Deep learning-based electrical fire hazard current waveform recognition method and system

Abstract

This invention relates to the field of data processing, and more specifically, to a method and system for identifying electrical fire hazards based on deep learning. The method includes: constructing multiple sampling windows containing current waveform time series, temperature time series, and corresponding labels; calculating the distinguishing weights of various labels based on the sampling windows, extracting multi-dimensional features, and filtering to obtain distinguishing feature vectors, wherein the distinguishing weights are determined based on the concentration and overlap area of ​​the energy entropy distribution; constructing a sample set based on the sampling windows and distinguishing feature vectors, and training a network model using a weighted loss function; collecting real-time data to extract real-time distinguishing feature vectors, and inputting them into the model to obtain the identification result. This invention, by fusing multi-dimensional features of current and temperature, and combining label distinguishing weights and a weighted loss function, solves the problem of low identification accuracy for weak faults in existing technologies, achieving accurate identification of four types of hazards: short circuit, overload, poor contact, and leakage.

Description

Technical Field

[0001] This invention relates to the field of data processing. More specifically, this invention relates to a method and system for identifying electrical fire hazard current waveforms based on deep learning. Background Technology

[0002] Electrical fires are one of the main causes of fire accidents, characterized by their wide distribution, long duration, and strong initial concealment. Their main causes include four typical hazards: short circuits, overloads, poor contact, and leakage current. These hazards will be reflected in the electrical signal for a long time before escalating into an open flame.

[0003] In practical engineering applications, the number of labeled samples of fires that have already occurred is extremely small, making it difficult to directly train identification models using large-sample fire data. Early warning can only be achieved based on electrical hazard data that has not yet resulted in a fire. Existing technologies rely solely on a single current signal for detection, failing to fully utilize the physical characteristics of current and temperature, such as hysteresis response, trend correlation, and energy coupling. This results in low identification accuracy for weak faults such as poor contact and leakage, and is prone to misjudgment and missed detection.

[0004] Therefore, how to construct a high-precision electrical fire hazard identification method in scenarios where electrical fire samples are scarce and hazard categories are easily confused has become an urgent technical problem to be solved. Summary of the Invention

[0005] To address the technical problem of low fault identification due to reliance on a single current signal in existing technologies, this invention provides solutions in the following aspects.

[0006] In the first aspect, the deep learning-based method for identifying electrical fire hazard current waveforms includes: The current waveform timing and temperature timing of the electrical circuit to be monitored are acquired. The current waveform timing and temperature timing are divided according to a preset sampling window to obtain the current waveform timing and temperature timing of multiple sampling windows, and the status label corresponding to each sampling window is acquired. Based on the energy entropy distribution of the current waveform time sequence in each sampling window, the concentration of each type of label is calculated, and the overlap area of ​​the energy entropy distribution of any two types of labels is calculated. Then, by combining the concentration of each type of label and the average overlap area between that type of label and other labels, the distinguishing weight of each type of label is determined. Based on the distinguishing weight of each type of label, the multi-dimensional features of the electrical circuit to be monitored are extracted and the distinguishing feature vector is obtained. A weighted loss function is constructed based on the discrimination weight of the label corresponding to each sampling window. A sample set is constructed based on the label and discrimination feature vector corresponding to each sampling window. The network model is trained using the weighted loss function and the sample set. Real-time current waveform and real-time temperature time series of the electrical circuit to be monitored are collected, real-time distinguishing feature vectors are extracted, and input into the trained network model to obtain the identification results of electrical fire hazards.

[0007] Optionally, the status labels include normal status labels and potential hazard status labels, wherein potential hazard status labels include short circuit, overload, poor contact, and leakage.

