Non-intrusive load monitoring method and system based on dual-channel convolutional neural network

Abstract

This invention proposes a non-intrusive load monitoring method and system based on a dual-channel convolutional neural network, belonging to the field of load monitoring technology. The method includes: extracting comprehensive electrical features from voltage and current measurement data of the electrical equipment to be identified, processing them into one-dimensional sequence features; performing multi-scale morphological gradient detection based on current signals to construct a nonlinear time-frequency sampling grid; extracting time-frequency features from the sampling grid to obtain a two-dimensional time-frequency coefficient matrix; constructing a dual-channel time-frequency image based on harmonic membership degrees and the two-dimensional time-frequency coefficient matrix; using a two-stage dual-channel convolutional neural network to process the one-dimensional sequence features and the dual-channel time-frequency image respectively, extracting features and performing bidirectional enhancement fusion to obtain fused features; inputting the fused features into a classification network to obtain the load identification result of the electrical equipment to be identified. This invention can adaptively allocate time-frequency resolution, suppress interference harmonics, and effectively capture local and global features, improving load identification accuracy.

Description

Technical Field

[0001] This invention relates to the field of load monitoring technology, and in particular to a non-intrusive load monitoring method and system based on a dual-channel convolutional neural network. Background Technology

[0002] With the development of smart grids, non-intrusive load monitoring (NILM) is an important tool for electricity consumption analysis and demand-side management. Existing NILM technologies have significant drawbacks: traditional signal processing methods have weak generalization capabilities; deep learning solutions have poor practicality; and two-dimensional feature methods have high hardware requirements. Furthermore, at the feature processing level, single features are insufficient for identifying devices with similar power outputs, fixed-weighted multi-feature fusion lacks adaptability, and it's difficult to balance recognition accuracy and computational efficiency. This results in poor differentiation of similar loads and makes it difficult to meet the high-precision, high-robust monitoring needs of real-world scenarios.

[0003] Furthermore, existing time-frequency analysis methods suffer from the following specific technical challenges. Current time-frequency images employ a uniform wavelet scale sampling density for all current segments, resulting in steep transient regions and smooth steady-state regions exhibiting similar edge sharpness in the grayscale image. In reality, the sudden rise of the current during appliance startup requires extremely high temporal resolution, while steady-state harmonics require extremely high frequency resolution. These two requirements for time-frequency image resolution are contradictory, and fixed sampling leads to information loss. Simultaneously, the frequency domain characteristics of harmonics and noise in actual current signals often overlap. Conventional modulus grayscale methods cannot distinguish between high-order harmonics generated by the device itself and interference harmonics from grid frequency drift or coupling with nearby devices, leading to the inclusion of false features in the grayscale image and affecting recognition accuracy. Additionally, existing convolutional neural network architectures struggle to fully capture the coupling relationship between long-distance harmonic evolution patterns and local texture abrupt changes during appliance transient processes, resulting in insufficient mining of load transient features and reduced detection accuracy for non-invasive load identification. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a non-intrusive load monitoring method and system based on a dual-channel convolutional neural network. By setting up a dual-channel architecture, it can simultaneously process one-dimensional electrical features and dual-channel time-frequency images. Furthermore, by employing a dynamic weighted fusion mechanism, it can automatically adjust the feature contribution based on the characteristics of different electrical equipment, thereby improving the accuracy of load identification.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a non-intrusive load monitoring method based on a dual-channel convolutional neural network, comprising: Extract comprehensive electrical features from the voltage and current measurement data of the electrical equipment to be identified, and process them into one-dimensional sequence features; Multi-scale morphological gradient detection is performed based on the current signal of the electrical equipment to be identified, the true transient edge position and its steepness are identified, and a nonlinear time-frequency sampling grid is constructed. Time-frequency features are extracted using continuous wavelet transform based on a nonlinear time-frequency sampling grid, resulting in a two-dimensional time-frequency coefficient matrix; Harmonic membership is calculated based on phase information, and a dual-channel time-frequency image containing amplitude and phase consistency channels is constructed by combining the two-dimensional time-frequency coefficient matrix. A two-stage convolutional neural network is used to process one-dimensional sequence features and dual-channel time-frequency images respectively, and bidirectional enhancement fusion is performed to obtain fused features; The fused features are input into the classification network to obtain the load identification results of the electrical equipment to be identified.

