A non-intrusive household appliance identification and monitoring method and system

Abstract

This invention belongs to the field of smart home technology, specifically disclosing a non-intrusive method and system for identifying and monitoring household appliances. The method includes: acquiring the output current signal of the main power strip in real time through a current acquisition device installed on the power cord of the main power strip and converting it into a discretized digital signal; dividing the discretized digital signal into multiple time windows; calculating the maximum adjacent difference peak value for the discrete signal in each time window; extracting transient and steady-state auxiliary features in each time window; and inputting the frequency domain harmonic features, maximum adjacent difference peak value, transient auxiliary features, and steady-state auxiliary features extracted from all time windows into a trained appliance identification model to obtain the operating status of each household appliance on the main power strip. This invention can focus on waveform features at key moments to obtain the fingerprint of the transient operation of appliances, and can accurately distinguish appliances with similar steady-state power but different starting characteristics.

Description

Technical Field

[0001] This invention relates to the field of smart home technology, and in particular to a non-invasive method and system for identifying and monitoring household appliances. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] In today's rapidly developing field of smart homes and energy management, non-invasive home appliance identification and monitoring technologies are becoming important tools for energy management, equipment automation, and power safety due to their unique flexibility and efficiency. However, traditional monitoring methods are mostly based on invasive sensor designs, requiring the one-to-one installation of detection equipment next to each device. These methods are mainly suitable for power monitoring of fixed and single devices, which greatly limits the technology's application potential in dynamic and complex environments.

[0004] Non-invasive household appliance identification technology can install a monitoring device at the bus end and combine it with intelligent algorithms to decompose and identify the operating status of multiple devices on the bus from the electrical characteristic parameters of the bus. For example, existing technologies disclose methods that use current transformers to collect bus current data and use Long Short-Term Memory (LSTM) network models to identify the switching events and types of appliances. However, simple current acquisition schemes are difficult to adapt to the wide dynamic range of household electricity consumption, from weak standby power consumption to the start-up of high-power appliances, resulting in insufficient or distorted signal acquisition accuracy in high-current or low-current regions. Although there are techniques that use operational amplifiers to set the gain, the current amplitude dynamic feedback in the household environment is wide and fluctuates drastically. The amplifier does not always operate in a single amplification stage but needs to switch between multiple states. In addition, traditional time-series models like LSTM are difficult to effectively capture long-range dependencies and multi-scale local features in the superimposed signals of multiple appliances. Under conditions where multiple devices are operating simultaneously or have similar load characteristics, it is difficult to distinguish the state of individual devices, limiting the performance upper limit of the identification system and resulting in deficiencies in the model's feature extraction capabilities and application specificity. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a non-invasive method and system for identifying and monitoring household appliances. Based on a current acquisition scheme using a current transformer and a fast relay-based intelligent dynamic switching transimpedance amplifier, it solves the accuracy and stability challenges of current acquisition over a wide dynamic range, ensuring lossless sampling across the entire dynamic range from weak standby to high-power startup. By performing Fast Fourier Transform and maximum adjacent difference peak detection on the raw electrical parameter data, a composite feature extraction scheme is constructed that considers both steady-state frequency domain harmonics and transient waveform characteristics. The Transformer-based identification model intelligently learns and integrates appliance features from different time scales, overcoming the identification bottleneck of traditional time-series models in scenarios with multiple concurrent appliances.

[0006] In some implementations, the following technical solutions are adopted:

[0007] A non-invasive method for identifying and monitoring household appliances, comprising:

[0008] The output current signal of the main power strip is collected in real time by a current acquisition device installed on the power cord of the main power strip, and the collected current signal is converted into a discrete digital signal.

[0009] The discretized digital signal is divided into multiple time windows using a local window of a predetermined length;

[0010] Fast Fourier Transform is performed on the discrete signal within each time window to extract its frequency domain harmonic features; the maximum adjacent difference peak value is obtained by calculating the difference between adjacent sampling points in the discrete signal sequence within the time window; and transient auxiliary features and steady-state auxiliary features within each time window are extracted simultaneously.

[0011] The frequency domain harmonic features, maximum neighbor difference peak values, transient auxiliary features, and steady-state auxiliary features extracted within all time windows are input into the trained appliance recognition model to obtain the working status of each household appliance on the main power strip.

[0012] As a further embodiment, the current acquisition device includes a current detection circuit, which includes: a current transformer, a power bus passing through the loop of the primary coil of the current transformer, the secondary coil of the current transformer being connected to a transimpedance amplifier, the non-inverting input terminal of the transimpedance amplifier being grounded, the inverting input terminal being connected to the secondary coil of the current transformer and a feedback resistor network respectively, and the output terminal generating a voltage signal proportional to the input current.

[0013] The feedback resistor network includes: resistor R10, resistor R11, and dual-channel relay U4; by controlling the switch of the dual-channel relay, resistor R10 or resistor R11 can be connected between the inverting input and output terminals of the transimpedance amplifier; the resistance value of resistor R10 is less than the resistance value of resistor R11.

[0014] As a further embodiment, the coil of the dual-channel relay U4 is connected to the source of the switching transistor Q2, and the gate of the switching transistor Q2 is connected to the control signal input terminal H7 through the resistor R14; the gate of the switching transistor Q2 is grounded after being connected to the pull-down resistor R15, and the drain of the switching transistor Q2 is grounded; a freewheeling diode D4 is connected in parallel across the coil of the dual-channel relay U4.

[0015] As a further embodiment, the current acquisition device also includes a step-down circuit. The step-down circuit includes a power input terminal P4, with a resistor R16, an optocoupler U7, and an output interface H1 connected in series at the live wire connection of the power input terminal P4. A diode D5 is connected in anti-parallel to the input terminal of the optocoupler U7, and the output terminal of the optocoupler U7 and the output interface H1 are connected to 5V through a pull-up resistor R17 to achieve a weak pull-up of the line, so that the output voltage is lower than 5V. The degree of conduction of the phototransistor inside the optocoupler U7 determines the magnitude of the output voltage drop.

[0016] As a further embodiment, the transient auxiliary features within each time window include the maximum instantaneous rate of change of current and the absolute amplitude of the transient peak current;

[0017] The maximum instantaneous rate of change of current is specifically the maximum change of current per unit time at the location of the maximum adjacent peak point; the absolute amplitude of the transient peak current is specifically the maximum absolute value of the current within the time window.

[0018] As a further embodiment, the steady-state auxiliary characteristics within each time window include the effective value of the current, active power, reactive power, multiple harmonics, and total harmonic current distortion rate; the total harmonic current distortion rate is specifically the ratio of the sum of the effective values ​​of all higher harmonic currents within the time window to the effective value of the fundamental current.