[0008] Optionally, calculating the concentration of various types of labels includes: Perform a fast Fourier transform on the current waveform timing of any sampling window, divide the spectrum into multiple frequency bands, calculate the energy proportion of each frequency band, and calculate the energy entropy of the sampling window based on the Shannon entropy formula; For any type of label, obtain its corresponding window subset, obtain the energy entropy distribution of all sampling windows within the window subset, calculate the kurtosis of the energy entropy distribution of the label type, normalize the kurtosis, and use the normalization result as the concentration of the label type.

[0009] Optionally, extracting multi-dimensional features of the electrical circuit to be monitored and filtering them to obtain distinguishing feature vectors includes: For any sampling window, extract current waveform features and current-temperature fusion features to construct a candidate feature set; Calculate the difference and separation of feature values ​​for any feature in the candidate feature set between any two different subsets of label windows, and multiply the difference and separation as the sensitivity of the feature to the difference between the two labels. Based on the difference sensitivity of each feature to all pairwise unique label combinations, the overall discriminative power of each feature is calculated, and the features are filtered according to the overall discriminative power. The filtered features are then concatenated in a preset order to obtain the discriminative feature vector.

[0010] Optionally, the current waveform characteristics include RMS value, waveform factor, peak value, and trend slope; The current-temperature fusion features include peak time difference ratio, trend consistency, peak ratio, and rate of change ratio.

[0011] Optionally, the calculation of the difference includes: From any two different types of window subsets corresponding to different labels, select one window subset from each to form a window pair, and calculate the absolute value of the difference between the feature values ​​in the window pair. Iterate through all window pairs to obtain the absolute value of the difference between the feature values ​​in all window pairs, then calculate the mean of all absolute differences and perform normalization to obtain the degree of difference of the feature values ​​between the two types of label window subsets.

[0012] Optionally, the calculation of the separation degree includes: For any two different types of window subsets, construct the feature probability density function for that feature respectively; Calculate the KL divergence of these two characteristic probability density functions, and calculate the separation degree based on the KL divergence. The separation degree is equal to 1 minus the value of the exponential function with the natural constant as the base and the inverse of the KL divergence as the exponent.

[0013] Optionally, the effective value is obtained by squaring the amplitudes of each sampling point of the current waveform time sequence within the sampling window, summing them, taking the average, and then taking the square root; the waveform factor is obtained by calculating the ratio of the effective value of the current waveform time sequence to the average of the absolute values ​​of all sampling points of the current waveform time sequence; the peak value is taken as the maximum current value in the current waveform time sequence; and the trend slope is obtained by linearly fitting the curve of the current waveform time sequence using the least squares method.

[0014] Optionally, the peak time difference ratio is obtained by calculating the time difference between the peak times of the current waveform timing and the temperature timing within the sampling window, and then dividing it by the total length of the window. The trend consistency is obtained by solving the linear correlation between the current waveform time series and the temperature time series within the sampling window using the Pearson correlation coefficient formula; The peak ratio is obtained by calculating the ratio of the peak current to the peak temperature within the sampling window; The ratio of the rates of change is obtained by calculating the mean of the time-domain rate of change of the current waveform and the mean of the time-domain rate of change of the temperature waveform, and then taking the ratio of the two.

[0015] Secondly, a deep learning-based electrical fire hazard current waveform recognition system includes: a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the deep learning-based electrical fire hazard current waveform recognition method described in any one of the claims is implemented.

[0016] The present invention has the following beneficial effects: 1. This invention constructs a multi-dimensional feature system by integrating current waveforms and temperature time series, and quantifies the degree of confusion of various hidden dangers based on the concentration and overlap area of ​​energy entropy distribution, constructing label differentiation weights. This solves the problem of low identification of weak faults such as poor contact and leakage in existing technologies that rely solely on a single current signal. Through a screening mechanism of the product of difference and separation, the most discriminative features for hidden danger categories are selected from the multi-dimensional features, eliminating the dimensional disaster risk brought by redundant features. On this basis, a weighted loss function is used to train the network model, enabling the model to adaptively focus on the differentiation of easily confused categories under the condition of scarce fire samples, achieving accurate and stable identification of four types of electrical fire hazards: short circuit, overload, poor contact, and leakage.