[0006] Secondly, the present invention provides a non-intrusive load monitoring system based on a dual-channel convolutional neural network, comprising: The one-dimensional sequence feature acquisition module is configured to extract comprehensive electrical features from the voltage and current measurement data of the electrical equipment to be identified and process them into one-dimensional sequence features. The sampling network construction module is configured to perform multi-scale morphological gradient detection based on the current signal of the electrical equipment to be identified, identify the true transient edge position and its steepness, and construct a nonlinear time-frequency sampling grid. The wavelet transform extraction module is configured to extract time-frequency features based on a nonlinear time-frequency sampling grid using continuous wavelet transform to obtain a two-dimensional time-frequency coefficient matrix. The time-frequency image construction module is configured to calculate harmonic membership based on phase information and construct a dual-channel time-frequency image containing amplitude and phase consistency channels by combining the two-dimensional time-frequency coefficient matrix. The fusion module is configured to use a two-stage convolutional neural network to process one-dimensional sequence features and dual-channel time-frequency images respectively, and perform bidirectional enhancement fusion to obtain fused features; The identification module is configured to input the fused features into the classification network to obtain the load identification result of the electrical equipment to be identified.

[0007] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the non-intrusive load monitoring method based on a dual-channel convolutional neural network described in the first aspect.

[0008] Fourthly, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the non-intrusive load monitoring method based on a dual-channel convolutional neural network described in the first aspect.

[0009] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention first extracts comprehensive electrical features from voltage and current measurement data to form a one-dimensional sequence feature. Based on multi-scale morphological gradient detection, it locates the true transient edges and their steepness, and constructs a nonlinear time-frequency sampling grid. This grid is optimized for density in transient regions and sparse in steady-state regions, with fine sampling in low-frequency bands and moderate sparse sampling in high-frequency bands, achieving differentiated allocation of time-frequency resources. A two-dimensional time-frequency coefficient matrix is ​​obtained from this sampling grid using continuous wavelet transform. Phase information is further introduced to calculate harmonic membership, constructing a dual-channel time-frequency image with amplitude and phase consistency, taking into account both amplitude variation and harmonic phase correlation characteristics. A two-stage convolutional neural network is used to perform targeted feature mining and bidirectional enhancement fusion on the one-dimensional sequence feature and the dual-channel time-frequency image, fully integrating one-dimensional electrical statistical information and time-frequency local-global complementary features. This data is then fed into a classification network to output the load identification result. This invention not only fully captures the details of transient changes in equipment switching and harmonic phase features but also suppresses the computational overhead caused by redundant steady-state data. This effectively improves the ability to distinguish different electrical equipment under complex operating conditions, enhances the robustness of the load identification model to noise and operating condition fluctuations, and improves load identification accuracy.

[0010] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0011] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute a limitation thereof.

[0012] Figure 1 This is a flowchart illustrating the main steps of a non-intrusive load monitoring method based on a dual-channel convolutional neural network, as provided in an embodiment of the present invention. Figure 2 This is a flowchart illustrating an improved time-frequency image construction process provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of a dual-channel CNN network structure provided in an embodiment of the present invention, which includes a time-frequency self-attention enhancement module and a dual-stream cross-attention fusion layer. Detailed Implementation

[0013] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0014] Example 1 like Figure 1 As shown, this embodiment discloses a non-intrusive load monitoring method based on a dual-channel convolutional neural network, including the following steps: S1: Extract comprehensive electrical features from the voltage and current measurement data of the electrical equipment to be identified and process them into one-dimensional sequence features; S2: Perform multi-scale morphological gradient detection based on the current signal of the electrical equipment to be identified, identify the true transient edge position and its steepness, and construct a nonlinear time-frequency sampling grid; S3: Based on the nonlinear time-frequency sampling grid, continuous wavelet transform is used to extract time-frequency features, resulting in a two-dimensional time-frequency coefficient matrix; S4: Calculate the harmonic membership degree based on the phase information, and construct a dual-channel time-frequency image containing the amplitude channel and the phase consistency channel by combining the two-dimensional time-frequency coefficient matrix; S5: A two-stage convolutional neural network is used to process one-dimensional sequence features and dual-channel time-frequency images respectively, and bidirectional enhancement fusion is performed to obtain fused features; S6: Input the fused features into the classification network to obtain the load identification results of the electrical equipment to be identified.

