is a time series containing multiple data points, after performing FFT transformation on it, a frequency domain signal is obtained. By setting an appropriate frequency threshold, the low-frequency signal is extracted, and converted back to the time domain through inverse Fourier transform to obtain the conductive material degradation coefficient sequence; the high-frequency signal is extracted and also converted back to the time domain through inverse Fourier transform to obtain the insulation layer damage rate sequence.

[0029] In an optional implementation, in order to overcome the decomposition failure caused by low-frequency sampling and the noise contained in the high-frequency components, empirical mode decomposition (EMD) can be used instead of FFT to adapt to non-uniform sampling data and extract more reasonable low-frequency (conductivity degradation) and high-frequency (insulation damage) components.

[0030] Step S232: Time-align the degradation coefficient of the conductive material with the current fluctuation peak-to-valley difference in the first dynamic eigenvector, calculate the phase correlation between the two within the sliding window, and generate a current-degradation interaction eigenvector.

[0031] Time alignment is to match the data of different time series in the time dimension so that they have a corresponding relationship at the same time point. In an embodiment of the present invention, the conductive material degradation coefficient sequence is time-aligned with the current fluctuation peak-to-valley difference sequence in the first dynamic feature vector. For example, if the sampling time interval of the conductive material degradation coefficient sequence is 1 hour and the sampling time interval of the current fluctuation peak-to-valley difference sequence is 30 minutes, one of the sequences is interpolated or sampled so that the time points of the two sequences are consistent.

[0032] Phase correlation refers to the degree of correlation between two signals in phase. The phase correlation between the degradation coefficient of the conductive material and the peak-to-valley difference of the current fluctuation is calculated within the sliding window, for example, the cross-correlation function is used to calculate the phase correlation between the two. The current-degradation interaction feature vector is generated by the calculated phase correlation result. This vector reflects the interaction between the degradation of the conductive material and the current fluctuation. For example, if the phase correlation is strong, it means that the degradation of the conductive material has a greater impact on the current fluctuation; if the phase correlation is weak, it means that the impact between the two is small.

[0033] Step S233: spatially match the insulation layer damage rate with the temperature gradient change in the second dynamic feature vector, detect the synchronous offset of the two in the extreme point distribution, and generate a temperature-damage interaction feature vector.

[0034] Spatial matching is to correspond and compare different data in the spatial dimension. In an embodiment of the present invention, the insulation layer breakage rate is spatially matched with the temperature gradient change in the second dynamic feature vector. The space here can be understood as different positions of the charging pile line. For example, the insulation layer breakage rate and the temperature gradient change are measured on different line sections of the charging pile, and the insulation layer breakage rate and the temperature gradient change on the same line section are matched.

[0035] Detecting the synchronous offset of the two in the distribution of extreme points is to analyze the time or space offset between the extreme points (maximum or minimum) of the insulation damage rate and the extreme points of the temperature gradient change. For example, if the maximum value of the insulation damage rate occurs at a certain time point, and the maximum value of the temperature gradient change occurs at a later time point, there is a time synchronous offset between the two. By detecting the synchronous offset, a temperature-damage interaction feature vector is generated. This vector reflects the interaction between insulation damage and temperature change.

[0036] Step S234: performing multi-scale convolution processing on the current-degradation interaction feature vector, extracting the hysteresis response pattern of the degradation coefficient of the conductive material to the change of the current fluctuation amplitude, and generating a first cross-influence factor sequence.

[0037] Multi-scale convolution processing is a method of performing convolution operations on data at different scales, which can capture the characteristics of data at different scales. In an embodiment of the present invention, multi-scale convolution processing is performed on the current-degradation interaction feature vector. For example, convolution operations are performed on the current-degradation interaction feature vector using convolution kernels of different sizes (such as 3×3, 5×5, etc.) to obtain feature representations at different scales.

[0038] Extracting the hysteresis response pattern of the degradation coefficient of the conductive material to the change of the current fluctuation amplitude refers to analyzing how the change of the degradation coefficient of the conductive material affects the change of the current fluctuation amplitude, and whether there is a time lag in this effect. Through multi-scale convolution processing, this hysteresis response pattern can be extracted from different scales. For example, at a certain scale, it may be found that the change of the degradation coefficient of the conductive material will cause a significant change in the current fluctuation amplitude after a period of time. This is a hysteresis response pattern. The extracted hysteresis response pattern is sorted to generate the first cross-influence factor sequence. This sequence reflects the degree and pattern of the influence of the conductive material degradation on the current fluctuation.

[0039] Step S235: performing bidirectional gated loop processing on the temperature-damage interaction feature vector to capture the cumulative effect of the insulation layer damage rate on the temperature extreme value distribution and generate a second cross-influence factor sequence.

[0040] Bidirectional gated recurrent processing is a processing method based on a recurrent neural network, which can consider both past and future information of the data. In an embodiment of the present invention, bidirectional gated recurrent processing is performed on the temperature-damage interaction feature vector. For example, a bidirectional gated recurrent unit (GRU) is used to process the temperature-damage interaction feature vector. The bidirectional GRU can process the input data from two directions (forward and reverse), which can better capture the context information of the data.

[0041] Capturing the cumulative effect of the insulation damage rate on the temperature extreme value distribution refers to analyzing how the change in the insulation damage rate has a cumulative impact on the temperature extreme value distribution over time. Through bidirectional gated loop processing, the temperature-damage interaction feature vector can be modeled to capture this cumulative effect. For example, as the insulation damage rate gradually increases, the range of the temperature extreme value distribution may gradually expand. This change is a cumulative effect. The captured cumulative effect is quantified to generate a second cross-influence factor sequence. This sequence reflects the degree and pattern of the impact of insulation damage on temperature changes.

[0042] Step S240: Dynamically weight the cross-influence factors through a multi-channel attention mechanism to generate an association mapping relationship between the current fluctuation feature sequence, the temperature change feature sequence, and the line aging index sequence.

[0043] The multi-channel attention mechanism can dynamically adjust the importance of each feature channel according to the specific situation of the data. In an embodiment of the present invention, a multi-channel attention mechanism is used to dynamically weight the first cross-influence factor sequence and the second cross-influence factor sequence. For example, an attention module based on a convolutional neural network is used, which can automatically learn the importance weight of each factor according to the input cross-influence factor sequence.

[0044] Generating the correlation mapping relationship between the current fluctuation feature sequence, the temperature change feature sequence, and the line aging index sequence means establishing the correlation between the three sequences by dynamically assigning weights to the cross-influence factors. For example, the weights calculated by the attention mechanism are weighted to combine the first cross-influence factor sequence and the second cross-influence factor sequence with the current fluctuation feature sequence, the temperature change feature sequence, and the line aging index sequence to obtain a mapping matrix that reflects the correlation between them. This mapping matrix can be used for subsequent analysis and processing, such as predicting abnormal risks of lines.

[0045] As an implementation mode, step S240, dynamically weighting the cross-influence factors through a multi-channel attention mechanism, generating an association mapping relationship between the current fluctuation feature sequence, the temperature change feature sequence, and the line aging index sequence, may specifically include the following steps S241-S245: Step S241: Channel-joining the first cross-influence factor sequence and the current fluctuation feature sequence to form a current modal association feature block, and channel-joining the second cross-influence factor sequence and the temperature change feature sequence to form a temperature modal association feature block.

[0046] In an embodiment of the present invention, the first cross-influence factor sequence is channel-joined with the current fluctuation feature sequence. Similarly, the second cross-influence factor sequence is channel-joined with the temperature change feature sequence to form a temperature modal associated feature block. Channel jointing can integrate related feature information together to facilitate subsequent processing and analysis. For example, by jointing the first cross-influence factor sequence with the current fluctuation feature sequence, the information on the impact of conductive material degradation on current fluctuation can be combined with the information on the current fluctuation itself, thereby more comprehensively reflecting the characteristics of the current mode.

[0047] The following processing is performed on the current mode associated feature block and the temperature mode associated feature block through the cross-modal attention gating mechanism: Step S242: In the current-dominated attention branch, the weighted influence value of each time step of the current mode-related feature block on the temperature mode-related feature block is calculated to generate an attention distribution diagram from current to temperature.

[0048] The current-dominant attention branch is a branch in the cross-modal attention gating mechanism, which focuses on the influence of the current mode-related feature block on the temperature mode-related feature block. In this branch, the weighted influence value of each time step of the current mode-related feature block on the temperature mode-related feature block is calculated. For example, through an attention calculation module, which adopts a dot product attention mechanism, takes the current mode-related feature block and the temperature mode-related feature block as input, calculates the similarity between the feature vector of each time step in the current mode-related feature block and the feature vector of the temperature mode-related feature block, and uses the similarity as the weighted influence value.

[0049] The calculated weighted influence values ​​are arranged according to the time step and spatial position to generate the attention distribution diagram from current to temperature. This distribution diagram can intuitively show the distribution of the influence of the current mode-related feature block on the temperature mode-related feature block in time and space. For example, if the weighted influence value is large at a certain time step and spatial position, it means that at this time and position, the current mode-related feature has a greater influence on the temperature mode-related feature.

[0050] Step S243: In the temperature-dominated attention branch, the weighted influence value of each spatial node of the temperature modal associated feature block on the current modal associated feature block is calculated to generate an attention distribution diagram from temperature to current.

[0051] The temperature-dominated attention branch is another branch in the cross-modal attention gating mechanism, which focuses on the influence of the temperature modal-related feature block on the current modal-related feature block. In this branch, the weighted influence value of each spatial node of the temperature modal-related feature block on the current modal-related feature block is calculated. The dot product attention mechanism can also be used to take the temperature modal-related feature block and the current modal-related feature block as input, calculate the similarity between the feature vector of each spatial node in the temperature modal-related feature block and the feature vector of the current modal-related feature block, and use the similarity as the weighted influence value.

[0052] The calculated weighted influence values ​​are arranged according to spatial nodes and time steps to generate a temperature-to-current attention distribution diagram. This distribution diagram can intuitively show the spatial and temporal distribution of the influence of the temperature modal association feature block on the current modal association feature block. For example, if the weighted influence value is large at a certain spatial node and time step, it means that at this position and time, the temperature modal association feature has a greater influence on the current modal association feature.