[0019] As a further embodiment, the appliance recognition model includes: an input layer, a preprocessing layer, a feature extraction layer, a feature fusion layer, a self-attention module, a Transformer encoding module, a classifier, and an output layer connected in sequence. The self-attention module incorporates residual connections. The feature fusion layer concatenates and compresses the multi-scale features extracted by the feature extraction layer into a unified feature representation, which serves as the input to the self-attention module. The self-attention module calculates the dependencies between features across different dimensions. The residual connections then add the output of the self-attention module to the feature representation of the original input to preserve the original feature information. The final output serves as the input to the Transformer encoding module, which captures the complex relationships and contextual information between features and weights them according to their correlation, thereby learning a higher level of correlation between different appliance behavior patterns.

[0020] As a further embodiment, the Transformer encoding module includes a time-frequency aware dynamic position encoder and a Transformer encoder; the time-frequency aware dynamic position encoder generates a fixed or time-based code, adds the code to the feature embedding of the input sequence, and then sends the result to the Transformer encoder.

[0021] As a further embodiment, the input of the time-frequency sensing dynamic position encoder includes: phase information of the fundamental frequency and important harmonic components within each time window. The physical feature vector consists of the maximum adjacent difference peak value, transient auxiliary features, and steady-state auxiliary features. and the gain switching status of the current acquisition circuit. ;

[0022] These input features are deeply fused using a gating network to obtain the fused encoding. ;

[0023] The obtained time-frequency sensing dynamic position code Specifically:

[0024] ;

[0025] in, Indicates a time step. Indicates the basic timing position code. The weights for the basic temporal position coding, To enhance the embedding of physical features, it is obtained through the physical prior information in the current signal.

[0026] In other embodiments, the following technical solutions are adopted:

[0027] A non-invasive home appliance identification and monitoring system, comprising:

[0028] The current acquisition device is installed on the power cord of the main power strip to acquire the output current signal of the main power strip in real time and convert the acquired current signal into a discrete digital signal.

[0029] The signal processing module is used to divide the discretized digital signal into multiple time windows through a local window of a predetermined length;

[0030] The feature extraction module performs a fast Fourier transform on the discrete signal within each time window to extract its frequency domain harmonic features; it calculates the difference between adjacent sampling points in the discrete signal sequence within the time window to obtain the maximum adjacent difference peak value; and it extracts transient auxiliary features and steady-state auxiliary features within each time window.

[0031] The state recognition module is used to input the frequency domain harmonic features, maximum adjacent difference peak values, transient auxiliary features and steady-state auxiliary features extracted within all time windows into the trained appliance recognition model to obtain the working status of each household appliance on the main power strip.

[0032] Compared with the prior art, the beneficial effects of the present invention are:

[0033] (1) This invention divides the original signal into multiple time windows, extracting frequency domain harmonic features, maximum adjacent difference peak values, transient auxiliary features, and steady-state auxiliary features within each time window. This enables the precise identification of the timing of appliance switching events. When an appliance turns on or off, the total load current undergoes significant transient changes. Maximum adjacent difference peak value detection can capture these dramatic points on the current waveform, accurately marking the time points of appliance state switching and providing a time reference for subsequent appliance event segmentation and identification. Simultaneously, it can better extract transient features. In addition to steady-state harmonic features, appliances also generate unique transient current waveforms at the moment of startup or shutdown. By identifying these peak change points, the system can focus on the waveform features at these critical moments, obtaining the fingerprint of the appliance's transient operation. This is crucial for distinguishing different types of appliances, especially those with similar steady-state power but different startup characteristics.

[0034] (2) The current acquisition device of this invention features a smart relay that dynamically switches the feedback resistor of the transimpedance amplifier. By pre-setting multiple high-precision feedback resistors and having a low-latency microprocessor quickly switch the relay according to the real-time current amplitude, the gain of the transimpedance amplifier is adjusted to the optimal state in a very short time. This ensures accurate sampling with high linearity and high signal-to-noise ratio, from the milliampere-level microcurrent in standby mode to the peak current of tens of amperes when starting high-power appliances. It effectively avoids signal distortion and sampling blind spots introduced when the gain is different; it not only improves the adaptability and measurement accuracy of the circuit, but also supports diverse power application scenarios, providing an efficient data foundation for energy management systems.

[0035] (3) The appliance identification model of this invention, based on Transformer, with its adaptive multi-scale dilated convolution and cross-scale attention mechanism optimized for non-intrusive load monitoring tasks, can intelligently learn and fuse appliance features at different time scales, overcoming the identification bottleneck of traditional time-series models in scenarios with multiple appliances operating concurrently. This not only significantly improves the system's identification accuracy when load features are similar or electricity consumption behavior fluctuates frequently, but also enhances its dynamic adaptability and compatibility when facing unknown appliances and complex environments, enabling it to achieve wide and flexible applications in home, commercial, and even industrial environments.

[0036] Other features and advantages 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

[0037] Figure 1 This is a flowchart of the non-invasive household appliance identification and monitoring method in an embodiment of the present invention;

[0038] Figure 2 This is a schematic diagram of the current detection circuit structure in the current acquisition device of the present invention.

[0039] Figure 3 This is a schematic diagram of the step-down circuit structure in the current acquisition device of this invention.

[0040] Figure 4 This is a schematic diagram of the electrical appliance identification model structure in an embodiment of the present invention. Detailed Implementation

[0041] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0042] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0043] Example 1

[0044] In one or more embodiments, a non-intrusive method for identifying and monitoring household appliances is disclosed, which identifies all electrical devices in operation by aggregating power signals acquired from the main power strip at the power inlet. Given an instantaneous current sequence i collected at the main inlet... t The goal is to predict the set of active electrical appliances A={a1,a2,…,a3} within a specific time period. k} and its operating status. Among them, the specific time period refers to the time range covered by the current signal data currently used for analysis. In this embodiment, the model performs inference and prediction within each local time window; active appliances refer to the set of all electrical devices that the system judges to be in an on or non-standby operating state within the specific time period; operating status refers to the specific operating mode of each "active appliance" within the current "specific time period", such as air conditioners having cooling and heating modes.

[0045] Since the voltage amplitude of the power supply bus is relatively stable, its phase characteristics are of crucial significance and can be used as a phase reference for the current signal. This embodiment employs high-sampling-rate data acquisition to capture transient features. After preprocessing and feature extraction, the raw data is converted into a low-dimensional feature vector. This is used to characterize the aggregated load characteristics. After training, the constructed appliance identification model can output an N-dimensional binary vector. Where N is the total number of identifiable appliance types, and the elements in the vector are... This indicates that the j-th type of electrical appliance is in a working state (1) or a non-working state (0).

[0046] As a specific implementation method, combined with Figure 1 The non-invasive household appliance identification and monitoring method in this embodiment specifically includes the following process:

[0047] S101: The current acquisition device installed on the power cord of the main power strip acquires the output current signal of the main power strip in real time and converts the acquired current signal into a discrete digital signal.

