Non-intrusive household electrical appliance identification and monitoring method and system

Abstract

The invention belongs to the technical field of smart home, and particularly discloses a non-intrusive household electrical appliance identification and monitoring method and system, and the method comprises the steps: collecting an output current signal of a total patch board in real time through a current collection device disposed on a power line of the total patch board, and converting the output current signal into a discrete digital signal; dividing the discretized digital signal into a plurality of time windows; calculating a maximum adjacent difference peak value for the discrete signals in each time window; extracting transient auxiliary features and steady auxiliary features in each time window; and inputting the frequency domain harmonic features, the maximum adjacent difference peak value, the transient auxiliary features and the steady-state auxiliary features extracted in all the time windows into a trained electric appliance identification model to obtain the working state of each household electrical appliance on the total patch board. The method can focus on the waveform characteristics at the critical moment, obtains the fingerprints of transient operation of the electric appliances, and can accurately distinguish the electric appliances with similar steady-state power but different starting characteristics.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of smart home, and particularly relates to a non-intrusive household appliance identification and monitoring method and system. BACKGROUND

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

[0003] In the field of smart home and energy management that is developing today, non-intrusive household appliance identification and monitoring technology is becoming an important tool for energy management, device automation and power safety with its unique flexibility and efficiency. However, most of the traditional monitoring methods are based on the design of intrusive sensors, which need to install detection devices one by one beside each device, and are mainly suitable for power monitoring of fixed and single devices, which greatly limits the application potential of the technology in dynamic and complex environments.

[0004] Non-intrusive household appliance identification technology can install a monitoring device at the bus end, and combine intelligent algorithms to decompose and identify the running conditions of multiple devices under the bus from bus electrical characteristic parameters. For example, the prior art discloses a method of collecting bus current data by using a current transformer, and using a long short-term memory (LSTM) network model to identify the switching events and types of electric appliances. However, the simple current collection scheme is difficult to adapt to the wide dynamic range of household power consumption from weak standby power consumption to high-power appliance starting, resulting in insufficient or distorted signal collection accuracy in high or low current regions; although there is a technology of setting the gain of an operational amplifier, the current amplitude dynamic feedback in the household environment is wide and fluctuates sharply, and the amplifier does not work in a single multiple amplification stage, but needs to switch between multiple states. In addition, traditional time sequence models similar to LSTM are difficult to effectively capture long-range dependencies and multi-scale local features in the superimposed signal of multiple appliances, and it is difficult to distinguish the state of individual devices under the condition of multiple devices running or similar load characteristics, which limits the performance upper limit of the identification system, resulting in deficiencies in feature extraction capability and application specificity of the model. SUMMARY

[0005] To solve the above problems, the application provides a non-invasive household appliance identification and monitoring method and system, which is based on a current collection scheme of an intelligent dynamic switching trans-impedance amplifier of a current transformer and a fast relay, solves the precision and stability problems of wide dynamic range current collection, and ensures lossless sampling in the full dynamic range from weak standby to high-power start; a composite feature extraction scheme considering steady-state frequency domain harmonics and transient waveform characteristics is constructed by performing fast Fourier transform and maximum neighbor difference peak detection on original electric parameter data; and a recognition model based on a Transformer intelligently learns and fuses electric appliance features of different time scales, and overcomes the recognition bottleneck of a traditional time sequence model in a multi-appliance concurrent scene.

[0006] In some embodiments, the following technical solutions are adopted: A non-invasive household appliance identification and monitoring method, comprising: Real-time collection of an output current signal of a total plug-in board by a current collection device installed on a power line of the total plug-in board, and conversion of the collected current signal into a discretized digital signal; Division of the discretized digital signal into a plurality of time windows through a local window of a predetermined length; Fast Fourier transform of the discretized signal in each time window to extract frequency domain harmonic features thereof; calculation of the difference between adjacent sampling points in the discretized signal sequence in the time window to obtain a maximum neighbor difference peak; and extraction of transient auxiliary features and steady-state auxiliary features in each time window; Input of the frequency domain harmonic features, the maximum neighbor difference peak, the transient auxiliary features, and the steady-state auxiliary features extracted in all time windows into a trained appliance identification model to obtain the working states of various household appliances on the total plug-in board.

[0007] As a further scheme, the current collection device comprises a current detection circuit, and the current detection circuit comprises: a current transformer, a circular ring through which a power bus passes a primary coil of the current transformer, a secondary coil of the current transformer connected to a trans-impedance amplifier, a non-inverting input terminal of the trans-impedance amplifier grounded, an inverting input terminal of the trans-impedance amplifier 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 an input current; The feedback resistor network comprises: a resistor R10, a resistor R11, and a two-way relay U4; the resistor R10 or the resistor R11 is connected between the reverse input terminal and the output terminal of the trans-impedance amplifier by controlling the switch of the two-way relay; and the resistance value of the resistor R10 is smaller than that of the resistor R11.