[0053] Step S244: perform a dot multiplication operation on the attention distribution map from current to temperature and the temperature mode associated feature block to obtain a temperature feature enhancement vector, and perform a dot multiplication operation on the attention distribution map from temperature to current and the current mode associated feature block to obtain a current feature enhancement vector.

[0054] In an embodiment of the present invention, a dot multiplication operation is performed on the attention distribution diagram from current to temperature and the temperature modal association feature block. Through the dot multiplication operation, the influence information of the current modal association feature block on the temperature modal association feature block is integrated into the temperature modal association feature block to obtain a temperature feature enhancement vector. Similarly, a dot multiplication operation is performed on the attention distribution diagram from temperature to current and the current modal association feature block to obtain a current feature enhancement vector. In this way, the association information between different modal features can be enhanced, so that subsequent analysis and processing can more accurately reflect the actual operating conditions of the line.

[0055] Step S245: cross-project and fuse the temperature feature enhancement vector and the current feature enhancement vector to generate an association mapping relationship among the current fluctuation feature sequence, the temperature change feature sequence, and the line aging index sequence.

[0056] In an embodiment of the present invention, the temperature feature enhancement vector and the current feature enhancement vector are cross-projected and fused. For example, a linear projection method can be used to project the temperature feature enhancement vector and the current feature enhancement vector into a common feature space, and then the projected vectors are added or concatenated to obtain a fused feature vector.

[0057] Through cross-projection fusion, the characteristic information of temperature mode and current mode is integrated to generate the correlation mapping relationship between current fluctuation characteristic sequence, temperature change characteristic sequence and line aging index sequence. This correlation mapping relationship can more comprehensively reflect the relationship between the three sequences and provide a more accurate basis for subsequent fire risk prediction.

[0058] Step S250: performing spatial interpolation fusion on the first dynamic feature vector and the second dynamic feature vector based on the association mapping relationship to generate a spatiotemporal association feature matrix containing a timestamp mark.

[0059] In an embodiment of the present invention, based on the previously generated association mapping relationship, the first dynamic feature vector and the second dynamic feature vector are spatially interpolated and fused. The association mapping relationship reflects the relationship between the current fluctuation characteristics, the temperature change characteristics and the line aging index. By utilizing this relationship, the first dynamic feature vector and the second dynamic feature vector can be more reasonably fused. Generate a spatiotemporal association feature matrix containing a timestamp mark, which is to arrange the feature vectors after spatial interpolation and fusion according to the time and space dimensions, and add a timestamp mark to each element. The timestamp mark can record the collection time of each feature data, so that the spatiotemporal association feature matrix can more accurately reflect the operating status of the target charging pile cluster line at different times and spaces.

[0060] As an implementation manner, step S250, based on the association mapping relationship, performs spatial interpolation fusion on the first dynamic feature vector and the second dynamic feature vector to generate a spatiotemporal association feature matrix containing a timestamp mark, which may specifically include the following steps: Step S251: performing timestamp alignment processing on the current fluctuation peak-to-valley difference sequence in the first dynamic feature vector and the temperature gradient variation sequence in the second dynamic feature vector to generate a time-synchronized current-temperature feature pair sequence.

[0061] The timestamp alignment process is to match the data of different time series in the time dimension so that they have a corresponding relationship at the same time point. In an embodiment of the present invention, the current fluctuation peak-to-valley difference sequence in the first dynamic feature vector and the temperature gradient change sequence in the second dynamic feature vector are timestamp aligned. For example, if the sampling time interval of the current fluctuation peak-to-valley difference sequence is 10 minutes and the sampling time interval of the temperature gradient change sequence is 15 minutes, one of the sequences needs to be interpolated or sampled so that the timestamps of the two sequences are consistent.

[0062] The current fluctuation peak-to-valley difference sequence and the temperature gradient change sequence after timestamp alignment are combined into pairs of feature pairs to generate a time-synchronized current-temperature feature pair sequence. This sequence can more accurately reflect the corresponding relationship between current fluctuation and temperature change at the same time point.

[0063] Step S252: extract the interpolation weight distribution of the current fluctuation peak-to-valley difference sequence in the spatial dimension according to the attention distribution diagram of the current to temperature in the association mapping relationship, and generate a current-dominated spatial interpolation mask matrix.

[0064] According to the attention distribution diagram of current to temperature in the association mapping relationship, the distribution diagram reflects the distribution of the degree of influence of the current mode association feature block on the temperature mode association feature block in time and space. In an embodiment of the present invention, the interpolation weight distribution of the current fluctuation peak-valley difference sequence in the spatial dimension is extracted from this attention distribution diagram. For example, the interpolation weight of the current fluctuation peak-valley difference sequence at each spatial position can be obtained by performing a statistical analysis on the attention distribution diagram in the spatial dimension. The extracted interpolation weights are arranged according to the spatial position to generate a current-dominated spatial interpolation mask matrix. This matrix can be used for subsequent spatial interpolation operations to achieve smoothing and fusion of the current fluctuation peak-valley difference sequence in the spatial dimension.

[0065] Step S253: According to the attention distribution diagram from temperature to current in the association mapping relationship, the interpolation weight distribution of the temperature gradient change sequence in the time dimension is extracted to generate a temperature-dominated time interpolation mask matrix.

[0066] Similarly, according to the attention distribution diagram of temperature to current in the association mapping relationship, the distribution diagram reflects the spatial and temporal distribution of the influence of the temperature modal association feature block on the current modal association feature block. In an embodiment of the present invention, the interpolation weight distribution of the temperature gradient change sequence in the time dimension is extracted from this attention distribution diagram. For example, the interpolation weight of the temperature gradient change sequence at each time point can be obtained by performing a statistical analysis on the attention distribution diagram in the time dimension.

[0067] The extracted interpolation weights are arranged according to time points to generate a temperature-dominated time interpolation mask matrix. This matrix can be used in subsequent time interpolation operations to achieve smoothing and fusion of the temperature gradient change sequence in the time dimension.

[0068] Step S254: performing a spatial dimension convolution operation on the current-dominated spatial interpolation mask matrix and the time-synchronized current-temperature feature pair sequence to generate an intermediate fusion matrix with enhanced current features.

[0069] The spatial dimension convolution operation is a method of performing a convolution operation on data in the spatial dimension, which can achieve spatial smoothing and feature extraction of data. In an embodiment of the present invention, a spatial dimension convolution operation is performed on a current-dominated spatial interpolation mask matrix and a time-synchronized current-temperature feature pair sequence. For example, a two-dimensional convolution kernel is used to perform a convolution operation on the current-temperature feature pair sequence in the spatial dimension, and the weight of the convolution kernel is determined by the current-dominated spatial interpolation mask matrix.

[0070] Through spatial dimension convolution operation, the spatial interpolation weight information of the current fluctuation peak-to-valley difference sequence is integrated into the time-synchronized current-temperature feature pair sequence to generate an intermediate fusion matrix with enhanced current features. This matrix enhances the current features in the spatial dimension and can more accurately reflect the spatial distribution of current fluctuations.

[0071] Step S255: superimpose the temperature-dominated time interpolation mask matrix and the current feature enhanced intermediate fusion matrix in a time dimension sliding window to generate a spatiotemporal joint feature tensor.

[0072] The time dimension sliding window superposition is a method for processing data in the time dimension, which can fuse data at different time points. In an embodiment of the present invention, the temperature-dominated time interpolation mask matrix and the current feature enhanced intermediate fusion matrix are superimposed in the time dimension sliding window. For example, a sliding window of a fixed length is used to slide the current feature enhanced intermediate fusion matrix in the time dimension, and the data in the window is weighted summed with the temperature-dominated time interpolation mask matrix each time it slides.

[0073] Through the time dimension sliding window superposition, the time interpolation weight information of the temperature gradient change sequence is integrated into the intermediate fusion matrix of the current feature enhancement to generate a spatiotemporal joint feature tensor. This tensor integrates the feature information in the time and space dimensions, and can more comprehensively reflect the operating status of the target charging pile cluster line.

[0074] The following interpolation operations are performed on the spatiotemporal joint feature tensor: Step S256: In the spatial dimension, according to the cross-projection fusion results in the associated mapping relationship, bidirectional weighted interpolation is performed on the feature vectors of adjacent line segments to generate a spatial continuity feature layer.

[0075] In the spatial dimension, according to the cross-projection fusion result in the association mapping relationship, the result reflects the comprehensive association relationship between the current fluctuation characteristics, the temperature change characteristics and the line aging index. In the embodiment of the present invention, this result is used to perform bidirectional weighted interpolation on the feature vectors of adjacent line segments.

[0076] By performing bidirectional weighted interpolation on the feature vectors of all adjacent line segments, a spatial continuity feature layer is generated. This feature layer smoothes the feature information in the spatial dimension, making the feature changes between adjacent line segments more continuous and more accurately reflecting the overall operation status of the line in space.

[0077] Step S257: In the time dimension, based on the timestamp interval of the original multi-source time series monitoring data, the spatial continuity feature layer is interpolated and completed in the time domain to generate a timestamp-aligned feature distribution map.

[0078] In the time dimension, based on the timestamp interval of the original multi-source time series monitoring data, the interval records the time interval of data collection. In the embodiment of the present invention, the spatial continuity feature layer is interpolated and supplemented in the time domain.

[0079] Through time domain interpolation and completion, a timestamp-aligned feature distribution map is generated. This distribution map completes the feature information in the time dimension, making the feature data more continuous and accurate in time, and can more comprehensively reflect the operation status of the line at different times.

[0080] Step S258: Channel superposition of the spatial continuity feature layer and the feature distribution map aligned with the timestamp is performed to generate a spatiotemporal correlation feature matrix containing timestamp marks, where each timestamp corresponds to the original collection moment of the multi-source time series monitoring data.

[0081] Channel superposition refers to merging different feature layers in the channel dimension. In an embodiment of the present invention, the spatial continuity feature layer and the feature distribution map aligned with the timestamp are channel superimposed. For example, if the spatial continuity feature layer is a two-dimensional matrix and the feature distribution map aligned with the timestamp is also a two-dimensional matrix, then after superimposing them in the channel dimension, a three-dimensional matrix is ​​obtained, which is the spatiotemporal correlation feature matrix containing the timestamp mark.