[0017] 2. This invention uses the kurtosis of the energy entropy distribution of various labels as the concentration to quantify the internal waveform regularity, and combines the overlapping area of ​​the energy entropy distribution to quantify the degree of external confusion. The constructed distinguishing weight can effectively reflect the identification difficulty of easily confused categories, provide accurate weight basis for the weighted loss function, and further improve the model's recognition performance under imbalanced sample conditions. Attached Figure Description

[0018] Figure 1 This is a flowchart of steps S1-S4 in the electrical fire hazard current waveform identification method based on deep learning in an embodiment of the present invention.

[0019] Figure 2 This is a structural block diagram of the electrical fire hazard current waveform recognition system based on deep learning, according to an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0021] Reference Figure 1 The deep learning-based method for identifying electrical fire hazard current waveforms includes steps S1-S4, as detailed below: S1: Construct a historical window set containing the current waveform timing, temperature timing, and corresponding category labels of the electrical circuit to be monitored.

[0022] Current technologies rely solely on a single current signal, resulting in low accuracy in identifying minor faults such as poor contact and leakage. To fully explore the correlation between current and temperature, it is first necessary to construct a historical dataset containing current waveforms, temperature time series, and corresponding hazard category labels as the data foundation for subsequent feature extraction.

[0023] Specifically, the sampling frequency is set to 50Hz, and the sampling window duration is 1 second. Using the meter's built-in current sensor module and temperature sensing unit, the current waveform and temperature of the monitored electrical circuit located between the power source and the electrical equipment are simultaneously collected. Each sampling window corresponds to a current waveform time series and a temperature time series. The current and temperature data are then subjected to Min-Max normalization processing and mapped to... Interval.

[0024] For each sampling window, the current waveform timing sequence is sampled, and its label is obtained. This label is either manually labeled or automatically generated by the meter monitoring module. The label categories include normal state and potential danger state, with potential danger state including four typical types: short circuit, overload, poor contact, and leakage.

[0025] Furthermore, the above labels are converted into unique integer codes according to a preset mapping relationship. For example, label 0 corresponds to normal state, label 1 corresponds to short circuit, label 2 corresponds to overload, label 3 corresponds to leakage, and label 4 corresponds to poor contact.

[0026] The current waveform timing, temperature timing, and labels of each sampling window are strictly aligned in the time dimension, and all sampling windows are summarized to form a historical window set.

[0027] S2: Calculate the distinguishing weight of each label based on the historical window set, extract multi-dimensional features of the electrical circuit to be monitored, and filter to obtain distinguishing feature vectors.

[0028] The historical window set constructed in S1 contains the current waveform, temperature, and corresponding tags of the electrical circuit to be monitored. However, these raw signals are continuous high-dimensional time-series data, which are mixed with noise and redundant information. Moreover, a single current signal is insufficient to comprehensively characterize the electrothermal coupling characteristics of different types of potential hazards. Therefore, further feature processing is needed on the historical window set to select the most discriminative features and provide high-quality input for subsequent analysis.

[0029] First, considering the differences in the regularity of current waveforms across different categories, and the similarity in waveform characteristics between certain potential hazards such as leakage and poor contact, treating all categories equally in subsequent classifications would limit the accuracy of identifying easily confused categories. Therefore, it is necessary to quantify the distinguishability of each category, giving greater attention to easily confused categories in subsequent analysis.

[0030] The specific process for obtaining the distinguishability of the above categories is as follows: Any sampling window in the historical window set is divided into 5 window subsets according to its corresponding window label, for example: normal window subset - label 0, short circuit window subset - label 1, overload window subset - label 2, leakage window subset - label 3, and poor contact window subset - label 4.