[0048] In this embodiment, the current acquisition device is responsible for high-precision, high-speed acquisition of aggregated current signals on the main power supply line. Installed on the power line of the main power strip, the device monitors current changes at the sockets in real time via a current detection circuit, ensuring accurate monitoring of the current flowing through the line. The acquired signal is transmitted to a data acquisition card, where a high-precision analog-to-digital converter converts it into a digital signal, which is then transmitted to the RK3588 processor chip for processing and analysis via the SPI data transmission protocol. This design allows the current acquisition device to be seamlessly integrated into existing home or industrial electrical systems without requiring large-scale modifications to the original circuit layout. It enables real-time monitoring of the current information of all appliances in the entire circuit, providing a solid data foundation for subsequent energy consumption analysis and load identification, and supporting equipment management and optimization in complex environments.

[0049] This embodiment's ultra-wide dynamic range current detection circuit employs a 20 / 1 current transformer to reduce the current parameter, effectively converting large currents into small current signals under high current conditions. This ensures that the measurement and control system obtains accurate and stable data during analysis and control. The current transformer uses magnetic coupling between the primary and secondary coils on the power line. When properly connected and configured, it can effectively transfer large currents to the secondary coil, outputting a small current signal proportional to the original current, while also exhibiting natural isolation from the mains power. This technology provides strong adaptability to current variations; even when current fluctuates significantly when appliances are turned on, the current transformer can still provide stable and accurate measurements, ensuring the functionality of the circuit system.

[0050] This embodiment employs a current sensing scheme using a current transformer (CT) and a transimpedance amplifier (TIA). The design securely and isolatedly acquires the proportional signal of the bus load current through the current transformer, which is then converted into a voltage signal by a transimpedance amplifier. To achieve a wide dynamic range and high-precision current sampling, the circuit uses a relay to dynamically switch the feedback resistor of the transimpedance amplifier, thereby adjusting the gain to accommodate different current input amplitudes. Ultimately, it outputs a conditioned voltage signal that precisely corresponds to the original load current waveform, providing a basis for non-intrusive load monitoring and analysis in subsequent systems.

[0051] Combination Figure 2 The current detection circuit specifically includes: a current transformer, a power bus passing through the loop of the primary coil of the current transformer, the secondary coil of the current transformer connected to a transimpedance amplifier, the non-inverting input terminal of the transimpedance amplifier grounded, the inverting input terminal connected to the secondary coil of the current transformer and the feedback resistor network respectively, and the output terminal generating a voltage signal proportional to the input current.

[0052] Specifically, Figure 2 In this circuit, P6 is the input terminal of the current transformer (CT). Pins 1 and 2 of P6 are connected to the secondary coil of the current transformer to receive the AC current signal induced from the primary coil. U6.1 (LM358ADR2G) is an operational amplifier configured as a transimpedance amplifier (TIA) in this circuit. The non-inverting input (pin 3) of operational amplifier U6.1 is connected to ground (GND) to provide a stable reference potential; its inverting input (pin 2) is connected to pin 2 of P6 and the feedback resistor network controlled by relay U4. Its output (pin 1) generates a voltage signal proportional to the input current. The LM358ADR2G is a low-power, dual operational amplifier suitable for battery-powered applications, with a wide supply voltage range, meeting the low power consumption and accuracy requirements of this device.

[0053] Resistors R10 and R11 form the feedback resistor network of the transimpedance amplifier. These resistors are switched via relay U4 to change the gain of the transimpedance amplifier. The value of the feedback resistor determines the current-to-voltage conversion ratio, i.e., the conversion factor.

[0054] Specifically, the feedback resistor network includes: resistors R10 and R11, and a dual-channel relay U4; by controlling the switch of the dual-channel relay, either resistor R10 or R11 is connected between the inverting input and output of the transimpedance amplifier; the resistance value of resistor R10 is less than the resistance value of resistor R11; the coil of the dual-channel relay U4 is connected to the source of the switching transistor Q2, and the gate of the switching transistor Q2 is connected to the control signal input terminal H7 through resistor R14; the gate of the switching transistor Q2 is grounded after being connected to the pull-down resistor R15, and the drain of the switching transistor Q2 is grounded; a freewheeling diode D4 is connected in parallel across the coil of the dual-channel relay U4.

[0055] Dual-channel relay U4 is a dual-channel relay used to dynamically switch the feedback resistor of a transimpedance amplifier. The relay coil is controlled by switching transistor Q2 and resistor R14, with the drive signal coming from control signal input H7. When the relay is energized, its normally closed contact opens and its normally open contact closes; when the relay is de-energized, its normally closed contact closes and its normally open contact opens. When control signal input H7 energizes the relay coil, the normally open contact of dual-channel relay U4 closes. At this time, resistor R10 is connected to the inverting input of U6.1 through pins 4-5 of dual-channel relay U4, forming a small feedback resistance path, providing lower gain, suitable for measuring large currents. When control signal input H7 de-energizes the relay coil, the normally closed contact of dual-channel relay U4 closes. At this time, resistor R11 is connected to the inverting input of U6.1 through pins 3-6 of dual-channel relay U4, forming a larger feedback resistance path, providing higher gain, suitable for measuring small currents, thus extending the dynamic range of current measurement.

[0056] Switch Q2 is an N-channel MOSFET. Resistors R14 and R15 drive the dual-channel relay U4. The control signal input terminal H7 provides the control signal, which is connected to the gate of switch Q2 through resistor R14. Resistor R15 is a gate pull-down resistor, ensuring that switch Q2 is reliably turned off when there is no control signal. When the control signal input terminal H7 outputs a high level, switch Q2 turns on, providing a path to the relay coil, and the relay engages. When H7 outputs a low level, switch Q2 turns off, and the relay coil is de-energized. Diode D4 is a freewheeling diode, connected in parallel with the relay coil, used to release the induced electromotive force when the relay coil is de-energized, protecting switch Q2 from reverse voltage surges.

[0057] H4 is the signal output interface. Pin 1 of H4 is connected to the conditioned analog voltage signal output terminal, and pin 2 is grounded. The output signal is an analog voltage after being converted by a transimpedance amplifier, gain adjusted, and level biased. Its waveform accurately corresponds to the waveform of the load current and can be sampled, analyzed, and processed by a subsequent analog-to-digital converter (ADC) or microcontroller (MCU) to achieve non-intrusive load monitoring.