[0082] Each timestamp corresponds to the original collection moment of the multi-source time series monitoring data, which enables the spatiotemporal correlation feature matrix to accurately reflect the operating status of the target charging pile cluster line at different times and spaces.

[0083] Step S259: Perform edge feature compensation processing on the spatiotemporal correlation feature matrix, detect the feature attenuation area at the matrix boundary, and repair the feature value according to the dynamic weight allocation ratio in the correlation mapping relationship to generate a spatiotemporal correlation feature matrix that completely covers the target charging pile line topology.

[0084] The edge feature compensation process is to compensate for the problem of feature information loss or attenuation that may exist at the boundary of the spatiotemporal correlation feature matrix. In an embodiment of the present invention, the edge feature compensation process is performed on the spatiotemporal correlation feature matrix. First, the feature attenuation area at the matrix boundary is detected, for example, by calculating the difference between the eigenvalue at the matrix boundary and the internal eigenvalue, to determine the location and degree of feature attenuation.

[0085] Then, the feature value is repaired according to the dynamic weight distribution ratio in the association mapping relationship. The dynamic weight distribution ratio in the association mapping relationship reflects the importance and correlation between different features, and this ratio is used to adjust and repair the feature values ​​of the feature attenuation area. For example, for a certain feature attenuation area, according to the weight distribution ratio of the relevant features of the area in the association mapping relationship, the preset information is extracted from the feature values ​​of the adjacent areas to supplement the feature values ​​of the area.

[0086] Through edge feature compensation processing, a spatiotemporal correlation feature matrix that completely covers the target charging pile line topology is generated. This matrix can more accurately reflect the overall operating status of the target charging pile cluster line, providing more reliable data support for subsequent fire risk prediction.

[0087] Step S300: Input the spatiotemporal correlation feature matrix into a pre-trained fire risk prediction model to generate a line abnormality risk probability distribution map. The fire risk prediction model is trained by the mapping relationship between the historical fire event data set and the multi-source monitoring data.

[0088] The pre-trained fire risk prediction model is a trained model that can predict the abnormal risk probability of the line based on the input spatiotemporal correlation feature matrix. The model is trained through the mapping relationship between the historical fire event dataset and the multi-source monitoring data. The historical fire event dataset contains relevant data when fires occurred in the past, such as the time, location, and operating status of the line; the multi-source monitoring data refers to various monitoring data collected from the target charging pile cluster, such as current fluctuation characteristics, temperature change characteristics, etc.

[0089] By establishing a mapping relationship between historical fire event data sets and multi-source monitoring data, the model is trained using machine learning or deep learning algorithms. For example, a model structure such as a convolutional neural network (CNN) or a recurrent neural network (RNN) can be used to take historical fire event data sets and multi-source monitoring data as input. After multiple iterations of training, the model parameters are adjusted so that the model can learn the relationship between fire occurrence and line operation status.

[0090] The spatiotemporal correlation feature matrix is ​​input into the pre-trained fire risk prediction model, and the model predicts the abnormal risk probability of the line based on its internal parameters and learned relationships. The prediction results are presented in the form of a line abnormal risk probability distribution map, which can intuitively show the abnormal risk probability of the target charging pile cluster line at different locations and times. For example, in the distribution map, the darker the color, the higher the abnormal risk probability, and the lighter the color, the lower the abnormal risk probability.

[0091] As an implementation method, the training process of the fire risk prediction model may specifically include the following steps: Step S10: Acquire a historical fire event dataset, which includes multi-source monitoring data slices within a set time range before the fire occurs and corresponding fire location annotation information.

[0092] The historical fire event dataset is an important data source for training the fire risk prediction model. In an embodiment of the present invention, a historical fire event dataset is obtained, which includes multi-source monitoring data slices and corresponding fire location annotation information within a set time range before the fire occurs. The set time range can be determined according to actual conditions, such as 1 hour or 2 hours before the fire occurs.

[0093] Multi-source monitoring data slices refer to various monitoring data collected within a set time range, such as current fluctuation characteristics, temperature change characteristics, line aging indicators, etc. These data slices record the operating status of the line before the fire. The corresponding fire location annotation information clarifies the specific location where the fire occurred, such as the specific number of a charging pile or the specific location of a line segment. For example, in a historical fire event data set, it contains multi-source monitoring data slices within 2 hours before the fire, as well as the annotation information that the fire occurred in the No. 5 charging pile line segment of a certain charging station. By collecting a large amount of such historical fire event data, a complete historical fire event data set is formed.

[0094] Step S20: Perform cross-modal feature alignment processing on multi-source monitoring data slices to generate a historical spatiotemporal correlation feature matrix set.

[0095] The cross-modal feature alignment process is performed on the multi-source monitoring data slices, which is similar to the cross-modal feature alignment process performed on the multi-source time series monitoring data in the previous step S200. Through this process, different types of monitoring data (such as current fluctuation characteristics, temperature change characteristics, line aging indicators, etc.) are aligned and associated in the time and space dimensions.

[0096] The specific processing process includes dynamically adjusting the weight distribution of temperature change characteristics based on the current fluctuation amplitude, and correcting the correlation between current and temperature characteristics based on line aging indicators. Through these processes, a historical spatiotemporal correlation feature matrix set is generated. This matrix set contains the line operation status information at different times and spaces before the historical fire incident, providing more accurate and comprehensive data for subsequent model training.

[0097] Step S30: construct a deep spatiotemporal convolutional network, which includes a local feature extraction branch and a global association branch in parallel, wherein the local feature extraction branch uses a three-dimensional hole convolution kernel to capture local abnormal patterns of the line, and the global association branch uses a graph attention mechanism to model the topological relationship across line segments.

[0098] A deep spatiotemporal convolutional network is a deep learning network structure for processing spatiotemporal data. In an embodiment of the present invention, a deep spatiotemporal convolutional network is constructed, which includes a local feature extraction branch and a global association branch in parallel.

[0099] The local feature extraction branch uses a three-dimensional hole convolution kernel to capture local abnormal patterns of the line. The three-dimensional hole convolution kernel can perform convolution operations on data in three-dimensional space (two dimensions of time and space), and by setting the hole rate, the receptive field can be expanded without increasing the convolution kernel parameters, thereby better capturing the local abnormal patterns of the line. For example, when a sudden increase in current or abnormal increase in temperature occurs in a local area of ​​the line, the local feature extraction branch can extract these abnormal patterns through the three-dimensional hole convolution kernel.

[0100] The global association branch uses the graph attention mechanism to model the topological relationship across line segments. The graph attention mechanism can dynamically assign attention weights based on the relationship between nodes, thereby better modeling the topological relationship across line segments. In this scenario, the charging pile line segments are regarded as nodes of the graph, and the electrical connection relationship between adjacent line segments is regarded as the edge of the graph. The graph attention mechanism can be used to learn the mutual influence and association between different line segments. For example, when an abnormality occurs in a line segment, the graph attention mechanism can be used to analyze the degree of impact of the abnormality on the adjacent line segments.

[0101] Step S40: Input the historical spatiotemporal correlation feature matrix set into the deep spatiotemporal convolutional network for forward propagation, output the predicted fire probability distribution map, and calculate the cross entropy loss with the actual fire location annotation information.

[0102] The historical spatiotemporal correlation feature matrix set is input into the deep spatiotemporal convolutional network for forward propagation. Forward propagation refers to the process in which data passes from the input layer of the network through each hidden layer and finally reaches the output layer. In this process, the local feature extraction branch and global association branch of the deep spatiotemporal convolutional network will process the input data respectively, extract local features and model global topological relationships.

[0103] Output the predicted fire probability distribution map, which reflects the model's prediction results of the probability of historical fire events at different times and spaces. Then, compare the predicted fire probability distribution map with the actual fire location labeling information and calculate the cross entropy loss. Cross entropy loss is a loss function used to measure the difference between the predicted result and the actual label. By calculating the cross entropy loss, the prediction accuracy of the model can be evaluated. For example, if the predicted probability of a certain location in the predicted fire probability distribution map is inconsistent with the actual fire location labeling information, the cross entropy loss will increase.

[0104] Step S50: Iteratively optimize the parameters of the deep spatiotemporal convolutional network through the back propagation algorithm until the spatial overlap between the predicted fire probability distribution map and the actual annotation reaches a preset threshold.

[0105] In an embodiment of the present invention, the parameters of the deep spatiotemporal convolutional network are iteratively optimized by a back-propagation algorithm. In each iteration, the gradient of the loss function to the network parameters is calculated based on the calculated cross entropy loss, and then the network parameters are updated using an optimization algorithm (such as a stochastic gradient descent method). The iterative process is repeated until the spatial overlap between the predicted fire probability distribution map and the actual annotation reaches a preset threshold. The spatial overlap can be measured by calculating the degree of spatial overlap between the predicted fire probability distribution map and the actual fire location annotation information. The preset threshold can be determined based on actual conditions. For example, when the spatial overlap reaches 80%, it is considered that the prediction result of the model is sufficiently accurate and the iterative optimization is stopped.

[0106] As an implementation, the global association branch of the deep spatiotemporal convolutional network performs the following processing: Step S31: construct a line topology diagram according to the physical connection relationship of the charging pile line, the node of the line topology diagram represents a single charging pile line segment, and the edge represents the electrical connection relationship between adjacent line segments.

[0107] The line topology diagram is a graph structure used to represent the connection relationship between charging pile lines. In an embodiment of the present invention, a line topology diagram is constructed based on the physical connection relationship of the charging pile lines. A single charging pile line segment is regarded as a node of the graph, and the electrical connection relationship between adjacent line segments is regarded as an edge of the graph. For example, in a charging station containing multiple charging piles, the line segments of the incoming and outgoing ends of each charging pile can be regarded as a node. If there is an electrical connection between the two line segments, then an edge is added between their corresponding nodes.

[0108] By constructing a line topology graph, the topological structure of the charging pile line can be more intuitively represented, providing a basis for subsequent graph attention mechanism processing.