[0031] A Fast Fourier Transform (FFT) is performed on the current waveform timing of any sampling window within the historical window set, dividing the spectrum into six frequency bands. The energy proportion of each frequency band is calculated, and the energy entropy of that sampling window is obtained based on the Shannon entropy formula. It should be noted that both the energy proportion calculation and the energy entropy calculation based on the Shannon entropy formula are existing techniques and will not be elaborated upon here.

[0032] Furthermore, for any type of label, obtain its corresponding window subset, obtain the energy entropy distribution of all sampling windows within that window subset, calculate the kurtosis of the energy entropy distribution of that type of label, and perform Min-Max normalization on the kurtosis to map the value range to... The range is then used, and the result of kurtosis normalization is taken as the concentration of this type of label. By calculating the concentration, we can reflect the degree of concentration of the waveform regularity within this type of label. The higher the concentration, the stronger the waveform regularity and the lower the internal diversity of this type of label. Conversely, the lower the concentration, the weaker the waveform regularity and the higher the internal diversity of this type of label.

[0033] Furthermore, for any two different types of labels, the probability density function of the energy entropy distribution of the corresponding window subsets of these two types of labels is obtained by kernel density estimation. Then, the overlap area of ​​the two types of label distributions is calculated by integrating these two probability density functions. The larger the overlap area, the more similar the overall waveform regularity of the two types of labels is, and the easier it is to confuse them.

[0034] Furthermore, for each type of label, the overlap area between it and other labels is calculated, and then the average of all overlapping areas is taken as the average overlap area of ​​that type of label.

[0035] Finally, combining the concentration of the tags themselves with the average overlap area calculated above, the distinguishing weight of the tags is calculated, which satisfies the following relationship: In the formula, For the first Distinguishing weights of class tags For the first Concentration of class tags For the first Average overlap area of ​​class tags This represents the number of label types.

[0036] Among them, adopt Converting concentration into an indicator reflecting the degree of internal confusion makes and They have the same directionality - the larger the value, the easier it is to confuse them. and The effects of the two on distinguishability are independent of each other, and the addition method can integrate information from both internal and external confusion dimensions.

[0037] Through the above calculations, easily confused labels such as leakage and poor contact are given greater distinguishing weight.

[0038] Furthermore, considering that relying solely on a single current waveform characteristic is insufficient to fully characterize the electrothermal coupling properties of electrical fire hazards—for example, poor contact may lead to localized temperature rises without significant current changes, while overload may cause simultaneous increases in both current and temperature—it is necessary to construct a multi-dimensional characteristic system for the monitored electrical circuit that integrates current and temperature.

[0039] The specific process for constructing the multi-dimensional features of the aforementioned electrical circuit to be monitored is as follows: For any sampling window in the historical window set, extract the following features: (1) Current waveform characteristics: RMS value: The RMS value is obtained by squaring the amplitude of each sampling point of the current waveform timing within the sampling window, summing the squared results of all sampling points, taking the average, and then taking the square root. This RMS value reflects the energy magnitude and thermal effect characteristics of the current waveform timing.

[0040] Waveform factor: Calculated by the mean of the absolute values ​​of the effective value and the values ​​of all sampling points of the current waveform timing sequence. This waveform factor indicates the fullness of the current waveform.

[0041] Peak value: The maximum current value in the current waveform timing sequence, which directly reflects the maximum transient impact energy of the current waveform timing sequence.

[0042] Trend slope: The least squares method is used to linearly fit the current waveform time sequence curve to obtain the trend slope of the fitted line, which reflects the direction and magnitude of the current change with time.

[0043] (2) Current-temperature fusion characteristics: Peak time difference ratio: The generation, transfer and accumulation of heat have a certain lag. The time lag between the peak current and the peak temperature may vary under different labels. Therefore, the peak times of the current waveform and temperature waveform within the sampling window are located, the time difference between the two is calculated, and then divided by the total length of the window to obtain the normalized ratio result. This result is used as the peak time difference ratio.