[0058] This embodiment combines a current transformer with a smart relay to dynamically switch the feedback resistor of the transimpedance amplifier. This design addresses the issues of insufficient acquisition accuracy and distortion caused by the wide dynamic range and significant fluctuations of current amplitude in household environments. Compared to the accuracy loss of a simple transformer solution over a wide dynamic range, this embodiment uses multiple preset high-precision feedback resistors and a low-latency microprocessor to rapidly switch the relay based on the real-time current amplitude. This allows the transimpedance amplifier gain to be adjusted to its optimal state in a very short time with extremely high accuracy. This ensures accurate sampling with high linearity and high signal-to-noise ratio, from milliampere-level microcurrents in standby mode to peak currents of tens of amperes when starting high-power appliances. This intelligent dynamic switching mechanism effectively avoids signal distortion and sampling blind spots introduced when the gain is different. This method not only improves the adaptability and measurement accuracy of the circuit but also supports diverse power application scenarios, providing an efficient data foundation for energy management systems.

[0059] In this embodiment, the current acquisition device also includes a power supply system, comprising a regulated power supply, a step-down circuit, and a phase detection circuit. The regulated power supply ensures a stable power supply to the entire circuit; the step-down circuit converts the 220V AC power to a voltage range of less than 20V suitable for the chip and operational amplifier; and the phase detection circuit ensures that each sampling cycle begins at the zero-phase point of the power frequency voltage, achieving accurate acquisition of voltage phase information.

[0060] The step-down circuit reduces the 220V AC power to an appropriate level while preserving the power frequency characteristics of the AC signal. Utilizing half-wave characteristics ensures accurate phase of the output AC signal, enhancing the overall stability and reliability of the circuit and preventing equipment damage due to voltage overload. Furthermore, optocouplers provide electrical isolation to ensure the safe operation of the microprocessor; combined with… Figure 3 The step-down circuit specifically includes a power input terminal P4. The live wire of the power input terminal P4 is connected in series with a resistor R16, an optocoupler U7, and an output interface H1. An anti-parallel diode D5 is connected to the input of the optocoupler U7. A pull-up resistor connects the output of the optocoupler U7 to the 5V output, providing a weak pull-up. When the phototransistor inside the optocoupler's output side is turned on, it shunts current through the pull-up resistor. The current through the pull-up resistor creates a voltage drop, causing the output voltage to be lower than 5V. The degree of conduction of the phototransistor inside the optocoupler U7 (i.e., the magnitude of the current flowing from the collector to the emitter) determines the magnitude of the output voltage drop. Therefore, the pull-up resistor converts the photosensitive signal into a measurable voltage change.

[0061] Specifically, power input terminal P4 is the AC power input terminal. Pins 1 and 2 of P4 are connected to the AC neutral (N) and live (L) wires, respectively, which are high-voltage hazardous areas. Resistor R16 is a current-limiting resistor, its main function being to limit the current flowing through the optocoupler U7 and the protection diode D5; for a 220V AC input, its peak voltage is approximately 220V*. ≈311V. When the AC voltage is at its peak, the extremely large current can burn out the optocoupler and protection diode D5. Resistor R16 ensures that the current is within the safe operating range of these devices.

[0062] In this circuit, the protection diode D5, acting as a fast-switching diode, is connected in reverse parallel with the input of optocoupler U7 (between pins 1 and 2). During the positive AC half-cycle, the voltage at pin 1 of the power input terminal P4 is higher than that at pin 2. At this time, the optocoupler diode of optocoupler U7 conducts, and protection diode D5 is in reverse cutoff. During the negative AC half-cycle, the voltage at pin 1 of the power input terminal P4 is lower than that at pin 2. At this time, D5 conducts in the forward direction, clamping the reverse voltage across the internal diode of optocoupler U7 to its forward conduction voltage (approximately 0.7V). Optocoupler U7 provides electrical isolation between the high-voltage AC side and the low-voltage DC digital side (typically capable of withstanding several kilovolts of isolation voltage). It consists of an infrared light-emitting diode (LED) and a phototransistor, packaged together but electrically isolated. This effectively protects the subsequent ADC chip from high-voltage surges and eliminates ground loop problems. The core function of protection diode D5 is to protect the LED diode of optocoupler U7; the internal LED of optocoupler U7 typically has a very low reverse breakdown voltage (e.g., around 5V). During the negative half-cycle of AC, without the protection diode D5, the LED would be damaged by a reverse voltage far exceeding its tolerance; D5 provides a path for the current during the negative half-cycle and protects the LED.

[0063] Output interface H1 provides an interface for the subsequent processor chip. Pin 1 outputs +5V power (which can power the MCU or other peripherals, or serve as a reference only). Pin 2 is used as a signal output, connected to the collector of optocoupler U7, and is also an input pin of the subsequent processor chip. This signal is a square wave (or approximately a square wave): during the positive half-cycle of the AC current (and when the voltage is higher than the optocoupler's turn-on threshold), the output is low (close to 0V); during the negative half-cycle and the positive half-cycle of the AC current, when the voltage is low near zero, the output is high (+5V).

[0064] This embodiment uses optocouplers to achieve electrical isolation, ensuring the safe operation of the microprocessor. This not only avoids the losses and errors that may occur in traditional voltage conversion, but also improves the deployment efficiency of the circuit system in smart homes, ensuring that the system can respond quickly and process each electrical cycle accurately, providing a stable foundation for subsequent current acquisition and data analysis.

[0065] After collecting the aggregated current waveform it is digitized with a high sampling rate, and the sampling frequency is set to 3.2 kHz to ensure that harmonics up to the 30th harmonic (corresponding to the 50 Hz fundamental frequency) can be accurately captured. With this high sampling rate, not only can harmonic components be accurately captured, but transient characteristics during electrical appliance startup or shutdown can also be effectively captured.

[0066] The digitized discrete-time current signal is represented as , where n is the sampling point index. The discrete signal is subjected to a Fourier transform to obtain its frequency-domain representation. In this embodiment, the fast Fourier transform (FFT) algorithm is used for processing, and its calculation formula is as follows:

[0067] ;

[0068] where X[k] represents the complex value of the spectral component at frequency index k, and N is the number of sampling points used in the FFT calculation. After obtaining the spectrum X[k], its harmonic components are extracted as features. Specifically, in this embodiment, since the contrast of high-order harmonics is not strong, and in order to optimize the subsequent data processing speed and reduce the storage capacity requirement of the controller, the amplitude information of the first 7 harmonics (i.e., the fundamental wave to the 7th harmonic) is selected for extraction and analysis. These harmonic components can effectively characterize the unique harmonic feature spectra of different types of electrical appliances (e.g., devices using switched power supplies, devices with motors, or purely resistive loads) in their operating states.

[0069] S102: The discretized digital signal is divided into multiple time windows through a local window of a predetermined length.

[0070] In this embodiment, to ensure the robustness of the algorithm, a local window mechanism is introduced. Its specific implementation is as follows: When extracting features from the digitized discrete-time current signal the system does not process the entire long time series at once, but divides the data stream into a series of overlapping time windows of a fixed length.