[0008] As a further scheme, the coil of the two-way relay U4 is connected to the source of the switching tube Q2, the gate of the switching tube Q2 is connected to the control signal input end H7 through the resistance R14; the gate of the switching tube Q2 is connected to the ground through the pull-down resistance R15, and the drain of the switching tube Q2 is grounded; the coil of the two-way relay U4 is connected in parallel with the continuous diode D4.

[0009] As a further scheme, the current collection device further comprises a voltage reduction circuit, the voltage reduction circuit comprises a power input terminal P4, the live wire input end of the power input terminal P4 is connected in series with the resistance R16, the optocoupler U7 and the output interface H1 in sequence; the input end of the optocoupler U7 is connected in anti-parallel with the diode D5, and the output end of the optocoupler U7 and the output interface H1 are connected to 5V through the pull-up resistance R17 to realize weak pull-up of the line, so that the output voltage is lower than 5V; the degree of conduction of the photosensitive transistor inside the optocoupler U7 determines the drop amplitude of the output voltage.

[0010] As a further scheme, the transient auxiliary features in each time window include the maximum instantaneous current change rate and the absolute amplitude of the transient peak current. The maximum instantaneous current change rate is specifically the maximum current change amount in unit time at the position of the maximum adjacent difference peak value; and the absolute amplitude of the transient peak current is specifically the maximum value of the absolute value of the current in the time window.

[0011] As a further scheme, the steady-state auxiliary features in each time window include the current effective value, the active power, the reactive power, the multiple harmonic current and the total harmonic current distortion rate; and the total harmonic current distortion rate is specifically the ratio of the sum of the effective values of all high-order harmonic currents to the effective value of the fundamental wave current in the time window.

[0012] As a further scheme, the 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 in sequence; wherein, the self-attention module introduces a residual connection, the feature fusion layer splices and compresses the multi-scale features extracted by the feature extraction layer into a unified feature representation as the input of the self-attention module, the self-attention module calculates the dependency between features in different dimensions, and the residual connection adds the output of the self-attention module to the feature representation of the original input to retain the original feature information; the final output is the input of the Transformer encoding module, the Transformer encoding module can capture the complex relationships and context information between features and weight them according to their correlation, thereby learning higher-level relevance between different appliance behavior patterns.

[0013] As a further solution, the Transformer encoding module comprises a time-frequency-aware dynamic position encoder and a Transformer encoder; the time-frequency-aware dynamic position encoder generates a fixed or time-based encoding, adds the encoding to the feature embedding of the input sequence, and sends the result to the Transformer encoder.

[0014] As a further solution, the input of the time-frequency-aware dynamic position encoder comprises phase information of the fundamental frequency and important harmonic components in each time window The physical feature vector is composed of the maximum neighbor difference peak value, the transient auxiliary feature, and the steady-state auxiliary feature , and the gain switching state of the current acquisition circuit ; The input features are deeply fused through a gating network to obtain a fused encoding ; The obtained time-frequency-aware dynamic position encoding Specifically, ; Wherein, t represents a time step, represents a basic time sequence position encoding, is the weight of the basic time sequence position encoding, is an enhanced physical feature embedding, which is obtained by using physical prior information in the current signal.

[0015] In some other embodiments, the following technical solutions are adopted: A non-invasive household appliance identification and monitoring system comprises: A current acquisition device is installed on the power line of the total plug-in board, and is used to acquire the output current signal of the total plug-in board in real time, and convert the acquired current signal into a discretized digital signal; A signal processing module is used to divide the discretized digital signal into a plurality of time windows through a local window of a predetermined length; A feature extraction module is used to perform fast Fourier transform on the discrete signal in each time window to extract the frequency domain harmonic feature thereof; the maximum neighbor difference peak value is obtained by calculating the difference between adjacent sampling points in the discrete signal sequence in the time window; and the transient auxiliary feature and the steady-state auxiliary feature in each time window are extracted at the same time; A state identification module is used to input the frequency domain harmonic feature, the maximum neighbor difference peak value, the transient auxiliary feature, and the steady-state auxiliary feature extracted in all time windows into a trained appliance identification model to obtain the working state of each household appliance on the total plug-in board.