[0044] Trend consistency: Pairwise correlation analysis is performed on all sampling points of the current waveform time series and temperature time series within the sampling window. The linear correlation between the two is solved by the Pearson correlation coefficient formula to obtain the trend consistency. The closer the absolute value of the correlation coefficient is to 1, the stronger the synchronicity of the current and temperature change trends.

[0045] Peak ratio: Calculate the peak value of the current waveform and the peak value of the temperature waveform within the sampling window, and then calculate the ratio of the current peak value to the temperature peak value to obtain the peak ratio, which reflects the temperature response intensity corresponding to a unit current impact.

[0046] Rate of Change Ratio: The mean time-domain rate of change for the current waveform (i.e., the mean of the absolute values ​​of the differences between adjacent sampling points in the current waveform time series) and the mean time-domain rate of change for the temperature time series are calculated separately. Then, the ratio of the mean current rate of change to the mean temperature rate of change is obtained to quantify the response relationship between current and temperature changes. The mean time-domain rate of change for the temperature time series is calculated in the same way as that for the current waveform time series.

[0047] For any sampling window in the historical window set, the above feature extraction operation is performed to obtain the current waveform features and current-temperature features of each sampling window. Min-Max normalization is then performed on each feature to map the feature values ​​to... The interval is then used to construct a candidate feature set from the aforementioned multidimensional features.

[0048] In summary, through the above multi-dimensional feature extraction, a complete feature system covering current waveform morphology, energy characteristics, and electrothermal coupling relationship has been constructed.

[0049] Furthermore, the candidate feature set constructed above is screened because there may be a large number of redundant features in the candidate feature set that do not contribute much to distinguishing different hidden dangers. If they are used directly for subsequent identification, it will not only increase the computational burden, but may also lead to a decline in identification performance due to the curse of dimensionality.

[0050] For any feature in the candidate feature set, select one window subset from each of the window subsets corresponding to any two different labels to form a window pair. Calculate the absolute value of the difference between the feature values ​​in the window pair. Iterate through all window pairs to obtain the absolute value of the difference between the feature values ​​in all window pairs. Then calculate the mean of all absolute values ​​of the difference and perform Min-Max normalization to obtain the degree of difference of the feature values ​​between any two different label window subsets.

[0051] Furthermore, for any feature in the candidate feature set, feature probability density functions are constructed for the window subsets corresponding to any two different labels. This is achieved by using kernel density estimation, fitting the probability density of the feature values ​​of the window subsets for each of the two label classes using a Gaussian kernel function, and obtaining their respective feature probability density functions. Then, the KL (Kullback-Leibler) divergence of these two feature probability density functions is calculated. Finally, based on the calculated KL divergence, the separation degree of the feature values ​​between any two different label window subsets is calculated, satisfying the following relationship: In the formula, Features in the above candidate feature set In the label With tags The separation of feature values ​​between window subsets Features In the label With tags The KL divergence between the feature probability density functions corresponding to the window subsets.

[0052] Among them, through The formal transformation maps the unbounded KL divergence to a bounded one. Interval.

[0053] Finally, multiplying the difference degree by the separation degree yields the feature. For tags and tags The difference sensitivity is specifically measured by the difference measure, which only reflects the average numerical difference of a feature across two categories of labels. When the feature values ​​of two categories of labels are similar, the difference measure may be small, but the distribution patterns of their feature values ​​may differ significantly. For example, a certain feature of two potential hazards, leakage current and poor contact, may be similar, but one category's distribution may be concentrated while the other's is dispersed. Therefore, combining the separation measure with the KL divergence measure to quantify the overall difference between the two distributions can effectively capture this difference in distribution patterns and compensate for the inadequacy of the difference measure. Through multiplication, it is ensured that the difference sensitivity is only high when both indicators are simultaneously large; if either indicator is small, the product result will be significantly suppressed, thus rigorously selecting features with true discriminative power.