[0071] Specifically, a fixed window length L is set, and L corresponds to several power frequency cycles and a sliding step S (S < L). The current sampling sequence is divided into a series of subsequences, that is, time windows . The kth time window , contains data from sampling point k S to k S + L 1, where k is the window index.

[0072] ;

[0073] To reduce spectral leakage, before performing the Fourier transform on the subsequences within each window, they are first multiplied by a window function w[n] (Hamming window), and the processed window data is .

[0074] For each independent time window processed by the window function and the data it contains , the system independently performs a fast Fourier transform (FFT), extracts its harmonic features in the frequency domain, and applies the maximum adjacent difference peak detection to obtain its transient and steady-state features in the time domain.

[0075] In this embodiment, through the design of the local window mechanism, the system can effectively handle local transient changes and non-steady signals, such as current spikes when an electrical appliance starts or shuts down, local harmonic distortions caused by load type switching, or short-term power grid disturbances. By continuously sliding these overlapping windows on the time axis, the system can achieve continuous monitoring and dynamic capture of signal features, and update the operating state of the electrical appliance in real time. In addition, the overlapping design of the windows (S < L) allows the algorithm to share information in adjacent time periods, enhancing the continuity and stability of the features, effectively reducing the risk of feature loss at the window boundaries, and facilitating the system's real-time response to dynamic power load changes, improving the flexibility and practicality of data processing, thereby ensuring that feature detection remains efficient and reliable in a complex power environment.

[0076] S103: Perform a fast Fourier transform on the discrete signals within each time window to extract their harmonic features in the frequency domain; obtain the maximum adjacent difference peak by calculating the differences between adjacent sampling points in the discrete signal sequence within the time window; meanwhile, extract the transient auxiliary features and steady-state auxiliary features within each time window.

[0077] In this embodiment, for all electrical devices to be identified, the amplitude information of the first 7 harmonics (i.e., the fundamental wave to the 7th harmonic) is uniformly extracted. These harmonic components (including the fundamental wave amplitude, second harmonic amplitude,..., seventh harmonic amplitude) constitute the main features of the device in the frequency domain, which are sufficient to characterize the unique harmonic feature spectra of different types of electrical appliances (e.g., devices using a switching power supply, devices with a motor, or pure resistive loads) in the operating state. Considering that the contrast of high-order harmonics is not strong, and to optimize the subsequent data processing speed and reduce the storage capacity requirement of the controller, selecting the first 7 harmonics can balance the calculation efficiency while maintaining the discrimination.

[0078] Simultaneously, by using maximum adjacent peak difference detection, the precise timing and transient characteristics of appliance switching events can be accurately captured. When an appliance turns on or off, the total load current undergoes significant transient changes. Maximum adjacent peak difference detection can capture these dramatic points on the current waveform, and its characteristics contain the precise timing of appliance state switching, providing a benchmark for subsequent neural network models to segment and identify appliance events.

[0079] For each local window Current sampling data within Maximum adjacent difference peak detection calculates the difference between adjacent sampling points in the sequence and identifies the location point with the largest difference.

[0080] For a discrete current sequence The difference sequence of its adjacent points Defined as:

[0081] ;

[0082] in, n This is the index for the sampling point. The goal of the algorithm is to find the index that makes... Index of the maximum value:

[0083] ;

[0084] in, It is the point with the largest adjacent difference peak.

[0085] In addition, to further enhance the ability to identify multiple appliances operating concurrently, with similar load characteristics, and weak events in complex home scenarios, this embodiment innovatively introduces a collaborative extraction mechanism of multidimensional transient waveforms and steady-state auxiliary features.

[0086] Specifically, transient auxiliary features can capture and quantify the details of non-stationary, high-frequency or low-frequency transient waveforms at the moment of appliance startup or shutdown with high resolution. This effectively distinguishes the differences between startup ripple and smooth changes in resistive loads, overcoming the limitations of FFT in analyzing non-stationary signals. These features can quantify the rapid changes in current signals when an appliance undergoes state changes (such as turning on, turning off, or mode switching).

[0087] Transient auxiliary features specifically include:

[0088] (1) Maximum instantaneous rate of change of current .

[0089] The maximum instantaneous rate of change of current reflects the intensity of the impact of an electrical appliance on the mains current over a very short period of time. High-power inductive loads (such as motor starting) or switching power supplies (such as computers and chargers) cause a rapid rise in current at the moment of startup, with significant differences in their characteristics. Different electrical appliances have their specific ranges and patterns of instantaneous rate of change of current.

[0090] The maximum instantaneous rate of change of current is specifically defined as the maximum adjacent peak value detected. At this point, calculate the maximum change in current per unit time:

[0091] ;

[0092] in, This indicates the point of maximum adjacent difference peak. Current sampling value at the location, This represents the current sample value in the sampling period preceding the point of maximum adjacent difference peak. The sampling period.

[0093] (2) Absolute amplitude of transient peak current .

[0094] The absolute amplitude of the transient peak current is used to capture the inrush current during appliance startup. Specifically, it is calculated as follows: the absolute maximum value of the current within the local window containing the point of maximum adjacent peak difference:

[0095] ;

[0096] in, This represents the current value within a local window.

[0097] Steady-state auxiliary features reflect the electrical characteristics of an appliance under stable operating conditions and are crucial for identifying the type and mode of continuous operation of appliances. These features are typically extracted within a stable operating cycle of the appliance; specific steady-state auxiliary features include the following:

[0098] (1) Effective value of current .

[0099] The effective value of the current is the root mean square value of the current signal calculated over one or more power frequency cycles.

[0100] ;

[0101] in, This indicates the number of sampling points collected within one complete power frequency cycle. Using multiple power frequency cycles will yield a smoother RMS current value.

[0102] (2) Active power .

[0103] Active power specifically refers to the active power of the bus within one or more power frequency cycles:

[0104] ;

[0105] in, , , These represent the number of sampling points per power frequency cycle, the voltage sampling value, and the current sampling value, respectively.

[0106] (3) Reactive power .

[0107] Reactive power is calculated using the 90-degree phase difference component of current and voltage, representing the reactive component of bus power. Common methods include utilizing the imaginary part of the Fourier transform or calculating apparent power and active power.

[0108] ;

[0109] in, , , , and represent the amplitude and phase angle of the h-th harmonic voltage and current, respectively.

[0110] (4) Multiple harmonic characteristics, calculated based on the aforementioned FFT algorithm.

[0111] (5) Total harmonic current distortion rate .

[0112] The total harmonic current distortion rate is the ratio of the effective value of higher harmonic currents to the effective value of the fundamental current within the local time window of calculation.

[0113] ;

[0114] in, This is the effective value of the fundamental current. This is the effective value of the h-th harmonic current.