[0016] Compared with the prior art, the beneficial effects of the present application are: (1) The present application divides the original signal into multiple time windows, extracts frequency domain harmonic features, maximum neighbor difference peak value, transient auxiliary features and steady state auxiliary features in each time window, can identify the accurate time of the appliance switching event, when an appliance is turned on or off, the total load current will have a significant transient change. The maximum neighbor difference peak value detection can capture the sharp points on the current waveform, accurately mark the time point of the appliance state switching, provide a time reference for the subsequent appliance event segmentation and identification. At the same time, it can better extract the transient features. In addition to the steady-state harmonic features, the appliance will also produce a unique transient current waveform at the moment of starting or closing. By identifying these peak change points, the system can focus on the waveform features at these key moments to obtain the fingerprint of the transient operation of the appliance, which is crucial for distinguishing different types of appliances, especially those with similar steady-state power but different starting characteristics.

[0017] (2) The design of the current acquisition device intelligent relay dynamic switching of the trans-impedance amplifier feedback resistor in the present application, by pre-setting multiple high-precision feedback resistors, and by a low-delay microprocessor, the relay is quickly and responsively switched according to the real-time current amplitude, so that the gain of the trans-impedance amplifier is adjusted to the best state in a very short time. Thus, it can ensure that from the standby state of the micro-current of several milliamperes to the peak current of several tens of amperes of the starting of high-power appliances, accurate sampling can be carried out with high linearity and high signal-to-noise ratio. Effectively avoid the signal distortion and sampling blind area introduced when the gain is different; not only improve the adaptability and measurement accuracy of the circuit, but also support diversified power application scenarios, and provide an efficient data basis for energy management systems.

[0018] (3) The appliance recognition model based on Transformer in the present application, 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 of different time scales, overcoming the recognition bottleneck of traditional time series models in multi-appliance concurrent scenarios. This not only greatly improves the recognition accuracy of the system when the load characteristics are similar or the power consumption behavior fluctuates frequently, but also enhances its dynamic adaptability and compatibility when facing unknown appliances and complex environments, enabling it to be widely and flexibly applied in homes, businesses and even industrial environments.

[0019] Other features and advantages of the additional aspects of the present application will be partially given in the following description, partially will become apparent from the following description, or will be understood by the practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 Flow chart of the non-intrusive household appliance identification and monitoring method in the embodiments of the present application; Figure 2A structure diagram of a current detection circuit in the current acquisition device in the embodiment of the present application is shown in the figure; Figure 3 A structure diagram of a voltage reduction circuit in the current acquisition device in the embodiment of the present application is shown in the figure; Figure 4 A structure diagram of an appliance recognition model in the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0021] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise indicated, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs.

[0022] It should be noted that the terms used herein are only for the purpose of describing specific embodiments, and are not intended to limit the exemplary embodiments according to the present application. As used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and furthermore, it should be understood that when the terms "comprise" and / or "include" are used in the specification, there is a feature, step, operation, device, component, and / or combination thereof.

[0023] Embodiment one In one or more embodiments, a non-invasive household appliance identification and monitoring method is disclosed, which identifies all electrical equipment in the working state through the aggregated power signal obtained at the total plug-in board of the power entry. Given the instantaneous current sequence i t , the goal is to predict the active appliance set A={a1,a2,…,a k} and its working state within a certain time period. Wherein, the certain time period refers to the time range covered by the current signal data currently used for analysis, and the embodiment model performs inference and prediction within each local time window; active appliances refer to the set of all electrical equipment that the system judges to be in the on or non-standby running state within a certain time period; the working state refers to the specific running mode of each "active appliance" within the current "certain time period", for example, the air conditioner has cooling and heating modes.

[0024] Since the supply bus voltage amplitude is relatively stable, its phase feature has key significance and can be used as the phase reference of the current signal. The embodiment adopts high sampling rate data acquisition to capture transient characteristics, and the original data is converted into a low-dimensional feature vector after preprocessing and feature extraction to represent the aggregated load characteristics. The constructed appliance recognition model can output an N-dimensional binary vector after training, where N is the total number of identifiable appliance types, and the element indicates that the jth type of electrical appliance is in an operating state (1) or a non-operating state (0).

[0025] As a specific embodiment, in combination with Figure 1 The non-intrusive household appliance identification and monitoring method of the embodiment specifically includes the following processes: S101: Real-time acquisition of output current signals of the total plug-in board through the current acquisition device installed on the power line of the total plug-in board, and conversion of the acquired current signals into discrete digital signals.