[0054] Furthermore, for any feature in the candidate feature set, the difference sensitivity of that feature to all pairwise non-repeating label combinations is calculated sequentially. The difference sensitivity of that feature to all non-repeating label combinations is then summed to obtain the overall discriminative power of that feature.

[0055] Then, the overall discriminative power of each feature in the candidate feature set is calculated, and all features are sorted in descending order according to the overall discriminative power value. The screening threshold is set by percentage method. For example, the overall discriminative power corresponding to the top 10% of features after sorting is selected as the screening threshold. Features with an overall discriminative power greater than the screening threshold are retained. The discriminative feature vector is obtained by concatenating the corresponding overall discriminative power in descending order.

[0056] Through the above feature screening, redundant and invalid features were eliminated, and only the features that contribute most to distinguishing various hidden dangers were retained, effectively reducing the input dimensions of subsequent analysis.

[0057] S3: Construct a sample set based on the historical window set and the distinguishing feature vector, and train a weighted BP neural network model.

[0058] The discriminative feature vectors obtained in step S2 have transformed the original signal into a feature representation with strong discriminative power, but the features themselves do not possess classification decision-making capabilities. To achieve automatic identification of the types of hidden dangers in the monitored electrical circuits, a classification model needs to be constructed to learn from the features. Considering the varying difficulty in identifying different types of hidden dangers, more attention needs to be paid to easily confused categories during model training.

[0059] For any sampling window in the historical window set, retrieve the distinguishing feature vector of that sampling window after the S2 feature filtering process described above. One sampling window corresponds to one sample, and each sample corresponds to one distinguishing feature vector and one label encoding vector. This label encoding vector is a 5-dimensional one-hot encoding, corresponding to five categories of labels: normal, short circuit, overload, leakage, and poor contact. Only the position corresponding to the true label is 1, and the other positions are 0. Summarize all samples to obtain the sample set used for training the BP neural network.

[0060] Furthermore, combining the discriminative weights of the labels calculated in S2 above, the discriminative weight of the label to which each sample belongs is used as the weight corresponding to that sample, and a weighted loss function is constructed based on cross-entropy loss: In the formula, For the sample set Cross-entropy loss for each sample, For the first The distinguishing weight of the label corresponding to each sample. For the number of label types, This is the weighted loss of the model.

[0061] The cross-entropy loss for each sample is a measure of the model's prediction bias for that sample, calculated by multiplying by... This can amplify or reduce the contribution of that sample to the total loss. When When the error is large, the model's prediction bias for that sample will be amplified, thus giving that sample a higher optimization priority in gradient updates.

[0062] The sample set constructed above was stratified and divided into a training set and a validation set at a ratio of 8:2, ensuring that the samples of each type of label were evenly distributed in the training set and the validation set during the division process.

[0063] Construct a 3-layer BP (Back Propagation) neural network: the number of neurons in the input layer is consistent with the dimension L of the distinguishing feature vector; the hidden layer has 1 layer with 64 neurons, using the ReLU activation function, and a dropout coefficient of 0.2 is introduced; the output layer has 5 neurons and uses the Softmax activation function.

[0064] Training samples are input into the BP neural network in batches, and the model's predicted output is calculated through forward propagation. A weighted loss function is used as the loss function, and the Adam optimizer is employed to calculate the gradient of the loss function with respect to each model parameter. The parameters are then updated using a gradient descent strategy. After each iteration of the full training set, the model performance is evaluated using the validation set, and the validation set loss is calculated. If the validation set loss value does not decrease for five consecutive iterations or the number of iterations reaches a preset maximum (e.g., 50 iterations), the model is considered to have converged, and training is terminated. After training terminates, all parameters of the converged BP neural network are saved.