[0115] The collaborative extraction mechanism of multidimensional transient waveforms and steady-state auxiliary features in this embodiment ensures that the system can not only identify the occurrence of events, but also accurately characterize the unique "electrical fingerprint" of electrical appliances during switching and steady-state operation from multiple dimensions. This greatly improves the recognition accuracy when load characteristics are highly similar or electrical behavior fluctuates frequently, especially demonstrating excellent performance when distinguishing complex loads such as motor-type appliances and switching power supply-type appliances. The fast FFT algorithm efficiently reduces the number of calculation multiplications, and the computational savings increase significantly, especially when the number of sampling points increases. This allows the system to quickly analyze complex frequency domain signals, ensuring signal integrity and analysis accuracy. At the same time, the maximum adjacent difference peak detection method quickly identifies the position with the largest difference by calculating the difference between adjacent points in the sequence, providing accurate feature points for electrical signal monitoring and recognition.

[0116] S104: Input the frequency domain harmonic features, maximum neighbor difference peak value, transient auxiliary features and steady-state auxiliary features extracted within all time windows into the trained appliance recognition model to obtain the working status of each household appliance on the main power strip.

[0117] In this embodiment, the frequency domain harmonic features, maximum neighbor difference peak value, transient auxiliary features and steady-state auxiliary features extracted within each time window are combined to form a multi-dimensional feature vector. The multi-dimensional feature vectors of all windows are used as input features of the electrical appliance identification model. The model will learn the complex nonlinear mapping relationship between these combined features and the working state of the electrical appliance, thereby achieving accurate identification and monitoring of the electrical appliance.

[0118] To address the fundamental shortcomings of existing traditional time-series models in handling complex household electricity scenarios with multiple concurrent appliances and long-range dependencies, as well as their inability to effectively capture subtle and critical transient features in current signals, leading to decreased recognition accuracy, this embodiment proposes an innovative, highly adaptive, and deeply integrated time-scale information-based Transformer-based recognition model. The designed Transformer model is specifically customized and optimized for electrical features, wide dynamic range signals, and multiple concurrent appliance detection problems in non-intrusive load monitoring (NILM) tasks.

[0119] Specifically, in combination Figure 4The appliance recognition model comprises an input layer, a preprocessing layer, a feature extraction layer, a feature fusion layer, a self-attention module, a Transformer encoding module, a classifier, and an output layer, connected sequentially. The self-attention module introduces residual connections. The feature fusion layer concatenates and compresses the multi-scale features extracted by the feature extraction layer into a unified feature representation, which serves as the input to the self-attention module. The self-attention module calculates the dependencies between features across different dimensions. The residual connections then add the output of the self-attention module to the original input feature representation to preserve original feature information and stabilize the training process, avoiding information loss and gradient problems. The final output serves as the input to the Transformer encoding module, which captures complex relationships and contextual information between features and weights them according to their correlations, thereby learning a higher level of correlation between different appliance behavior patterns.

[0120] Specifically, the input features are fed into the input layer and first preprocessed through a linear preprocessing layer to map the feature dimensions from the original dimensions to 64 dimensions (feature_dim, 64). Then, a non-linear transformation is performed using batch normalization (BatchNorm1d) and the ReLU activation function. Finally, a dropout layer is used to prevent overfitting. This preprocessing step aims to standardize the input data and increase the model's expressive power.

[0121] The feature extraction layer uses four convolutional layers (conv1d) with different dilation rates (1, 2, 4, and 8) to extract multi-scale features. The output of each convolutional layer is reduced in dimensionality through average pooling, and finally, the features from different scales are concatenated. This multi-dilation rate convolutional structure (or dilated convolution) allows the model to increase its receptive field without increasing parameters or computational cost, capturing dependencies and local patterns at different time scales, and effectively extracting both transient and steady-state features.

[0122] The original features are concatenated with the outputs of four dilated convolutions, and then fused through a fusion layer (containing a linear layer, batch normalization, ReLU activation, and Dropout) to finally output a 64-dimensional feature vector. This fusion strategy ensures that the model can comprehensively utilize the detailed information of the original features and the high-level features extracted by multi-scale convolutions.

[0123] In this embodiment, the combination of the self-attention module, the Transformer encoding module, and the residual connection network is key to the model's ability to capture long-range dependencies and achieve high-level semantic understanding. The 64-dimensional feature vector output from the feature fusion layer serves as the input to the self-attention module. (where B is the batch size) =64 is the feature dimension) First, a linear transformation is performed to generate query Q, key K, and value V, that is , , ,in , , The weight matrix is ​​learnable. These parameters are then fed into a multi-head self-attention mechanism.

[0124] To mitigate the impact of gradients during deep network training and allow the network to learn identity mappings, this model innovatively introduces residual connections in the self-attention module. For example... Figure 4 As shown, the input features of the self-attention module are connected via residual connections. (i.e., the output of the fusion layer) is directly related to the output of the multi-head self-attention mechanism. To perform element-wise addition. Its mathematical expression is:

[0125] ;

[0126] This embodiment, through the design of residual connections, ensures that the original input information can be directly passed to deeper layers of the network. Even if the attention mechanism fails to capture effective information in some layers, the model can still utilize the original features by "skipping" the attention layer. This significantly improves the stability and convergence speed of training while maintaining the model's expressive power, which is especially crucial for network structures that capture long-range dependencies. The output after residual connections is then processed through layer normalization and a feedforward network, further enhancing the expressive power of the features, together forming a sublayer of a Transformer encoder.

[0127] The core structure of the Transformer encoding module consists of a customized adaptive scale-time Transformer encoder and a NILM (Non-Intrusive Load Monitoring) specifically enhanced position encoder, working in conjunction with a multi-label classification head. The Transformer encoding module uses a four-layer Transformer encoder, with each layer containing at least four attention heads, aiming to progressively abstract and refine more discriminative patterns of electrical appliance behavior through a stacked encoder structure.

[0128] In the Transformer encoding module, the position encoder generates a fixed or time-based code, which is then added to the word embedding (or feature embedding) of the input sequence. The result is then fed into the Transformer encoder. In other words, the position encoder acts as a preprocessing step for the Transformer encoder.

[0129] In this embodiment, the NILM-specific enhanced position encoder is embodied as a time-frequency sensing dynamic position encoder. The inputs of the time-frequency sensing dynamic position encoder include:

[0130] Phase information of the fundamental frequency and important harmonic components within each time window The physical feature vector consists of the maximum adjacent difference peak value, transient auxiliary features, and steady-state auxiliary features. and the gain switching status of the current acquisition circuit. (Based on the actual gain of the circuit); these input features are deeply fused through a gated network to obtain... ;

[0131] Obtain time-frequency sensing dynamic position code Specifically:

[0132] ;

[0133] in, Indicates a time step. Indicates the basic timing position code. The weights for the basic temporal position coding, It is a basic temporal position code, learned through network training, which provides the most basic temporal information for each time step in the sequence. This is a learnable weight matrix used to linearly transform this basic positional encoding, allowing it to be more flexibly integrated into the feature space of the Transformer model. This represents a high-dimensional augmented physical feature embedding, which maps the physical prior information unique to the current signal (such as frequency domain features, event features, and gain states) to a representation consistent with the model feature dimension through a small fully connected network.