[0109] Step S32: assigning an initial feature vector to each node, the initial feature vector including the current fluctuation characteristics, temperature change characteristics and aging index statistics corresponding to the charging pile line segment.

[0110] Assign an initial feature vector to each node of the line topology diagram. The initial feature vector contains the current fluctuation characteristics, temperature change characteristics and aging index statistics corresponding to the charging pile line segment. For example, the statistics of the current fluctuation characteristics may include the average value and standard deviation of the current; the statistics of the temperature change characteristics may include the maximum and minimum values ​​of the temperature; the statistics of the aging index may include the average value of the degradation coefficient of the conductive material, the average value of the insulation layer damage rate, etc.

[0111] By assigning an initial feature vector to each node, the running status information of the line can be integrated into the graph structure, so that the graph attention mechanism can better process this information.

[0112] Step S33: Iteratively update the node features through a multi-layer graph attention network, where each layer of the graph attention network performs the following operations: Step S331: Calculate the attention coefficient between the current node and its neighboring nodes, and the attention coefficient is dynamically adjusted based on the current phase difference and temperature conduction rate between the nodes.

[0113] In a multi-layer graph attention network, each layer of the graph attention network first calculates the attention coefficient between the current node and its neighboring nodes. The attention coefficient is used to measure the degree of attention paid by the current node to the characteristics of the neighboring nodes. In an embodiment of the present invention, the attention coefficient is dynamically adjusted based on the current phase difference and temperature conduction rate between nodes.

[0114] The current phase difference between nodes reflects the phase difference of the current between adjacent line segments, and the temperature conduction rate reflects the temperature transfer speed between adjacent line segments. For example, if the current phase difference between two adjacent nodes is large, it means that their current operating states may be quite different, and the attention coefficient may be relatively small; if the temperature conduction rate between two adjacent nodes is fast, it means that the heat transfer relationship between them is relatively close, and the attention coefficient may be relatively large.

[0115] By dynamically adjusting the attention coefficient based on the current phase difference and temperature conduction rate between nodes, the mutual influence relationship between adjacent line segments can be more accurately reflected.

[0116] Step S332: Perform weighted aggregation on neighbor node features according to the attention coefficient, and perform residual connection with the current node features.

[0117] According to the calculated attention coefficient, the features of neighbor nodes are weighted and aggregated. Then, the feature vector after weighted aggregation is residually connected with the feature vector of the current node. Residual connection is a commonly used technique in deep learning, which can alleviate the gradient vanishing problem and make the network easier to train. Specifically, the feature vector after weighted aggregation is added to the feature vector of the current node to obtain a new node feature vector.

[0118] Step S334: input the updated node features into the gated linear unit for nonlinear transformation.

[0119] In an embodiment of the present invention, the updated node features are input into a gated linear unit for nonlinear transformation. GLU can control the transmission of input information through a gating mechanism, thereby realizing nonlinear processing of node features. For example, GLU can dynamically adjust the output feature values ​​according to the input node features, making the feature information richer and more complex. Through the nonlinear transformation of the gated linear unit, the expressive power of the node features can be enhanced, and the performance of the graph attention network can be improved.

[0120] Step S34: Channel-join the node features output by the last layer of graph attention network with the output features of the local feature extraction branch to generate a fused global-local joint feature.

[0121] The node features output by the last layer of the graph attention network are channel-joined with the output features of the local feature extraction branch. Channel jointing is to merge different feature vectors in the channel dimension. Through channel jointing, the global topological relationship information extracted by the global association branch is integrated with the local abnormal pattern information extracted by the local feature extraction branch to generate a fused global-local joint feature. This feature can more comprehensively reflect the operating status of the charging pile line and provide a more accurate basis for subsequent fire risk prediction.

[0122] Step S400: Generate a hierarchical warning signal set according to the risk level corresponding to each spatial node in the line abnormality risk probability distribution diagram, and the hierarchical warning signal set includes differentiated control instructions for different charging pile line segments.

[0123] The line abnormality risk probability distribution diagram shows the abnormal risk probability of the target charging pile cluster line at different locations and times. In an embodiment of the present invention, a hierarchical warning signal set is generated according to the risk level corresponding to each spatial node in the line abnormality risk probability distribution diagram. The risk level can be divided according to the size of the abnormal risk probability, for example, the risk level is divided into three levels: high, medium, and low. The hierarchical warning signal set contains differentiated control instructions for different charging pile line segments. Different risk levels correspond to different control instructions. For example, for line segments with high risk levels, more stringent control measures may need to be taken, such as cutting off the power supply; for line segments with medium risk levels, measures such as adjusting the load may need to be taken; for line segments with low risk levels, only monitoring and early warning prompts may be required.

[0124] By generating a hierarchical warning signal set, different control measures can be taken according to the actual risk situation of the line to improve the pertinence and effectiveness of fire warning.

[0125] As an implementation mode, step S400 generates a hierarchical warning signal set according to the risk level corresponding to each spatial node in the line abnormality risk probability distribution diagram, which may specifically include the following steps: Step S410: Divide the line abnormality risk probability distribution map into a plurality of risk level areas, each risk level area corresponds to a preset early warning response strategy.

[0126] The line abnormal risk probability distribution map is divided into multiple risk level areas. For example, the distribution map can be divided into high-risk areas, medium-risk areas and low-risk areas according to the size of the abnormal risk probability. Each risk level area corresponds to a preset early warning response strategy, which specifies the specific measures to be taken at that risk level. For example, for a high-risk area, the preset early warning response strategy may be to immediately cut off the power supply to the area and notify relevant personnel to conduct inspections and repairs; for a medium-risk area, the early warning response strategy may be to adjust the load in the area and reduce the operating pressure of the line; for a low-risk area, the early warning response strategy may be to strengthen monitoring of the area and issue an early warning prompt.

[0127] Step S420: For the first risk level area, extract the real-time current spectrum characteristics of the line segment therein to detect whether there is a harmonic resonance phenomenon; if harmonic resonance is detected, generate a first-level warning signal including a harmonic suppression instruction.

[0128] For the first risk level area (e.g., high risk area), the real-time current spectrum characteristics of the line segment are extracted. The real-time current spectrum characteristics reflect the distribution of the frequency components of the current in the line. The real-time current spectrum characteristics can be obtained by performing Fourier transform on the current signal in the line.

[0129] Detect whether there is harmonic resonance. Harmonic resonance is when the harmonic frequency in the line matches the inherent resonant frequency of the line, resulting in a sharp increase in harmonic energy, which may cause damage to the line. The presence of harmonic resonance can be detected by calculating the matching degree between the harmonic frequency in the real-time current spectrum characteristics and the inherent resonant frequency spectrum of the line.

[0130] If harmonic resonance is detected, a first-level warning signal containing a harmonic suppression instruction is generated. The harmonic suppression instruction is used to suppress the harmonic components in the line, for example, by starting a filter. The first-level warning signal will promptly notify relevant personnel to take measures to avoid damage to the line caused by harmonic resonance.

[0131] Step S430: For the second risk level area, calculate the deviation between the temperature change rate and the current load of the line segment in the area, and generate a second level warning signal including a load balancing instruction when the deviation exceeds a dynamic threshold.

[0132] For the second risk level area, the deviation between the temperature change rate and the current load of the line section in the area is calculated. The temperature change rate reflects how fast the line temperature changes over time, and the current load reflects the magnitude of the current in the line. The deviation is used to measure the difference between the temperature change rate and the current load. For example, the deviation can be obtained by calculating the ratio of the temperature change rate to the current load and comparing it with a standard ratio.

[0133] The dynamic threshold is a threshold that is dynamically adjusted according to the actual operation of the line. When the deviation exceeds the dynamic threshold, it means that the operation status of the line may be abnormal, and a second-level warning signal containing a load balancing instruction is generated. The load balancing instruction is used to adjust the load of the line so that the load of each line segment is more balanced, thereby reducing the operation risk of the line. For example, load balancing can be achieved by adjusting the charging power of the charging pile.

[0134] Step S440: For the third risk level area, monitor the cumulative change of the line aging index, and generate a third level warning signal including a preventive maintenance prompt when the cumulative change reaches a preset critical value.

[0135] For the third risk level areas (such as low risk areas), monitor the cumulative change of line aging indicators. The cumulative change of line aging indicators reflects the gradual increase in the degree of line aging. For example, the cumulative change of line aging indicators is obtained by continuously accumulating the changes in the degradation coefficient of the conductive material and the insulation layer breakage rate. The preset critical value is a threshold value determined based on factors such as the design life and safety standards of the line. When the cumulative change reaches the preset critical value, it means that the line may be close to the aging limit and preventive maintenance is required. At this time, a third-level early warning signal containing preventive maintenance prompts is generated to prompt relevant personnel to inspect and maintain the line to ensure the safe operation of the line.

[0136] Step S450: topologically sort the first-level warning signal, the second-level warning signal and the third-level warning signal according to their spatial positions to generate a hierarchical warning signal set that matches the physical layout of the charging pile line.

[0137] The first-level warning signal, the second-level warning signal, and the third-level warning signal are topologically sorted according to the spatial position. Topological sorting refers to sorting the warning signals according to the physical connection relationship and spatial position of the line. For example, sorting is performed according to the number or geographical location of the line segment so that the order of the warning signals matches the physical layout of the charging pile line.

[0138] Through topological sorting, a hierarchical warning signal set is generated that matches the physical layout of the charging pile line. This signal set can be more easily matched with the actual charging pile line, allowing relevant personnel to quickly and accurately take corresponding control measures based on the prompts of the signal set.

[0139] As an implementation manner, in step S420, detecting whether there is a harmonic resonance phenomenon may specifically include: Step S421: Acquire current waveform sampling data of the target line segment, and extract the frequency-amplitude pair set of the fundamental component and each harmonic component through time-frequency conversion processing.

[0140] Get the current waveform sampling data of the target line segment. You can use a current sensor to sample the current waveform of the target line segment to obtain a series of discrete current values. Time-frequency conversion processing is a method of converting the current waveform data in the time domain to the frequency domain for analysis. Common time-frequency conversion methods include fast Fourier transform (FFT).