[0015] Next, combined Figure 2 This embodiment provides a detailed description of a non-intrusive load monitoring method based on a dual-channel convolutional neural network.

[0016] (a) Extracting the comprehensive electrical characteristics of the electrical equipment to be identified For each electrical device instance, a comprehensive electrical feature set X is extracted from voltage and current measurements, including power-related metrics, waveform statistics, and frequency domain characteristics. These features capture the time and frequency domain characteristics of the load, providing a multifaceted representation of its electrical behavior.

[0017] Among them, power-related indicators include active power, apparent power, reactive power, and power factor. These indicators characterize the basic energy consumption patterns and phase relationships of different electrical devices.

[0018] Waveform statistical features include RMS value, peak value, crest factor, zero-crossing rate and slope index, as well as distribution measurements such as mean, standard deviation, variance, skewness, kurtosis, quartiles and entropy, which together capture the time-domain pattern and amplitude characteristics of the current waveform.

[0019] To quantify the characteristics of nonlinear loads, frequency domain characteristics are analyzed, including fundamental amplitude and total harmonic distortion (THD). The fundamental amplitude can be directly obtained through spectral decomposition, and the formula for calculating THD is as follows: (1) in, Indicates the first n The amplitude of the second harmonic component, It is the fundamental frequency component.

[0020] (ii) Principal Component Analysis Dimensionality Reduction To address the computational complexity of the 29-dimensional feature set, Principal Component Analysis (PCA) is employed. First, the feature matrix X is standardized: (2) in, and These are the mean and standard deviation vectors of the original features, respectively.

[0021] Then, the covariance matrix C of the standardized characteristic matrix is ​​calculated according to the equation: (3) Where T represents the transpose operation.

[0022] The optimal number of principal components is determined by analyzing the cumulative explained variance ratio through eigenvalue decomposition. (4) Where d is the original feature dimension. Let represent the i-th eigenvalue after eigenvalue decomposition, sorted in descending order, and the variance information carried by the i-th principal component. Let p represent the number of principal components to be selected. Selecting the top 9 principal components that explain at least 95% of the total variance reduces the dimension from 29 to 9 features, resulting in a one-dimensional sequence feature.

[0023] (III) Adaptive Dual-Channel Time-Frequency Image Construction Furthermore, to address the issues of uneven resolution distribution and interference harmonic aliasing in existing time-frequency analysis, an improved image construction method is proposed, such as... Figure 2 As shown.

[0024] 1. Multi-scale morphological gradient detection and nonlinear time-frequency sampling mesh construction Considering that the rising edge of the current signal during appliance startup requires extremely high time resolution, while the harmonics of the current signal in steady state require extremely high frequency resolution, adopting a uniform sampling scale, i.e., using a fixed sampling rate to collect the current signal throughout the entire process, presents a dilemma: if a high sampling rate is used throughout, although it can accurately capture the steep current edge at startup and has sufficient time resolution, the frequency resolution is poor. To finely distinguish harmonics, extremely long data segments need to be collected, increasing the data processing burden. If short data is truncated for FFT processing, problems such as blurred harmonic spectrum and severe picket fence effect will occur, resulting in the loss of steady-state harmonic information. If a low sampling rate is used throughout, good frequency resolution can be obtained with the same number of sampling points, and steady-state harmonic calculation is accurate, but the time resolution is insufficient. The steep current rising edge during appliance startup will be smoothed out, transient spikes will be lost, and the real startup impact process cannot be reproduced. It is impossible to simultaneously take into account the time capture capability of transient sudden current signal and the frequency resolution capability of steady-state harmonic current signal, resulting in information loss.

[0025] Therefore, this embodiment no longer uses uniform scale sampling. First, the continuous current signal is divided into multiple current signal segments according to a preset window length and segment step size, and each current signal segment serves as the input for multi-scale morphological gradient detection.

[0026] (5) in, i k [ m ] indicates the first k The first current signal segment m One sampling point, N Indicates the window length. L Indicates the step size of the segment.

[0027] For each current signal segment, construct multiple one-dimensional structuring elements of different scales. s The structural element of each scale is represented as: (6) in, B s Indicates the first s Structural elements at various scales r s This represents the half-width of the structuring element. M This represents the total number of scales of the structuring elements. Smaller-scale structuring elements are used to detect short-term abrupt changes, while larger-scale structuring elements are used to suppress short-term glitches and noise and detect waveform abrupt changes with longer durations.