[0134] When the model receives the input sequence, the gating network dynamically adjusts the weights and form of the positional encoding based on the frequency domain and event characteristics within the current time window. This allows the encoding to more strongly emphasize key transient moments and energy change cycles of electrical appliances. This dynamic and physically aware positional encoding enables the Transformer encoding module not only to understand the temporal order of the sequence but also to deeply understand the physical meaning and context of electrical events within the current signal. This significantly enhances the model's contextual understanding of electrical events and its sensitivity to specific appliance behavior patterns. This design allows the model to adjust and fuse temporal positional information with electrical characteristics at each computational step of the Transformer encoding module based on the real-time physical characteristics of the current signal. This method not only improves the Transformer's accuracy in capturing long-range electrical dependencies but also significantly enhances its ability to identify subtle transient features of electrical appliances.

[0135] Each Transformer encoder layer internally follows the basic structures of "multi-head self-attention + residual connections + layer normalization" and "feedforward network + residual connections + layer normalization". This encoder structure can effectively capture the complex relationships and contextual information between features, enabling the model to learn higher-level correlations between different appliance behavior patterns. Through its internal self-attention mechanism, the Transformer encoder allows each input feature in the model to "attention" to all other features in the input sequence and weights them according to their correlations, thereby capturing long-range dependencies. This has significant advantages in household appliance recognition tasks where the operating states of different appliances may influence each other, or where certain event features require a long time window to be fully understood. By stacking multiple encoder layers, the model can progressively abstract and refine more discriminative feature representations.

[0136] The model's classifier consists of two linear layers, with a non-linear transformation in between using batch normalization, the GELU activation function, and a Dropout layer to enhance the model's generalization ability and prevent overfitting. The final output is the predicted probability for each appliance category, activated by a Sigmoid function; this Sigmoid function maps the output to a value between 0 and 1.

[0137] The model's output layer outputs a set, where each element corresponds to a known appliance category and indicates the appliance's current operating state. This state can be a simple binary value (on / off) or a multi-valued value (such as gear position or operating mode), depending on the granularity of the training data and the model's classification ability.

[0138] In this embodiment, when training the appliance recognition model, air conditioners (heating / cooling), heaters, hair dryers, electric cookers, electric fans, table lamps, and electric mosquito coils were used as sample appliances. Data sampling was performed on individual appliances and combinations of different appliances to construct a training dataset. The model outputs an 8-dimensional vector, with each dimension corresponding to an appliance or specific state, representing whether the appliance is currently working and the probability of that working state. For example, if the model outputs an 8-dimensional vector indicating that the probability of the air conditioner cooling and the electric fan starting is 80%, it means that these two loads are currently running on the bus.

[0139] During training, the binary cross-entropy loss function (BCEWithLogitsLoss) is used to calculate the error between the predicted values ​​and the true labels to optimize the model parameters. For this type of task, the binary cross-entropy loss can independently evaluate the prediction accuracy of each category, avoiding the assumption of label mutual exclusivity made by traditional multi-class loss functions (such as Softmax cross-entropy).

[0140] The ReduceLROnPlateau learning rate scheduler is used to dynamically adjust the learning rate based on the F1 score on the validation set. If the F1 score on the validation set does not improve for 10 consecutive rounds, the learning rate will be reduced to half its original value. This adaptive learning rate adjustment strategy helps the model escape local optima, accelerates convergence, and improves the model's final performance.

[0141] During training, data augmentation is performed on categories with insufficient sample sizes by adding Gaussian noise and random scaling to generate new samples and balance the data distribution. This data augmentation strategy can effectively alleviate the data imbalance problem, improve the model's ability to recognize uncommon patterns, and enhance the model's generalization and robustness.

[0142] The system continuously collects feature information from training data, including the multi-label distribution of features during the dynamic process, as a basis for optimizing model performance. The system transmits loss and metric information during training to the backend in real time, providing data analysis and chart display for interactive data processing. Operators then execute appropriate strategies based on this information to ensure model stability and convergence. A validation set configured during model training monitors the model's generalization ability, ensuring its robustness and accuracy in real-world applications.

[0143] This embodiment, based on the Transformer recognition model, leverages its adaptive multi-scale dilated convolution and cross-scale attention mechanism optimized for NILM tasks to intelligently learn and fuse appliance features across different time scales, overcoming the recognition bottleneck of traditional time-series models in scenarios with multiple concurrent appliances. This significantly improves the system's recognition accuracy when load characteristics are similar or electricity consumption behavior fluctuates frequently, and enhances its dynamic adaptability and compatibility when facing unknown appliances and complex environments, enabling its wide and flexible application in home, commercial, and even industrial environments. The real-time analysis capabilities of the built-in data processing module, combined with continuous optimization of model training, further enhance the practicality of energy consumption pattern assessment. This series of forward-looking innovative designs and functional integrations constitutes a highly efficient, stable, high-precision, and robust home appliance identification and monitoring solution, providing solid technical support for smart homes and refined energy management, and possessing undeniable broad application prospects and significant promotional value.

[0144] Example 2

[0145] In one or more embodiments, a non-invasive home appliance identification and monitoring system is disclosed, comprising:

[0146] The current acquisition device is installed on the power cord of the main power strip to acquire the output current signal of the main power strip in real time and convert the acquired current signal into a discrete digital signal.

[0147] The signal processing module is used to divide the discretized digital signal into multiple time windows through a local window of a predetermined length;

[0148] The feature extraction module performs a fast Fourier transform on the discrete signal within each time window to extract its frequency domain harmonic features; it calculates the difference between adjacent sampling points in the discrete signal sequence within the time window to obtain the maximum adjacent difference peak value; and it extracts transient auxiliary features and steady-state auxiliary features within each time window.

[0149] The state recognition module is used to input the frequency domain harmonic features, maximum adjacent difference peak values, transient auxiliary features and steady-state auxiliary features extracted within all time windows into the trained appliance recognition model to obtain the working status of each household appliance on the main power strip.

[0150] The specific implementation methods of the above modules are the same as those in Example 1, and will not be described in detail again.