[0065] By introducing a weighted loss function, the model can adaptively adjust the degree of attention to different categories during training, effectively improving the identification accuracy of easily confused hidden dangers such as leakage and poor contact, while maintaining the identification performance of easily distinguishable categories such as normal and short circuit.

[0066] S4: Collect real-time data of the electrical circuits to be monitored, extract real-time distinguishing feature vectors, and input them into the trained BP model to complete the real-time detection of electrical fire hazards.

[0067] According to the sampling parameters set in S1 above, the current waveform and temperature time series of the electrical circuit to be monitored are collected in real time. The same data preprocessing method as in S1 is used to generate a real-time sampling window. The feature extraction operation in S2 above is performed on the real-time sampling window to extract the features in the distinguishing feature vector and obtain the real-time distinguishing feature vector.

[0068] The real-time distinguishing feature vector is input into the trained BP model. The model calculates and outputs the probability distribution corresponding to five labels through forward propagation. The label corresponding to the maximum probability is taken as the recognition result of the current waveform, where label 0 corresponds to the normal state, and labels 1-4 correspond to four abnormal states: short circuit, overload, leakage, and poor contact, respectively. If the recognition result is label 0, it is determined that there is no electrical fire hazard in the current electrical circuit; if the recognition result is label 1-4, it is determined that there is a corresponding type of electrical fire hazard in the current electrical circuit, thus completing the real-time detection.

[0069] In summary, this invention constructs a historical dataset containing the current, temperature, and corresponding labels of the electrical circuits to be monitored through step S1; calculates label discrimination weights based on the historical window set, extracts multi-dimensional features of the electrical circuits to be monitored, and filters effective features through step S2, constructing a highly discriminative feature vector; constructs a sample set based on the historical window set through step S3, builds a weighted loss function, and trains a BP neural network, enabling the model to focus on distinguishing easily confused hazards; and collects real-time data of the electrical circuits to be monitored through step S4, achieving real-time detection and early warning. The overall method solves the problems of scarce electrical fire samples and easily confused hazard categories in existing technologies, achieving accurate and stable identification of four types of electrical fire hazards: short circuit, overload, poor contact, and leakage.

[0070] This invention also provides a deep learning-based system for identifying electrical fire hazard current waveforms. For example... Figure 2 As shown, the system includes a processor and a memory. The memory stores computer program instructions. When the computer program instructions are executed by the processor, the method for identifying electrical fire hazard current waveforms based on deep learning according to the first aspect of the present invention is implemented.

[0071] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0072] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for identifying electrical fire hazard current waveforms based on deep learning, characterized in that, include: The current waveform timing and temperature timing of the electrical circuit to be monitored are acquired. The current waveform timing and temperature timing are divided according to a preset sampling window to obtain the current waveform timing and temperature timing of multiple sampling windows, and the status label corresponding to each sampling window is acquired. Based on the energy entropy distribution of the current waveform time sequence in each sampling window, the concentration of each type of label is calculated, and the overlap area of ​​the energy entropy distribution of any two types of labels is calculated. Then, by combining the concentration of each type of label and the average overlap area between that type of label and other labels, the distinguishing weight of each type of label is determined. Based on the distinguishing weight of each type of label, the multi-dimensional features of the electrical circuit to be monitored are extracted and the distinguishing feature vector is obtained. A weighted loss function is constructed based on the discrimination weight of the label corresponding to each sampling window. A sample set is constructed based on the label and discrimination feature vector corresponding to each sampling window. The network model is trained using the weighted loss function and the sample set. Real-time current waveform and real-time temperature time series of the electrical circuit to be monitored are collected, real-time distinguishing feature vectors are extracted, and input into the trained network model to obtain the identification results of electrical fire hazards.

2. The method for identifying electrical fire hazard current waveforms based on deep learning according to claim 1, characterized in that, Status labels include normal status labels and potential hazard status labels, with potential hazard status labels including short circuit, overload, poor contact, and leakage.