[0028] At each scale, morphological dilation and erosion operations are performed on the current signal segment, and the difference between the dilation and erosion results is used as the morphological gradient: (7) in, G s [ m ] indicates the first s Location at each scale m The morphological gradient is defined as follows: max represents the maximum value within the coverage area of ​​the structuring element, min represents the minimum value within the coverage area of ​​the structuring element, and q represents the position index within the structuring element. A larger morphological gradient indicates a more pronounced abrupt change in the current waveform near the corresponding position.

[0029] Morphological gradients at various scales are compared. When the same location exhibits significant gradient responses at multiple scales, that location is determined as the true transient edge location. The normalized gradient response at the true transient edge location is then determined as the transient edge steepness. This yields the true transient edge location and its steepness, which serve as the basis for subsequent construction of the nonlinear time-frequency sampling grid.

[0030] Furthermore, based on the detection results, a nonlinear time-frequency sampling grid is designed: Logarithmically dense sampling is used on the time axis near the transient edge to achieve higher time resolution; linear sparse sampling is used in the steady-state region far from the actual transient edge. On the frequency axis, a scaling sampling density factor is used to control the scaling sampling interval in different frequency regions. A larger scaling sampling density factor is used in the low-frequency region to reduce the scaling interval and increase the number of sampling points, so as to finely characterize the fundamental wave and its adjacent harmonics; a smaller scaling sampling density factor is used in the high-frequency region to reduce high-frequency redundant calculations.

[0031] In continuous wavelet transform, the relationship between scale and pseudofrequency is as follows: (8) in, f Indicates pseudo-frequency, f c Indicates the wavelet center frequency. a Indicates the wavelet scale. This indicates the sampling interval of the current signal.

[0032] make Represents frequency f The corresponding scale sampling density factor, If the reference scale interval is used, then the actual scale interval is: (9) (10) (11) in, The scale sampling density factor represents the low-frequency region. The scale sampling density factor represents the high-frequency region. This represents the boundary frequency between the low-frequency fine-sampling region and the high-frequency sparse-sampling region. A nonlinear time-frequency sampling grid is obtained by implementing non-uniform sampling on both the time and frequency axes.

[0033] The time-frequency coefficient matrix generated by the non-uniform sampling grid has a variable size. Finally, a differentiable elastic resampling layer maps the time-frequency coefficient matrix to a uniform 64×64 pixel image matrix. This design enables on-demand resolution allocation, allowing for more refined representation of key feature regions.

[0034] In this embodiment, a nonlinear time-frequency sampling grid is constructed through multi-scale morphological gradient detection, achieving on-demand resolution allocation. Without increasing image resolution, a limited number of pixels are automatically focused on the most information-rich time-frequency region, effectively resolving the conflict between resolution requirements in transient steep regions and steady-state smooth regions. By simultaneously implementing non-uniform sampling in both the time and frequency domains to obtain the nonlinear time-frequency sampling grid, the overall sampling data volume can be reduced while ensuring the precise capture of key information such as transient abrupt changes and low-frequency harmonics, balancing the resolution capability of key time-frequency regions with computational overhead.

[0035] 2. An Improved Morlet Wavelet Transform Based on Asymmetric Time Envelope Based on the aforementioned grid, an improved Morlet continuous wavelet transform based on asymmetric time envelope is used for time-frequency feature extraction. Its mathematical expression is: (12) in, W ( a,b ) represents wavelet coefficients. f ( t () represents the time-domain signal of the original current segment. a This is the scale factor (corresponding to the frequency component). b This is the translation factor (corresponding to the time node). It is the complex conjugate of the wavelet basis functions.

[0036] To address the transient asymmetric physical characteristics such as sudden rises and gradual falls exhibited during the start-up and shutdown phases of various electrical appliances in non-intrusive load monitoring, this embodiment abandons the traditional fully symmetric standard Morlet wavelet and constructs an improved Morlet wavelet basis function based on asymmetric time envelopes. Its time-domain mathematical expression is as follows: (13) in, f c For the center frequency, f b This is the bandwidth parameter.