[0141] Through time-frequency conversion processing, the current waveform sampling data of the target line section is converted to the frequency domain, and the frequency-amplitude pair set of the fundamental component and each harmonic component is extracted. The fundamental component refers to the component with the lowest frequency in the current waveform, and each harmonic component refers to the component whose frequency is an integer multiple of the fundamental frequency. The frequency-amplitude pair set records the frequency and corresponding amplitude information of each harmonic component.

[0142] Step S422: Calculate the matching degree between the harmonic frequencies in the frequency-amplitude pair set and the inherent resonant frequency spectrum of the line to generate the frequency offset and resonance tendency index corresponding to each harmonic component.

[0143] The matching degree of the harmonic frequencies in the frequency-amplitude pair set and the line's natural resonant frequency spectrum is calculated. The line's natural resonant frequency spectrum refers to the set of resonant frequencies that the line itself has, and these frequencies are determined by the line's physical parameters (such as inductance, capacitance, etc.). The matching degree calculation can be performed by calculating the difference between the harmonic frequency and the line's natural resonant frequency.

[0144] Based on the matching calculation results, the frequency offset and resonance tendency index corresponding to each harmonic component are generated. The frequency offset refers to the difference between the harmonic frequency and the inherent resonant frequency of the line, reflecting the proximity between the harmonic frequency and the resonant frequency. The resonance tendency index is an indicator calculated based on factors such as frequency offset and harmonic amplitude, which is used to measure the possibility of resonance of the harmonic component. For example, if the frequency offset is small and the harmonic amplitude is large, the resonance tendency index will be high.

[0145] Step S423: Screening the potentially dangerous harmonic components based on the resonance tendency index, and inputting the amplitudes of the potentially dangerous harmonic components into the energy accumulation predictor in time series to generate a harmonic energy accumulation trajectory.

[0146] Screening of potentially dangerous harmonic components based on the resonance tendency index. A resonance tendency index threshold can be set. When the resonance tendency index of a harmonic component exceeds the threshold, it is considered a potentially dangerous harmonic component. The amplitude of the potentially dangerous harmonic component is input into the energy accumulation predictor in time series. The energy accumulation predictor is a model for predicting the accumulation of harmonic energy. It can predict the cumulative change of harmonic energy over time based on the input harmonic amplitude time series.

[0147] The energy accumulation predictor generates a harmonic energy accumulation trajectory. This trajectory reflects the accumulation of energy of potentially dangerous harmonic components over a period of time. For example, if the harmonic energy accumulation trajectory shows a rapid upward trend, it means that the harmonic energy is accumulating and may cause harmonic resonance.

[0148] Step S424: Perform trend analysis on the harmonic energy accumulation trajectory, and when it is detected that the continuous rising slope exceeds the safety threshold, trigger the multi-band interference detection process.

[0149] Perform trend analysis on the trajectory of harmonic energy accumulation. Trend analysis can be performed by calculating the slope of the trajectory and other methods to determine whether the trend of harmonic energy accumulation is rising, falling, or stable. The safety threshold is a threshold determined based on factors such as line safety standards and design requirements, and is used to determine whether the rate of increase of harmonic energy accumulation is too fast.

[0150] When the continuous rising slope is detected to exceed the safety threshold, it means that the harmonic energy is accumulating too fast, which may cause harmonic resonance. At this time, the multi-band interference detection process is triggered. The multi-band interference detection process is used to further analyze the interaction between harmonics in different frequency bands to determine whether there is a risk of harmonic resonance.

[0151] Step S425: extracting harmonic phase difference sequences of adjacent frequency bands, and calculating phase synchronization index and energy superposition effect coefficient.

[0152] In the multi-band interference detection process, the harmonic phase difference sequence of adjacent frequency bands is extracted. The harmonic phase difference sequence records the phase difference between the harmonics of adjacent frequency bands. The harmonic phase difference sequence is obtained by performing phase analysis on the harmonic signals of adjacent frequency bands.

[0153] Calculate the phase synchronization index and energy superposition effect coefficient. The phase synchronization index is used to measure the phase synchronization degree between harmonics in adjacent frequency bands. If the phase synchronization degree is high, it means that the harmonics in adjacent frequency bands may be superimposed on each other, increasing the harmonic energy. The energy superposition effect coefficient is an index calculated based on factors such as the phase synchronization index and the harmonic amplitude, and is used to measure the degree of harmonic energy superposition in adjacent frequency bands. For example, if the phase synchronization degree is high and the harmonic amplitude is large, the energy superposition effect coefficient will be high.

[0154] Step S426: input the phase synchronization index and the energy superposition effect coefficient into the resonance risk classifier to generate a comprehensive resonance risk level.

[0155] The phase synchronization index and the energy superposition effect coefficient are input into the resonance risk classifier. The resonance risk classifier is a trained classification model that can classify the risk of harmonic resonance according to the input phase synchronization index and energy superposition effect coefficient. For example, the resonance risk classifier can classify the risk level into three levels: high, medium, and low.

[0156] The resonance risk classifier generates a comprehensive resonance risk level, which takes into account the phase synchronization and energy superposition effect between harmonics in adjacent frequency bands, and can more accurately assess the risk level of harmonic resonance.

[0157] Step S427: adjusting the response strength parameter of the harmonic suppression instruction according to the comprehensive resonance risk level, the response strength parameter including the filter startup priority and the attenuation depth gradient.

[0158] Adjust the response strength parameters of the harmonic suppression command according to the comprehensive resonance risk level. The response strength parameters include the filter startup priority and the attenuation depth gradient. The filter startup priority is used to determine the startup order of the filter. When the comprehensive resonance risk level is high, the filter startup priority can be increased to suppress harmonics faster. The attenuation depth gradient is used to control the filter's attenuation of harmonics. When the comprehensive resonance risk level is high, the attenuation depth gradient can be increased to suppress harmonics more effectively.

[0159] For example, if the comprehensive resonance risk level is high, the filter startup priority can be set to the highest value, and the attenuation depth gradient can be set to a larger value; if the comprehensive resonance risk level is medium, the filter startup priority is set to medium, and the attenuation depth gradient is appropriately adjusted; when the comprehensive resonance risk level is low, the filter startup priority is lowered, and the attenuation depth gradient is reduced. Exemplarily, for a high risk level, the filter startup priority can be set to immediately start all available filters, and the attenuation depth gradient is set to attenuate 3dB (voltage) per 100 Hz; for a medium risk level, the priority can be set to start some key filters first, and the attenuation depth gradient is attenuated 2dB (voltage) per 100 Hz; at a low risk level, the filter is started only when necessary, and the attenuation depth gradient is attenuated 1dB (voltage) per 100 Hz.

[0160] Step S428: Bind the adjusted response intensity parameter to the harmonic suppression instruction in the first-level warning signal to generate an optimized suppression instruction set that dynamically adapts to the electrical characteristics of the target line segment.

[0161] The adjusted response intensity parameter is parameter-bound with the harmonic suppression instruction in the first-level warning signal. In actual operation, the response intensity parameter can be associated with the harmonic suppression instruction in the form of a data structure, for example, in the form of a key-value pair, with the filter start priority and attenuation depth gradient as the key, and the corresponding specific parameter value as the value, integrated with the harmonic suppression instruction.

[0162] Generate an optimized suppression instruction set that dynamically adapts to the electrical characteristics of the target line segment. The electrical characteristics of the target line segment include parameters such as resistance, inductance, and capacitance of the line, which affect the propagation and characteristics of harmonics. By combining the electrical characteristics of the target line segment, the bound instructions are further optimized. For example, for line segments with larger resistance, it may be necessary to place more emphasis on the suppression of high-frequency harmonics, and the filter startup priority and attenuation depth gradient for high-frequency harmonics can be appropriately increased; for line segments with larger inductance, low-frequency harmonics may be more likely to resonate, and the suppression instructions for low-frequency harmonics can be optimized. The optimized suppression instruction set generated in this way can more accurately suppress harmonics based on the actual situation of the target line segment and improve the suppression effect.

[0163] Step S500: triggering a dynamic protection mechanism based on the hierarchical warning signal set, the dynamic protection mechanism including performing a current cutoff operation on a high-risk line segment and performing a power attenuation operation on an adjacent line segment.

[0164] The hierarchical warning signal set contains differentiated control instructions for line segments of different risk levels, based on which the dynamic protection mechanism is triggered. The core goal of the dynamic protection mechanism is to take timely measures to avoid accidents such as fire when abnormal risks are found in the line. By performing current cutoff operations on high-risk line segments and power attenuation operations on adjacent line segments, the risk of the entire charging pile cluster line is reduced.

[0165] When the dynamic protection mechanism is triggered, the system will first parse the hierarchical warning signal set to determine the line segment and specific control instructions corresponding to each warning signal. For high-risk line segments, the current cutoff operation is immediately performed, which can be achieved by controlling devices such as circuit breakers to quickly cut off the current supply to the line segment to prevent possible overheating, short circuits, etc. from further deteriorating. At the same time, in order to avoid excessive impact on adjacent line segments caused by current cutoff, power attenuation operations are performed on adjacent line segments. By adjusting the load of adjacent line segments, their power consumption is reduced to maintain the stability of the entire line system.

[0166] As an implementation manner, in step S500, triggering a dynamic protection mechanism based on a hierarchical warning signal set may specifically include the following steps: Step S510: parse the line segment identifiers and warning response strategy parameters corresponding to each warning signal in the hierarchical warning signal set, extract the line segment set covered by the first-level warning signal containing the harmonic suppression instruction, and the electrical connection topology relationship corresponding to the line segment set.

[0167] Parse the line segment identifier and warning response strategy parameters corresponding to each warning signal in the hierarchical warning signal set. Each warning signal in the hierarchical warning signal set is associated with a line segment and contains corresponding warning response strategy parameters, such as specific parameters of harmonic suppression instructions, adjustment amplitude of load balancing instructions, etc. The parsing process can be achieved by encoding and decoding the signal set, for example, encoding the signal using a preset protocol, decoding it according to the protocol rules during parsing, and extracting the line segment identifier and warning response strategy parameters.