[0151] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A non-invasive household appliance identification and monitoring method, characterized by, include: The output current signal of the main power strip is collected in real time by a current acquisition device installed on the power cord of the main power strip, and the collected current signal is converted into a discrete digital signal. The discretized digital signal is divided into multiple time windows using a local window of a predetermined length; Perform a Fast Fourier Transform on the discrete signal within each time window to extract its frequency domain harmonic features; The maximum adjacent difference peak value is obtained by calculating the difference between adjacent sampling points in the discrete signal sequence within the time window; at the same time, transient auxiliary features and steady-state auxiliary features within each time window are extracted. The frequency domain harmonic features, maximum neighbor difference peak values, transient auxiliary features and steady-state auxiliary features extracted within all time windows are input into the trained appliance recognition model to obtain the working status of each household appliance on the main power strip. The current acquisition device includes a current detection circuit, which includes a current transformer, a power bus passing through the loop of the primary coil of the current transformer, a secondary coil of the current transformer connected to a transimpedance amplifier, a non-inverting input terminal of the transimpedance amplifier grounded, an inverting input terminal connected to the secondary coil of the current transformer and a feedback resistor network respectively, and an output terminal generating a voltage signal proportional to the input current. The feedback resistor network includes: resistor R10, resistor R11, and dual-channel relay U4; by controlling the switch of the dual-channel relay, resistor R10 or resistor R11 can be connected between the inverting input terminal and the output terminal of the transimpedance amplifier; the resistance value of resistor R10 is less than the resistance value of resistor R11. The coil of the dual-channel relay U4 is connected to the source of the switching transistor Q2, and the gate of the switching transistor Q2 is connected to the control signal input terminal H7 through the resistor R14; the gate of the switching transistor Q2 is grounded after being connected to the pull-down resistor R15, and the drain of the switching transistor Q2 is grounded; a freewheeling diode D4 is connected in parallel across the coil of the dual-channel relay U4.

2. A non-intrusive household appliance identification and monitoring method as claimed in claim 1, characterized in that, The current acquisition device also includes a step-down circuit, which includes a power input terminal P4. The live wire of the power input terminal P4 is connected in series with a resistor R16, an optocoupler U7, and an output interface H1. The input terminal of the optocoupler U7 is connected in anti-parallel with a diode D5. The output terminal of the optocoupler U7 and the output interface H1 are connected to 5V through a pull-up resistor R17 to achieve a weak pull-up of the line, so that the output voltage is lower than 5V. The degree of conduction of the phototransistor inside the optocoupler U7 determines the magnitude of the output voltage drop.

3. The non-invasive household appliance identification and monitoring method as described in claim 1, characterized in that, The transient auxiliary features within each time window include the maximum instantaneous rate of change of current and the absolute amplitude of the transient peak current; The maximum instantaneous rate of change of current is specifically the maximum change of current per unit time at the location of the maximum adjacent peak point; the absolute amplitude of the transient peak current is specifically the maximum absolute value of the current within the time window.

4. A non-intrusive household appliance identification and monitoring method as claimed in claim 1, characterized in that, The steady-state auxiliary characteristics within each time window include the effective value of the current, active power, reactive power, multiple harmonics, and total harmonic current distortion rate; the total harmonic current distortion rate is specifically the ratio of the sum of the effective values ​​of all higher harmonic currents within the time window to the effective value of the fundamental current.

5. A non-intrusive household appliance identification and monitoring method as claimed in claim 1, characterized in that, The appliance recognition model comprises, in sequence, an input layer, a preprocessing layer, a feature extraction layer, a feature fusion layer, a self-attention module, a Transformer encoding module, a classifier, and an output layer. The self-attention module introduces residual connections. The feature fusion layer concatenates and compresses the multi-scale features extracted by the feature extraction layer into a unified feature representation, which serves as the input to the self-attention module. The self-attention module calculates the dependencies between features across different dimensions. The residual connections then add the output of the self-attention module to the original input feature representation to preserve the original feature information. The final output serves as the input to the Transformer encoding module, which captures the complex relationships and contextual information between features and weights them according to their correlation, thereby learning a higher level of correlation between different appliance behavior patterns.

6. A non-intrusive household appliance identification and monitoring method as claimed in claim 5, characterized in that, The Transformer encoding module includes a time-frequency aware dynamic position encoder and a Transformer encoder; the time-frequency aware dynamic position encoder generates a fixed or time-based code, adds the code to the feature embedding of the input sequence, and then sends the result to the Transformer encoder.

7. The non-invasive household appliance identification and monitoring method as described in claim 6, characterized in that, The input to the time-frequency sensing dynamic position encoder includes: phase information of the fundamental frequency and important harmonic components within each time window. The physical feature vector consists of the maximum adjacent difference peak value, transient auxiliary features, and steady-state auxiliary features. and the gain switching status of the current acquisition circuit. ; The input features are deeply fused through a gating network to obtain fused encoding ; The resulting time-frequency aware dynamic position encoding Specifically: in, Indicates a time step. Indicates the basic timing position code. The weights for the basic temporal position coding, To enhance the embedding of physical features, it is obtained through the physical prior information in the current signal.

8. A non-intrusive household appliance identification and monitoring system, characterized by, include: The current acquisition device is installed on the power cord of the main power strip to acquire the output current signal of the main power strip in real time and convert the acquired current signal into a discrete digital signal. The signal processing module is used to divide the discretized digital signal into multiple time windows through a local window of a predetermined length; The feature extraction module is used to perform fast Fourier transform on the discrete signal within each time window to extract its frequency domain harmonic features. The maximum adjacent difference peak value is obtained by calculating the difference between adjacent sampling points in the discrete signal sequence within the time window; at the same time, transient auxiliary features and steady-state auxiliary features within each time window are extracted. The state recognition module is used to input the frequency domain harmonic features, maximum adjacent difference peak values, transient auxiliary features and steady-state auxiliary features extracted within all time windows into the trained appliance recognition model to obtain the working status of each household appliance on the main power strip. The current acquisition device includes a current detection circuit, which includes a current transformer, a power bus passing through the loop of the primary coil of the current transformer, a secondary coil of the current transformer connected to a transimpedance amplifier, a non-inverting input terminal of the transimpedance amplifier grounded, an inverting input terminal connected to the secondary coil of the current transformer and a feedback resistor network respectively, and an output terminal generating a voltage signal proportional to the input current. The feedback resistor network includes: resistor R10, resistor R11, and dual-channel relay U4; by controlling the switch of the dual-channel relay, resistor R10 or resistor R11 can be connected between the inverting input terminal and the output terminal of the transimpedance amplifier; the resistance value of resistor R10 is less than the resistance value of resistor R11. The coil of the dual-channel relay U4 is connected to the source of the switching transistor Q2, and the gate of the switching transistor Q2 is connected to the control signal input terminal H7 through the resistor R14; the gate of the switching transistor Q2 is grounded after being connected to the pull-down resistor R15, and the drain of the switching transistor Q2 is grounded; a freewheeling diode D4 is connected in parallel across the coil of the dual-channel relay U4.

Images