[0037] The asymmetric time decay factor introduced in this embodiment is defined piecewise as follows: (14) The and The decay rate of the wavelet envelope before and after the zero point of time is controlled separately. To match the asymmetric decay characteristics of the current waveform during load startup, the following settings are configured: This improved structure in tThe <0 stage exhibits a steep decay, providing extremely high temporal resolution to precisely pinpoint the initial transient edge of abrupt changes in electrical state; in t The ≥0 stage exhibits a relatively gentle attenuation, matching the asymmetric attenuation characteristics of the load starting current. The complex wavelet coefficient matrix is ​​obtained. W ( a,b ), that is, a two-dimensional time-frequency coefficient matrix.

[0038] 3. Harmonic membership calculation and dual-channel image generation To distinguish between real harmonics and interference, this embodiment introduces phase information. The instantaneous phase change rate at each frequency channel along the time axis is calculated, thereby estimating the actual center frequency of that channel. f est ( t The actual center frequency is compared with the theoretical harmonic frequency to calculate the harmonic membership degree. H ( f , t The calculation formula is: (15) in, H ( f , t ) represents the harmonic membership degree. f est ( t )for t The actual center frequency is estimated based on the instantaneous phase change rate. f 0 is the fundamental frequency. n For harmonic order, λ This is the attenuation adjustment factor. This parameter decays rapidly when the frequency deviation exceeds ±0.5Hz.

[0039] Among them, based on complex wavelet coefficients W ( a , b The instantaneous phase is calculated using the real and imaginary parts of the equation: (16) in, Representing scale a Time and location b The instantaneous phase at that point, This indicates taking the imaginary part. Indicates taking the real part, This represents the arctangent function in the four quadrants.

[0040] Phase expansion along the time axis is performed on the instantaneous phase to obtain the continuous phase. And calculate the actual center frequency based on the phase difference between adjacent time positions: (17) in, Representing scale a Time and location b The actual center frequency at that location, This represents the continuous phase after phase expansion. Indicates the time sampling interval.

[0041] Subsequently, a dual-channel time-frequency image is generated: Channel 1 (Amplitude Channel): Converts the original wavelet coefficient magnitudes | W ( a,b Multiply by harmonic membership degree H ( f , t As a new grayscale mapping source, an amplitude-weighted feature map is obtained; Channel 2 (Phase Consistency Channel): Calculates the instantaneous phase change pattern and extracts tightly coupled regions across scales to generate a phase consistency channel map.

[0042] Specifically, for each scale a Select its adjacent scales to form a set N a And calculate the phase consistency between adjacent scales: (18) in, C ( a , b ) indicates scale a Time and location b Cross-scale phase consistency at the location. N a Representing scale a The set of adjacent scales, Indicates adjacent scales. j It represents the imaginary unit. C ( a , b The closer the value is to 1, the more consistent the phase changes of adjacent scales are, and the tighter the cross-scale coupling is.

[0043] The time-frequency points with phase consistency greater than a preset threshold are identified as tightly coupled regions across scales, and the phase consistency is normalized to generate a phase consistency channel map.

[0044] Finally, the two channels are superimposed to generate a dual-channel time-frequency image, replacing the traditional single-channel grayscale image. This allows the CNN to utilize both amplitude and phase coherence information simultaneously, thereby automatically suppressing non-harmonic interference.

[0045] This embodiment uses the instantaneous phase change rate to calculate the harmonic membership degree and generate a phase consistency channel, which can effectively distinguish between real harmonics and interfering harmonics, suppress non-harmonic interference, and enhance the model's ability to identify real harmonic structures.

[0046] (iv) Dual-channel network architecture with embedded attention mechanism like Figure 3 As shown, this embodiment constructs an improved two-stage dual-channel convolutional neural network.

[0047] For the PCA feature branch, a multi-layer fully connected network is used for processing to obtain deep features. .

[0048] For the dual-channel time-frequency image branch, a novel architecture incorporating an attention mechanism was designed. Specifically: First, the dual-channel time-frequency image is input into the feature segmentation sub-network CNN to extract key feature segmentation maps; then, the key feature segmentation maps are input into the parallel CNN classification sub-network.

[0049] The parallel CNN classification sub-network includes a first CNN classification network, a second CNN classification network, and a time-frequency self-attention enhancement module (TSAEM module). The downsampling strategies and effective receptive fields of the two CNN classification networks are different. The output of the first CNN focuses on feature maps of locally salient time-frequency patterns, while the output of the second CNN retains feature maps of low-level edges, textures, and global background information.