[0168] Extract the line segment set covered by the first-level warning signal containing the harmonic suppression instruction, and the electrical connection topology relationship corresponding to the line segment set. The warning signal can be screened to find the first-level warning signal containing the harmonic suppression instruction, and the line segment identification covered by it can be extracted from it to form a line segment set. The electrical connection topology relationship describes the connection mode and mutual relationship between these line segments, which can be obtained through a pre-built line topology map, which records the connection information of each line segment, such as the identification and connection mode of adjacent line segments.

[0169] Step S520: Send a current cutoff instruction to the target charging pile in the line segment set, the current cutoff instruction includes a cutoff phase angle range and duration parameters generated based on the harmonic suppression instruction, and collects in real time the current waveform distortion rate and temperature gradient change after the cutoff operation is performed.

[0170] Send a current cutoff instruction to the target charging pile in the line segment set. When sending the instruction, the instruction can be transmitted to the controller of the target charging pile through the communication network to ensure that the instruction can be accurately conveyed. The current cutoff instruction contains the cutoff phase angle range and duration parameters generated based on the harmonic suppression instruction. The cutoff phase angle range refers to the phase interval of the current waveform in which the cutoff operation is performed. This can be determined according to the requirements of harmonic suppression. For example, it can be selected to cut off in the phase interval with a larger harmonic amplitude. The duration parameter specifies the duration of the current cutoff to avoid unnecessary impact on the charging pile and line due to long-term cutoff.

[0171] Real-time acquisition of the current waveform distortion rate and temperature gradient change after the truncation operation. The current waveform and temperature can be monitored in real time using current sensors and temperature sensors, respectively. The current waveform distortion rate reflects the degree of deviation of the current waveform from the ideal sine wave. The distortion rate is calculated by analyzing the collected current waveform. The temperature gradient change refers to the rate of change of temperature before and after the truncation operation. The temperature gradient change is calculated by comparing the temperature values ​​at different time points. These data can be used to evaluate the effect of the current truncation operation and the real-time status of the line.

[0172] Step S530: When the current waveform distortion rate exceeds a preset distortion threshold or the temperature gradient change does not reach the expected drop, the power attenuation compensation process of the adjacent line segment is activated.

[0173] When the current waveform distortion rate exceeds the preset distortion threshold or the temperature gradient change does not reach the expected reduction, it means that the current cutoff operation may not achieve the expected effect and the line still has a high risk. The preset distortion threshold is a threshold determined according to the safety standards and design requirements of the line. When the current waveform distortion rate exceeds this threshold, it indicates that the degree of distortion of the current waveform is too large and may cause damage to the line and equipment. The expected reduction refers to the reduction in the temperature gradient change that should be achieved after the current cutoff operation is performed. If the temperature gradient change does not reach this reduction, it means that the heating of the line is not effectively controlled.

[0174] In this case, the power reduction compensation process of the adjacent line section is activated. The purpose of this process is to reduce the load pressure of the line by reducing the power consumption of the adjacent line section, thereby further reducing the risk of the line.

[0175] Step S540: determining a propagation path sequence of power attenuation according to the electrical connection topology relationship, and dynamically allocating a power attenuation ratio of each path node based on the load balancing instruction in the early warning response strategy parameters.

[0176] The propagation path sequence of power attenuation is determined based on the electrical connection topology. The electrical connection topology describes the connection mode and mutual relationship between line segments. By analyzing the topology relationship, the propagation path of power attenuation between adjacent line segments can be determined. For example, if line segment A is adjacent to line segments B and C, and B and C are connected to other line segments respectively, then the propagation path of power attenuation can be determined as AB-... and AC-... based on these connection relationships.

[0177] The power attenuation ratio of each path node is dynamically allocated based on the load balancing instruction in the early warning response strategy parameters. The load balancing instruction specifies how to distribute the load between different line segments to achieve the purpose of load balancing. According to the instruction, the power attenuation ratio of each path node can be calculated. For example, if the load balancing instruction requires that the total power attenuation be distributed to each adjacent line segment according to a preset ratio, the power attenuation ratio of each path node can be dynamically adjusted according to factors such as the load capacity and importance of the line segment.

[0178] Step S550: convert the power attenuation ratio into a pulse width modulation waveform, inject it into the power regulation loop of the corresponding charging pile, and generate a step attenuation signal synchronized with the cut-off phase angle range.

[0179] Convert the power attenuation ratio into a pulse width modulation (PWM) waveform. A PWM waveform is a waveform that controls the average power by adjusting the width of the pulse. The duty cycle of the PWM waveform, that is, the ratio of the pulse width to the period, can be calculated based on the power attenuation ratio. For example, if the power attenuation ratio is 50%, the duty cycle of the PWM waveform can be set to 50%. Inject the generated PWM waveform into the power regulation loop of the corresponding charging pile. The power regulation loop is a circuit used in the charging pile to control the power output. By injecting a PWM waveform, the power output of the charging pile can be regulated. Generate a step attenuation signal synchronized with the truncation phase angle range, that is, the starting and ending phases of the step attenuation signal are consistent with the truncation phase angle range in the current truncation instruction. This ensures that the power attenuation operation is performed synchronously with the current truncation operation, thereby improving the effect of the dynamic protection mechanism.

[0180] Step S560: Collect line status feedback data under the action of the step attenuation signal, which may specifically include the following steps: current harmonic component distribution and temperature field uniformity index after step attenuation, and input them into the fire risk prediction model for real-time risk reassessment.

[0181] Collect line status feedback data under the action of step attenuation signal. Through various sensors installed on the line, such as current sensors, temperature sensors, etc., collect the distribution of current harmonic components and temperature field uniformity index after attenuation in real time. The distribution of current harmonic components reflects the content and distribution of harmonics of different frequencies in the line after the step attenuation signal, which can be obtained by performing spectrum analysis on the collected current signal. The temperature field uniformity index describes the uniformity of the spatial distribution of line temperature, which can be evaluated by installing temperature sensors at different positions, measuring the temperature values ​​of each point, and then calculating the standard deviation of the temperature and other statistics. Input the collected line status feedback data into the fire risk prediction model for real-time risk re-evaluation. The fire risk prediction model is a previously trained model that can predict the abnormal risk probability of the line based on the input line status data. By inputting the line status feedback data in real time, the model can timely update the assessment results of the line risk and provide a basis for subsequent decision-making.

[0182] Step S570: Update the line abnormality risk probability distribution map according to the real-time risk reassessment result. If a diffusion trend of the harmonic energy accumulation trajectory is detected, the secondary protection strategy is triggered.

[0183] Update the line abnormality risk probability distribution map based on the real-time risk reassessment results. The line abnormality risk probability distribution map shows the abnormal risk probability of the target charging pile cluster line at different locations and times. The real-time risk reassessment results are merged with the original distribution map, and the risk probability value of each spatial node is updated so that the distribution map can more accurately reflect the current risk status of the line. If a diffusion trend of the harmonic energy accumulation trajectory is detected, the secondary protection strategy is triggered. The diffusion trend of the harmonic energy accumulation trajectory indicates that the harmonic energy is increasing and spreading to the surrounding line segments, which may cause more serious problems. The secondary protection strategy may include further strengthening of protection measures for high-risk line segments, such as increasing the range of current cutoff, power attenuation of more adjacent line segments, etc., to prevent further spread of harmonic energy and the occurrence of accidents.

[0184] Step S580: performing a voltage amplitude limiting operation on key nodes in the propagation path sequence to suppress the conduction path of abnormal harmonic components.

[0185] Perform voltage amplitude limiting operations on key nodes in the propagation path sequence. Key nodes refer to line segments or nodes that play an important role in harmonic conduction in the power attenuation propagation path. By limiting the voltage amplitude of these key nodes, the conduction path of abnormal harmonic components can be effectively suppressed. Voltage amplitude limiting can be achieved by using devices such as voltage regulators to control the voltage amplitude of key nodes within a safe range. For example, a voltage upper limit value is set. When the voltage of the key node exceeds the upper limit value, the voltage regulator automatically adjusts the voltage to keep it within a safe range.

[0186] Step S590: after the voltage amplitude limiting operation takes effect, the power attenuation ratio is recalculated, and a high-frequency oscillation suppression component is superimposed to eliminate residual harmonic interference.

[0187] After the voltage amplitude limit operation takes effect, the power attenuation ratio is recalculated. Since the voltage amplitude limit affects the power transmission and load conditions of the line, the power attenuation ratio needs to be recalculated according to the new line state to ensure the effectiveness of the power attenuation operation. The power attenuation ratio of each path node can be recalculated based on the line state data collected in real time, such as current, voltage, etc., combined with the load balancing instruction.

[0188] Superimpose high-frequency oscillation suppression components to eliminate residual harmonic interference. High-frequency oscillation refers to high-frequency harmonic components that may still exist after the voltage amplitude limiting operation, which may cause interference to the line and equipment. By superimposing high-frequency oscillation suppression components in the power regulation loop, such as using filters and other equipment, high-frequency harmonics can be further suppressed, residual harmonic interference can be eliminated, and the stability of the line can be improved.

[0189] Step S5100: perform timing alignment calibration on the updated power attenuation parameter and the current cutoff instruction to generate a dynamic protection instruction sequence including a multi-stage protection logic connection relationship.

[0190] The updated power decay parameters are time-aligned and calibrated with the current cut-off instruction. Timing alignment calibration is to ensure that the power decay operation and the current cut-off operation are coordinated in time to avoid operation conflicts or large time differences. The updated power decay parameters and the time parameters of the current cut-off instruction can be compared and adjusted to match their execution times.

[0191] Generate a dynamic protection instruction sequence containing the connection relationship of multi-stage protection logic. The dynamic protection mechanism includes multiple stages of protection measures, such as current cutoff, power attenuation, voltage amplitude limitation, etc. These measures need to have a reasonable connection relationship to ensure the effectiveness and stability of the entire protection process. Integrate the updated power attenuation parameters and current cutoff instructions to generate a dynamic protection instruction sequence containing the connection relationship of multi-stage protection logic, clarify the operation sequence, time interval and specific parameters of each stage, and provide clear guidance for subsequent control operations.