3. The method for identifying electrical fire hazard current waveforms based on deep learning according to claim 1, characterized in that, Calculating the concentration of various labels includes: Perform a fast Fourier transform on the current waveform timing of any sampling window, divide the spectrum into multiple frequency bands, calculate the energy proportion of each frequency band, and calculate the energy entropy of the sampling window based on the Shannon entropy formula; For any type of label, obtain its corresponding window subset, obtain the energy entropy distribution of all sampling windows within the window subset, calculate the kurtosis of the energy entropy distribution of the label type, normalize the kurtosis, and use the normalization result as the concentration of the label type.

4. The method for identifying electrical fire hazard current waveforms based on deep learning according to claim 3, characterized in that, Extracting multi-dimensional features of the electrical circuits to be monitored and filtering them to obtain distinguishing feature vectors includes: For any sampling window, extract current waveform features and current-temperature fusion features to construct a candidate feature set; Calculate the difference and separation of feature values ​​for any feature in the candidate feature set between any two different subsets of label windows, and multiply the difference and separation as the sensitivity of the feature to the difference between the two labels. Based on the difference sensitivity of each feature to all pairwise unique label combinations, the overall discriminative power of each feature is calculated, and the features are filtered according to the overall discriminative power. The filtered features are then concatenated in a preset order to obtain the discriminative feature vector.

5. The method for identifying electrical fire hazard current waveforms based on deep learning according to claim 4, characterized in that, The current waveform characteristics include RMS value, waveform factor, peak value, and trend slope; The current-temperature fusion features include peak time difference ratio, trend consistency, peak ratio, and rate of change ratio.

6. The method for identifying electrical fire hazard current waveforms based on deep learning according to claim 4, characterized in that, The calculation of the difference includes: From any two different types of window subsets corresponding to different labels, select one window subset from each to form a window pair, and calculate the absolute value of the difference between the feature values ​​in the window pair. Iterate through all window pairs to obtain the absolute value of the difference between the feature values ​​in all window pairs, then calculate the mean of all absolute differences and perform normalization to obtain the degree of difference of the feature values ​​between the two types of label window subsets.

7. The method for identifying electrical fire hazard current waveforms based on deep learning according to claim 4, characterized in that, The calculation of the separation degree includes: For any two different types of window subsets, construct the feature probability density function for that feature respectively; Calculate the KL divergence of these two characteristic probability density functions, and calculate the separation degree based on the KL divergence. The separation degree is equal to 1 minus the value of the exponential function with the natural constant as the base and the inverse of the KL divergence as the exponent.

8. The method for identifying electrical fire hazard current waveforms based on deep learning according to claim 5, characterized in that, The effective value is obtained by squaring the amplitude of each sampling point of the current waveform time sequence within the sampling window, summing the values, taking the average, and then taking the square root. The waveform factor is obtained by calculating the ratio of the effective value of the current waveform time sequence to the average of the absolute values ​​of all sampling points of the current waveform time sequence. The peak value is taken as the maximum current value in the current waveform time sequence. The trend slope is obtained by linearly fitting the curve of the current waveform time sequence using the least squares method.

9. The method for identifying electrical fire hazard current waveforms based on deep learning according to claim 5, characterized in that, The percentage of peak time difference is obtained by calculating the time difference between the peak times of the current waveform timing and temperature timing within the sampling window, and then dividing it by the total length of the window. The trend consistency is obtained by solving the linear correlation between the current waveform time series and the temperature time series within the sampling window using the Pearson correlation coefficient formula; The peak ratio is obtained by calculating the ratio of the peak current to the peak temperature within the sampling window; The ratio of the rates of change is obtained by calculating the mean of the time-domain rate of change of the current waveform and the mean of the time-domain rate of change of the temperature waveform, and then taking the ratio of the two.

10. A deep learning-based electrical fire hazard current waveform identification system, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the deep learning-based electrical fire hazard current waveform identification method according to any one of claims 1-9.

Images