[0050] The time-frequency self-attention enhancement module contains two parallel branches: Branch 1 (Frequency Axis): Performs sliding window self-attention calculations along the vertical axis of the image to capture long-range dependencies between different harmonic orders. This is particularly important for motor-type loads because their higher harmonic attenuation modes have global periodicity.

[0051] Branch 2 (Time Axis): Dilated Convolution Pyramid is used along the horizontal axis of the image to extract multi-scale transient structures.

[0052] The outputs of the two branches are gated and fused, and then residually connected with the original feature map to obtain the time-frequency enhanced fusion feature. In this embodiment, a time-frequency self-attention enhancement module is embedded in the network. This module significantly improves the model's ability to effectively model global time-frequency dependencies and multi-scale detailed features through frequency-axis self-attention and time-axis dilated convolutional pyramids, while strengthening the ability to capture global sequence dependencies and local high-frequency details.

[0053] Next, a dual-stream cross-attention fusion layer (DCFAL) is introduced. In this layer, the local salient feature map of the first CNN is used as the query, and the global background feature map of the second CNN is used as the key and value, resulting in a cross-enhanced feature map. Through the cross-attention computation mechanism, the global background feature map is used to enhance the context of the local salient feature map, achieving bidirectional enhancement of both local details and global context.

[0054] Finally, the cross-enhanced feature map and the time-frequency enhanced fusion feature are adaptively weighted and fused to obtain the time-frequency image enhancement feature. .

[0055] This embodiment introduces a dual-stream cross-attention fusion layer, which achieves bidirectional enhancement of local details and global context, strengthens temporal consistency and edge sharpness, improves the classification ability of nonlinear loads with large differences between transient and steady-state characteristics, and enhances the model's ability to identify different electrical appliance time-frequency coupling modes.

[0056] (v) Dynamic feature fusion and classification The time-frequency image enhancement features are fused with the deep features extracted by PCA. The fusion strategy employs a dynamic weighting mechanism, where weights are calculated and normalized using learnable parameters, adaptively adjusting the contribution of each channel based on different device characteristics. The weight calculation formula is as follows: (19) in, This represents the deep features extracted from the PCA-reduced electrical features. This represents the depth features extracted from the harmonic image. and The weights are dynamically calculated and satisfy... .

[0057] The weights are calculated in the following way: (20) (twenty one) in, This represents the sigmoid activation function. , , and These are learnable parameters.

[0058] Then the weights are normalized to ensure that their sum is 1: (twenty two) (twenty three) Finally, the fused features are input into a fully connected classifier, which outputs the load identification result.

[0059] (vi) Model training and optimization A combination of cross-entropy loss and focus loss is used as the optimization objective: (twenty four) (25) (26) Hyperparameters are optimized based on the principle of minimizing empirical risk. (27) Among them, L CE It is the standard cross-entropy loss, L focal To enhance learning on difficult samples, α and β are hyperparameters that balance the two losses, set to 0.7 and 0.3 respectively. This represents the true label of the i-th sample. This represents the predicted probability of the model output for the i-th sample. The modulation factor representing the focus loss is used to adjust the loss weights for easy and difficult samples. L val This represents the validation set loss. The base of the logarithm is typically 10.

[0060] The model is trained using the Adam optimizer with adaptive learning rate scheduling, implementing a multi-stage strategy that includes a warm-up phase, platform-based reduction, and minimum learning rate protection to ensure stable convergence.

[0061] To prevent overfitting and enhance generalization, several regularization techniques are integrated into the network. Dropout layers reduce neuron co-adaptation, while batch normalization after each layer stabilizes training dynamics. The model employs L2 weight regularization to penalize excessive model complexity, and data augmentation techniques for harmonic images, including geometric transformations and intensity adjustments. Gradient clipping ensures training stability throughout the optimization process. Early stopping is implemented to prevent overfitting, and the model is trained using mini-batch gradient descent until convergence.

[0062] During inference, the model processes the inputs through two channels and combines them via a feature fusion mechanism. Test-time enhancements generate multiple transformed versions of each input to improve recognition robustness. Performance is evaluated using standard classification metrics including accuracy, precision, recall, F1 score, and confusion matrix to provide a comprehensive assessment across different appliance types.