[0192] Step S5110: Control the target charging pile cluster to perform coordinated protection operations through a dynamic protection instruction sequence, and continuously monitor the dynamic response curve of the current waveform distortion rate and the temperature gradient change.

[0193] The target charging pile cluster is controlled to perform coordinated protection operations through a dynamic protection instruction sequence. The generated dynamic protection instruction sequence is sent to each controller of the target charging pile cluster. The controllers perform current cutoff, power attenuation, voltage amplitude limitation and other operations in sequence according to the requirements of the instruction sequence to achieve coordinated protection of the target charging pile cluster.

[0194] Continuously monitor the dynamic response curve of the current waveform distortion rate and temperature gradient change. By collecting current waveform and temperature data in real time, calculate the current waveform distortion rate and temperature gradient change, and plot their changes over time into a dynamic response curve. By analyzing the dynamic response curve, the effect of the coordinated protection operation can be evaluated and whether the risk of the line is effectively controlled. If the dynamic response curve shows that the current waveform distortion rate and temperature gradient change gradually decrease and tend to stabilize, it means that the coordinated protection operation has achieved good results; if the curve shows abnormal fluctuations or continues to rise, it is necessary to adjust the protection strategy in time to further reduce the risk of the line.

[0195] It should be noted that, in the process of reading the above-mentioned embodiments of the invention, those skilled in the art can implement the technical details that are not detailed according to their own technical knowledge in the field without obstacles. For example, when it comes to the calculation of variables of different dimensions, those skilled in the art can use general normalization or standardization means to uniformly eliminate the dimension difference and then perform subsequent operations. For another example, for the scenes not involved, the technical means disclosed in the present invention can be used to continue adaptive extension. For example, step S520 requires "truncation phase angle range" control. If the charging pile lacks fast phase control hardware in some scenes, the instruction cannot be executed. The "truncation phase angle" can be changed to time window truncation (such as cutting off the current within 10ms) to adapt to the response capability of conventional circuit breakers. For calculations with non-uniform dimensions, for example, the number of topological nodes (such as 10) is different from the number of sensors (such as 20), the sensor data can be aggregated to the topological node dimension through principal component analysis (PCA) to ensure spatial consistency, or the topological graph structure can be directly processed by graph convolution network (GCN) to avoid explicit spatial interpolation. As for the specific numerical values ​​of the parameter adjustment examples, more reasonable numerical values ​​can also be selected according to actual conditions. The present invention is only an example and is not limited to this.

[0196] Based on the foregoing embodiments, the embodiments of the present application provide a charging pile line fire warning device based on big data. The units included in the device and the modules included in each unit can be implemented by a processor in a computer device; of course, it can also be implemented by a specific logic circuit; in the implementation process, the processor can be a central processing unit (CPU), a microprocessor (MPU), a digital signal processor (DSP) or a field programmable gate array (FPGA), etc.

[0197] Figure 2 A schematic diagram of the composition structure of a charging pile line fire warning device based on big data provided in an embodiment of the present application is shown in FIG. Figure 2 As shown, the charging pile line fire warning device 200 based on big data includes: The data acquisition module 210 is used to collect the line operation data set of the target charging pile cluster, and the line operation data set includes multi-source time series monitoring data; A feature alignment module 220 is used to perform cross-modal feature alignment on multi-source time series monitoring data to generate a spatiotemporal correlation feature matrix, wherein the cross-modal feature alignment includes dynamically adjusting the weight allocation of temperature change features based on the current fluctuation amplitude, and correcting the correlation between current and temperature features based on the line aging index; The risk prediction module 230 is used to input the spatiotemporal correlation feature matrix into a pre-trained fire risk prediction model to generate a line abnormality risk probability distribution map, and the fire risk prediction model is trained by the mapping relationship between the historical fire event data set and the multi-source monitoring data; The warning signal generation module 240 is used to generate a hierarchical warning signal set according to the risk level corresponding to each spatial node in the line abnormality risk probability distribution diagram, and the hierarchical warning signal set includes differentiated control instructions for different charging pile line segments; The protection triggering module 250 is used to trigger a dynamic protection mechanism based on the hierarchical warning signal set, wherein the dynamic protection mechanism includes performing a current interruption operation on a high-risk line segment and performing a power reduction operation on an adjacent line segment.

[0198] The description of the above device embodiment is similar to the description of the above method embodiment, and has similar beneficial effects as the method embodiment. In some embodiments, the functions or modules provided by the device provided in the embodiment of the present application can be used to execute the method described in the above method embodiment. For the technical details not disclosed in the device embodiment of the present application, please refer to the description of the method embodiment of the present application for understanding. It should be noted that in the embodiment of the present application, if the above-mentioned charging pile line fire warning method based on big data is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present application is essentially or the part that contributes to the relevant technology can be embodied in the form of a software product, which is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), disk or optical disk, etc. Various media that can store program codes. In this way, the embodiment of the present application is not limited to any specific hardware, software or firmware, or any combination of hardware, software and firmware.

[0199] An embodiment of the present application provides a computer system, including a memory and a processor, wherein the memory stores a computer program that can be executed on the processor, and when the processor executes the program, some or all of the steps in the above method are implemented.

[0200] The embodiment of the present application provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, some or all of the steps in the above method are implemented. The computer-readable storage medium can be transient or non-transient.

[0201] An embodiment of the present application provides a computer program, including a computer-readable code. When the computer-readable code is run in a computer device, a processor in the computer device executes some or all of the steps for implementing the above method.

[0202] The embodiment of the present application provides a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program, and when the computer program is read and executed by a computer, some or all of the steps in the above method are implemented. The computer program product can be implemented specifically by hardware, software or a combination thereof. In some embodiments, the computer program product is specifically embodied as a computer storage medium, and in other embodiments, the computer program product is specifically embodied as a software product, such as a software development kit (SDK) and the like.

[0203] It should be noted here that the description of the various embodiments above tends to emphasize the differences between the various embodiments, and the same or similar aspects can be referenced to each other. The description of the above device, storage medium, computer program and computer program product embodiments is similar to the description of the above method embodiment, and has similar beneficial effects as the method embodiment. For technical details not disclosed in the embodiments of the device, storage medium, computer program and computer program product of this application, please refer to the description of the method embodiment of this application for understanding.

[0204] If the above-mentioned integrated unit of the present application is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can essentially or in other words, the part that contributes to the relevant technology can be embodied in the form of a software product, which is stored in a storage medium and includes a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0205] The above is only an implementation method of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.

Claims

1. A charging pile line fire warning method based on big data, characterized in that: The method comprises: Collecting a line operation data set of a target charging pile cluster, wherein the line operation data set includes multi-source time series monitoring data; Performing cross-modal feature alignment on the multi-source time series monitoring data to generate a spatiotemporal correlation feature matrix, wherein the cross-modal feature alignment includes dynamically adjusting the weight allocation of the temperature change feature based on the current fluctuation amplitude, and correcting the correlation relationship between the current and temperature features based on the line aging index; Inputting the spatiotemporal correlation feature matrix into a pre-trained fire risk prediction model to generate a line abnormality risk probability distribution map, wherein the fire risk prediction model is trained by a mapping relationship between a historical fire event data set and multi-source monitoring data; Generate a hierarchical warning signal set according to the risk level corresponding to each spatial node in the line abnormality risk probability distribution diagram, wherein the hierarchical warning signal set includes differentiated control instructions for different charging pile line segments; A dynamic protection mechanism is triggered based on the hierarchical warning signal set, and the dynamic protection mechanism includes performing a current interruption operation on a high-risk line segment and performing a power reduction operation on an adjacent line segment.

2. The method according to claim 1, characterized in that The multi-source time series monitoring data includes a current fluctuation feature sequence, a temperature change feature sequence and a line aging index sequence. The cross-modal feature alignment is performed on the multi-source time series monitoring data to generate a spatiotemporal correlation feature matrix, including: A time dimension sliding window is constructed according to the amplitude change rate of the current fluctuation characteristic sequence, and a current fluctuation peak-to-valley difference and a duration time interval are extracted in the sliding window as a first dynamic characteristic vector; Extracting the temperature gradient variation and extreme value distribution density of the temperature variation feature sequence in the sliding window as a second dynamic feature vector; Decomposing the line aging index sequence into two independent components, namely, a conductive material degradation coefficient and an insulation layer damage rate, and respectively calculating cross-influence factors between the independent components and the first dynamic eigenvector and the second dynamic eigenvector; Dynamically weight the cross-influencing factors through a multi-channel attention mechanism to generate an association mapping relationship between the current fluctuation feature sequence, the temperature change feature sequence and the line aging index sequence; The first dynamic feature vector and the second dynamic feature vector are spatially interpolated and fused based on the association mapping relationship to generate the spatiotemporal association feature matrix containing a timestamp mark.

3. The method according to claim 2, characterized in that Decomposing the line aging index sequence into two independent components, namely, a conductive material degradation coefficient and an insulation layer damage rate, and respectively calculating cross-influence factors between the independent components and the first dynamic feature vector and the second dynamic feature vector, comprises: Decomposing the line aging index sequence in the frequency domain, extracting the low-frequency fluctuation component as the conductive material degradation coefficient, and separating the high-frequency mutation component as the insulation layer damage rate; Time-aligning the conductive material degradation coefficient with the current fluctuation peak-to-valley difference in the first dynamic eigenvector, calculating the phase correlation between the two within the sliding window, and generating a current-degradation interaction eigenvector; Spatially matching the insulation layer damage rate with the temperature gradient change in the second dynamic feature vector, detecting the synchronous offset of the two in the extreme point distribution, and generating a temperature-damage interaction feature vector; Performing multi-scale convolution processing on the current-degradation interaction feature vector, extracting the hysteresis response mode of the degradation coefficient of the conductive material to the change of the current fluctuation amplitude, and generating a first cross-influence factor sequence; Performing bidirectional gated loop processing on the temperature-damage interaction feature vector to capture the cumulative effect of the insulation layer damage rate on the temperature extreme value distribution and generate a second cross-influence factor sequence; The method of dynamically allocating weights of the cross-influence factors through a multi-channel attention mechanism to generate an association mapping relationship between the current fluctuation feature sequence, the temperature change feature sequence, and the line aging index sequence includes: Channel-joining the first cross-influence factor sequence with the current fluctuation feature sequence to form a current modal correlation feature block, and channel-joining the second cross-influence factor sequence with the temperature change feature sequence to form a temperature modal correlation feature block; The following processing is performed on the current mode-related feature block and the temperature mode-related feature block through a cross-modal attention gating mechanism: In the current-dominated attention branch, the weighted influence value of each time step of the current mode-related feature block on the temperature mode-related feature block is calculated to generate an attention distribution diagram from current to temperature; In the temperature-dominated attention branch, the weighted influence value of each spatial node of the temperature mode-related feature block on the current mode-related feature block is calculated to generate an attention distribution map from temperature to current; Performing a dot multiplication operation on the attention distribution map from current to temperature and the temperature mode associated feature block to obtain a temperature feature enhancement vector, and performing a dot multiplication operation on the attention distribution map from temperature to current and the current mode associated feature block to obtain a current feature enhancement vector; The temperature feature enhancement vector and the current feature enhancement vector are cross-projected and fused to generate an associated mapping relationship among the current fluctuation feature sequence, the temperature change feature sequence and the line aging index sequence.

4. The method according to claim 2, characterized in that: The performing spatial interpolation fusion on the first dynamic feature vector and the second dynamic feature vector based on the association mapping relationship to generate the spatiotemporal association feature matrix containing a timestamp mark includes: Performing timestamp alignment processing on the current fluctuation peak-to-valley difference sequence in the first dynamic feature vector and the temperature gradient change sequence in the second dynamic feature vector to generate a time-synchronized current-temperature feature pair sequence; According to the attention distribution diagram of current to temperature in the association mapping relationship, extract the interpolation weight distribution of the current fluctuation peak-to-valley difference sequence in the spatial dimension to generate a current-dominated spatial interpolation mask matrix; According to the attention distribution diagram from temperature to current in the association mapping relationship, extract the interpolation weight distribution of the temperature gradient variation sequence in the time dimension, and generate a temperature-dominated time interpolation mask matrix; Performing a spatial dimension convolution operation on the current-dominated spatial interpolation mask matrix and the time-synchronized current-temperature feature pair sequence to generate an intermediate fusion matrix with enhanced current features; Performing a sliding window superposition of the temperature-dominated time interpolation mask matrix and the current feature enhanced intermediate fusion matrix in the time dimension to generate a spatiotemporal joint feature tensor; The following interpolation operation is performed on the spatiotemporal joint feature tensor: In the spatial dimension, according to the cross-projection fusion result in the association mapping relationship, bidirectional weighted interpolation is performed on the feature vectors of adjacent line segments to generate a spatial continuity feature layer; In the time dimension, based on the timestamp interval of the original multi-source time series monitoring data, the spatial continuity feature layer is interpolated and completed in the time domain to generate a timestamp-aligned feature distribution map; Channel-superimpose the spatial continuity feature layer and the feature distribution map aligned with the timestamp to generate the spatiotemporal correlation feature matrix containing timestamp marks, wherein each timestamp corresponds to the original acquisition time of the multi-source time series monitoring data; Edge feature compensation processing is performed on the spatiotemporal correlation feature matrix, feature attenuation areas at the matrix boundaries are detected, and feature value repair is performed according to the dynamic weight allocation ratio in the correlation mapping relationship to generate a spatiotemporal correlation feature matrix that completely covers the target charging pile line topology.

5. The method according to claim 1, characterized in that The training process of the fire risk prediction model includes: Acquire a historical fire event data set, wherein the historical fire event data set includes multi-source monitoring data slices within a set time range before the fire occurs and corresponding fire location annotation information; Performing cross-modal feature alignment processing on the multi-source monitoring data slices to generate a historical spatiotemporal correlation feature matrix set; Constructing a deep spatiotemporal convolutional network, the deep spatiotemporal convolutional network comprising a local feature extraction branch and a global association branch connected in parallel, wherein the local feature extraction branch uses a three-dimensional hole convolution kernel to capture local abnormal patterns of the line, and the global association branch uses a graph attention mechanism to model the topological relationship across line segments; Input the historical spatiotemporal correlation feature matrix set into the deep spatiotemporal convolutional network for forward propagation, output a predicted fire probability distribution map, and calculate the cross entropy loss with the actual fire location annotation information; The parameters of the deep spatiotemporal convolutional network are iteratively optimized through a back-propagation algorithm until the spatial overlap between the predicted fire probability distribution map and the actual annotation reaches a preset threshold.

6. The method according to claim 5, characterized in that The global association branch of the deep spatiotemporal convolutional network performs the following processing: Constructing a line topology diagram according to the physical connection relationship of the charging pile line, wherein the node of the line topology diagram represents a single charging pile line segment, and the edge represents the electrical connection relationship between adjacent line segments; Assigning an initial feature vector to each node, wherein the initial feature vector includes the current fluctuation characteristics, temperature change characteristics and aging index statistics corresponding to the charging pile line segment; The node features are iteratively updated through a multi-layer graph attention network, where each layer of the graph attention network performs the following operations: Calculate an attention coefficient between a current node and its neighboring nodes, wherein the attention coefficient is dynamically adjusted based on a current phase difference and a temperature conduction rate between the nodes; Perform weighted aggregation on neighbor node features according to the attention coefficient, and perform residual connection with the current node features; The updated node features are input into the gated linear unit for nonlinear transformation; The node features output by the last layer of graph attention network are channel-concatenated with the output features of the local feature extraction branch to generate fused global-local joint features.

7. The method according to claim 1, characterized in that The generating of a hierarchical warning signal set according to the risk level corresponding to each spatial node in the line abnormality risk probability distribution diagram includes: Dividing the line abnormality risk probability distribution map into a plurality of risk level areas, each risk level area corresponding to a preset early warning response strategy; For the first risk level area, the real-time current spectrum characteristics of the line section within it are extracted to detect whether there is harmonic resonance phenomenon; if harmonic resonance is detected, a first-level warning signal containing a harmonic suppression instruction is generated; For the second risk level area, the deviation between the temperature change rate and the current load of the line section in the area is calculated, and when the deviation exceeds the dynamic threshold, a second level warning signal including a load balancing instruction is generated; For the third risk level area, the cumulative change of the line aging index is monitored, and when the cumulative change reaches a preset critical value, a third level warning signal including a preventive maintenance prompt is generated; The first-level warning signal, the second-level warning signal and the third-level warning signal are topologically sorted according to their spatial positions to generate the hierarchical warning signal set that matches the physical layout of the charging pile line.

8. The method according to claim 7, characterized in that The detecting whether there is a harmonic resonance phenomenon comprises: Acquire current waveform sampling data of the target line section, and extract the frequency-amplitude pair set of the fundamental component and each harmonic component through time-frequency conversion processing; Calculate the matching degree between the harmonic frequencies in the frequency-amplitude pair set and the inherent resonant frequency spectrum of the line to generate the frequency offset and resonance tendency index corresponding to each harmonic component; Screening potentially dangerous harmonic components based on the resonance tendency index, and inputting the amplitudes of the potentially dangerous harmonic components into an energy accumulation predictor in a time series to generate a harmonic energy accumulation trajectory; The trend analysis is performed on the accumulated trajectory of harmonic energy. When it is detected that the continuous rising slope exceeds the safety threshold, the multi-band interference detection process is triggered: Extract the harmonic phase difference sequence of adjacent frequency bands, calculate the phase synchronization index and energy superposition effect coefficient; Inputting the phase synchronization index and the energy superposition effect coefficient into a resonance risk classifier to generate a comprehensive resonance risk level; adjusting a response strength parameter of a harmonic suppression instruction according to the comprehensive resonance risk level, the response strength parameter including a filter startup priority and an attenuation depth gradient; The adjusted response intensity parameters are bound to the harmonic suppression instructions in the first-level warning signal to generate an optimized suppression instruction set that dynamically adapts to the electrical characteristics of the target line segment.

9. The method according to claim 1, characterized in that: The triggering of a dynamic protection mechanism based on the hierarchical warning signal set includes: Parse the line segment identifiers and warning response strategy parameters corresponding to each warning signal in the hierarchical warning signal set, extract the line segment set covered by the first-level warning signal containing the harmonic suppression instruction, and the electrical connection topology relationship corresponding to the line segment set; Sending a current cutoff instruction to a target charging pile in the line segment set, the current cutoff instruction including a cutoff phase angle range and a duration parameter generated based on the harmonic suppression instruction, and collecting in real time the current waveform distortion rate and the temperature gradient change after the cutoff operation is performed; When the current waveform distortion rate exceeds a preset distortion threshold or the temperature gradient change does not reach the expected drop, the power attenuation compensation process of the adjacent line segment is activated: Determine a propagation path sequence of power attenuation according to the electrical connection topology, and dynamically allocate a power attenuation ratio of each path node based on a load balancing instruction in the early warning response strategy parameter; The power attenuation ratio is converted into a pulse width modulation waveform, injected into the power regulation loop of the corresponding charging pile, and a step attenuation signal synchronized with the cut-off phase angle range is generated; Collecting line status feedback data under the step attenuation signal, including the distribution of current harmonic components and temperature field uniformity index after attenuation, and inputting it into the fire risk prediction model for real-time risk reassessment; The line abnormality risk probability distribution diagram is updated according to the real-time risk reassessment result. If a diffusion trend of the harmonic energy accumulation trajectory is detected, the secondary protection strategy is triggered: performing a voltage amplitude limiting operation on key nodes in the propagation path sequence to suppress the conduction path of abnormal harmonic components; After the voltage amplitude limiting operation takes effect, recalculating the power attenuation ratio, and superimposing a high-frequency oscillation suppression component to eliminate residual harmonic interference; Performing timing alignment calibration on the updated power attenuation parameter and the current cutoff instruction to generate a dynamic protection instruction sequence including a multi-stage protection logic connection relationship; The target charging pile cluster is controlled to perform a coordinated protection operation through the dynamic protection instruction sequence, and the dynamic response curve of the current waveform distortion rate and the temperature gradient change is continuously monitored.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps in the method according to any one of claims 1 to 9 are implemented.

Images