[0063] This invention utilizes a dual-channel convolutional neural network architecture to simultaneously process one-dimensional electrical features and dual-channel time-frequency images, achieving comprehensive utilization of multimodal information of electrical equipment. Furthermore, a dynamic weighted fusion mechanism adaptively adjusts the feature contribution based on the characteristics of different equipment to obtain the final load identification result. The proposed dual-modal architecture effectively captures the time-domain dynamic characteristics and frequency-domain features of electrical equipment, and the adaptive feature fusion strategy significantly improves load identification accuracy, thus providing an effective technical solution for smart grid applications.

[0064] Example 2 This embodiment provides a non-intrusive load monitoring system based on a dual-channel convolutional neural network, including: The one-dimensional sequence feature acquisition module is configured to extract comprehensive electrical features from the voltage and current measurement data of the electrical equipment to be identified and process them into one-dimensional sequence features. The sampling network construction module is configured to perform multi-scale morphological gradient detection based on the current signal of the electrical equipment to be identified, identify the true transient edge position and its steepness, and construct a nonlinear time-frequency sampling grid. The wavelet transform extraction module is configured to extract time-frequency features based on a nonlinear time-frequency sampling grid using continuous wavelet transform to obtain a two-dimensional time-frequency coefficient matrix. The time-frequency image construction module is configured to calculate harmonic membership based on phase information and construct a dual-channel time-frequency image containing amplitude and phase consistency channels by combining the two-dimensional time-frequency coefficient matrix. The fusion module is configured to use a two-stage convolutional neural network to process one-dimensional sequence features and dual-channel time-frequency images respectively, and perform bidirectional enhancement fusion to obtain fused features; The identification module is configured to input the fused features into the classification network to obtain the load identification result of the electrical equipment to be identified.

[0065] Example 3 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of a non-intrusive load monitoring method based on a dual-channel convolutional neural network as described in Embodiment 1 above.

[0066] Example 4 This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the non-intrusive load monitoring method based on a dual-channel convolutional neural network as described in Embodiment 1 above.

[0067] The steps or modules involved in Embodiments 2 to 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0068] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A non-invasive load monitoring method based on a dual-channel convolutional neural network, characterized in that, include: Extract comprehensive electrical features from the voltage and current measurement data of the electrical equipment to be identified, and process them into one-dimensional sequence features; Multi-scale morphological gradient detection is performed based on the current signal of the electrical equipment to be identified, the true transient edge position and its steepness are identified, and a nonlinear time-frequency sampling grid is constructed. Time-frequency features are extracted using continuous wavelet transform based on a nonlinear time-frequency sampling grid, resulting in a two-dimensional time-frequency coefficient matrix; Harmonic membership is calculated based on phase information, and a dual-channel time-frequency image containing amplitude and phase consistency channels is constructed by combining the two-dimensional time-frequency coefficient matrix. A two-stage convolutional neural network is used to process one-dimensional sequence features and dual-channel time-frequency images respectively, and bidirectional enhancement fusion is performed to obtain fused features; The fused features are input into the classification network to obtain the load identification results of the electrical equipment to be identified.

2. The non-intrusive load monitoring method based on a dual-channel convolutional neural network as described in claim 1, characterized in that, The comprehensive electrical characteristics include power-related indicators, waveform statistical characteristics, and frequency domain characteristics.

3. The non-intrusive load monitoring method based on a dual-channel convolutional neural network as described in claim 1, characterized in that, The multi-scale morphological gradient detection based on the current signal of the electrical equipment to be identified, to identify the true transient edge position and its steepness, specifically includes: The current signal of the electrical equipment to be identified is segmented, and multiple one-dimensional structural elements of different scales are constructed for each current signal segment to detect waveform abrupt changes. Among the one-dimensional structural elements, small-scale structural elements are used to detect short-term abrupt changes, and large-scale structural elements are used to suppress short-term glitch noise and detect waveform abrupt changes whose duration exceeds a preset threshold. Based on each scale, morphological dilation and erosion operations are performed on the current signal segment, and the difference between the dilation and erosion results is used as the morphological gradient. When the morphological gradients at the same location at multiple scales all exceed the preset value, the current location is determined as the true transient edge location; the normalized gradient response at the true transient edge location is determined as the transient edge steepness.

4. The non-intrusive load monitoring method based on a dual-channel convolutional neural network as described in claim 1, characterized in that, The construction of the nonlinear time-frequency sampling grid specifically includes: Logarithmic encrypted sampling is used on the time axis near the actual transient edge to obtain high time resolution; linear sparse sampling is used on the time axis in the steady-state region far away from the actual transient edge. On the frequency axis, a large-scale sampling density factor is used for low-frequency bands and a small-scale sampling density factor is used for high-frequency bands. By performing non-uniform sampling on the time axis and frequency axis respectively, a nonlinear time-frequency sampling grid is obtained.

5. The non-intrusive load monitoring method based on a dual-channel convolutional neural network as described in claim 1, characterized in that, The calculation of harmonic membership degree based on phase information is specifically as follows: ； in, H ( f , t ) represents the harmonic membership degree. f est ( t )for t The actual center frequency is estimated based on the instantaneous phase change rate. f 0 is the fundamental frequency. n For harmonic order, λ This is the attenuation adjustment coefficient.

6. The non-intrusive load monitoring method based on a dual-channel convolutional neural network as described in claim 1, characterized in that, The construction of a dual-channel time-frequency image, including an amplitude channel and a phase consistency channel, by combining the two-dimensional time-frequency coefficient matrix specifically includes: In the first channel, the original wavelet coefficient magnitude is multiplied by the harmonic membership degree as a new grayscale mapping source to obtain the amplitude-weighted feature map. In the second channel, the instantaneous phase change pattern is calculated and the tightly coupled cross-scale region is extracted to generate a phase consistency channel map; The amplitude-weighted feature map and the phase-consistency channel map are superimposed to generate a dual-channel time-frequency image.

7. The non-intrusive load monitoring method based on a dual-channel convolutional neural network as described in claim 1, characterized in that, The process employs a two-stage convolutional neural network to process one-dimensional sequence features and dual-channel time-frequency images respectively, and performs bidirectional enhancement fusion to obtain fused features, specifically including: The first stage uses a multi-layer fully connected network to process one-dimensional sequence features and obtain deep features; The second stage includes a feature segmentation sub-network (CNN), a parallel CNN classification sub-network, and a dual-stream cross-attention fusion layer. The CNN extracts key feature segmentation maps from the dual-channel time-frequency image and inputs them into the parallel CNN classification sub-network. The parallel CNN classification sub-network includes a first CNN classification network, a second CNN classification network, and a time-frequency self-attention enhancement module. The first CNN classification network extracts local salient time-frequency pattern feature maps, and the second CNN classification network extracts global background information feature maps. The time-frequency self-attention enhancement module performs enhancement processing along both the frequency and time axes to obtain time-frequency enhanced fusion features. In the dual-stream cross-attention fusion layer, the local salient feature map of the first CNN classification network is used as the query, and the global background feature map of the second CNN classification network is used as the key and value. Cross-attention calculation is performed to obtain a cross-enhanced feature map, which is then adaptively weighted and fused with the time-frequency enhanced fusion features to obtain the time-frequency image enhancement features. The final fused feature is obtained by fusing deep features and time-frequency image enhancement features.

8. A non-intrusive load monitoring system based on a dual-channel convolutional neural network, characterized in that, include: The one-dimensional sequence feature acquisition module is configured to extract comprehensive electrical features from the voltage and current measurement data of the electrical equipment to be identified and process them into one-dimensional sequence features. The sampling network construction module is configured to perform multi-scale morphological gradient detection based on the current signal of the electrical equipment to be identified, identify the true transient edge position and its steepness, and construct a nonlinear time-frequency sampling grid. The wavelet transform extraction module is configured to extract time-frequency features based on a nonlinear time-frequency sampling grid using continuous wavelet transform to obtain a two-dimensional time-frequency coefficient matrix. The time-frequency image construction module is configured to calculate harmonic membership based on phase information and construct a dual-channel time-frequency image containing amplitude and phase consistency channels by combining the two-dimensional time-frequency coefficient matrix. The fusion module is configured to use a two-stage convolutional neural network to process one-dimensional sequence features and dual-channel time-frequency images respectively, and perform bidirectional enhancement fusion to obtain fused features; The identification module is configured to input the fused features into the classification network to obtain the load identification result of the electrical equipment to be identified.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the non-intrusive load monitoring method based on a dual-channel convolutional neural network as described in any one of claims 1-7.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the non-intrusive load monitoring method based on a dual-channel convolutional neural network as described in any one of claims 1-7.

Images