[0026] In the embodiment, the current acquisition device is responsible for high-precision and high-speed acquisition of the aggregated current signal on the main power supply line; the current acquisition device is installed on the power line of the total plug-in board, and the current change of the socket is monitored in real time through the current detection circuit to ensure accurate monitoring of the current flowing through the line; the acquired signal is transmitted to the data acquisition card, and a high-precision analog-to-digital converter converts it into a digital signal, which is transmitted to the RK3588 processor chip for processing and analysis through the SPI data transmission protocol. The structure design enables the current acquisition device to be seamlessly integrated into the existing home or industrial electrical system without the need for large-scale modification of the original circuit layout, realizing real-time monitoring of the current information of all electrical appliances in the entire circuit, providing a solid data foundation for subsequent energy consumption analysis and load identification, and supporting device management and optimization in complex environments.

[0027] The current detection circuit with super-wide dynamic range in the embodiment uses a current transformer with a transformation ratio of 20 / 1 to reduce the current parameter, which can effectively convert large current into small current signal under high current condition, ensuring that the measurement and control system obtains accurate and stable data in the analysis and control process. The current transformer is magnetically coupled through the primary coil and the secondary coil on the power line. When the two are connected and properly configured, they can effectively transfer large current to the secondary coil, output a small current signal proportional to the original current, and have the characteristic of being naturally isolated from the power supply. This technology has strong adaptability to current changes. Even if the current fluctuates greatly when the household appliance is turned on, the current transformer can still provide accurate measurement stably, ensuring the functions of the circuit system.

[0028] The current detection circuit in the embodiment uses a current detection scheme of current transformer (CT) and trans-impedance amplifier (TIA). This design safely and isolatedly obtains the proportional signal of the bus load current through the current transformer, which is then converted into a voltage signal by a trans-impedance amplifier. To achieve large dynamic range and high-precision current sampling, the circuit uses a relay to dynamically switch the feedback resistance of the trans-impedance amplifier, thereby adjusting the gain to adapt to different amplitude of current input. Finally, a voltage signal that is accurately corresponding to the original load current waveform is output after conditioning, for subsequent non-intrusive load monitoring and analysis.

[0029] In combination Figure 2 The current detection circuit specifically comprises: a current transformer, a circular ring through which the power bus passes a primary coil of the current transformer, a secondary coil of the current transformer connected to a trans-impedance amplifier, a non-inverting input terminal of the trans-impedance amplifier grounded, an inverting input terminal of the trans-impedance amplifier 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 an input current; Specifically, Figure 2 In the circuit, P6 is an input terminal of the current transformer (CT). Pin 1 and pin 2 of P6 are connected to the secondary coil of the current transformer respectively, for receiving an alternating current signal induced from the primary coil. U6.1 (LM358ADR2G) is an operational amplifier configured as a trans-impedance amplifier (TIA) in the circuit. The non-inverting input terminal (pin 3) of the operational amplifier U6.1 is connected to ground (GND), providing a stable reference potential; the inverting input terminal (pin 2) of the operational amplifier U6.1 is connected to pin 2 of P6 and a feedback resistor network controlled by the relay U4. The output terminal (pin 1) of the operational amplifier U6.1 generates a voltage signal proportional to the input current. LM358ADR2G is a low-power, dual-channel operational amplifier suitable for battery-powered applications, with a wide range of power supply voltages, which can meet the low-power and precision requirements of the device.

[0030] The resistors R10 and R11 constitute a feedback resistor network of the trans-impedance amplifier. These resistors are switched by the relay U4 to change the gain of the trans-impedance amplifier. The resistance value of the feedback resistor determines the conversion ratio of current to voltage, i.e., the conversion factor.

[0031] Specifically, the feedback resistor network comprises: the resistor R10, the resistor R11, and the dual-channel relay U4. By controlling the switch of the dual-channel relay, the resistor R10 or the resistor R11 is connected between the reverse input terminal and the output terminal of the trans-impedance amplifier. The resistance value of the resistor R10 is smaller than that of the resistor R11. The coil of the dual-channel relay U4 is connected to the source of the switching tube Q2. The gate of the switching tube Q2 is connected to the control signal input terminal H7 through the resistor R14. The gate of the switching tube Q2 is connected to the ground through the pull-down resistor R15. The drain of the switching tube Q2 is connected to the ground. The coil of the dual-channel relay U4 is connected in parallel with the continuous diode D4 across the coil.

[0032] Dual-relay U4 is a dual-relay, which is used to dynamically switch the feedback resistance of the trans-impedance amplifier. The relay coil is controlled by switch Q2 and resistor R14, whose driving signal comes from the control signal access terminal H7. When the relay is powered on, its normally closed contact is open and the normally open contact is closed; when the relay is powered off, its normally closed contact is closed and the normally open contact is open. When the control signal access terminal H7 controls the relay coil to be powered on, the normally open contact of the dual-relay U4 is closed. At this time, resistor R10 is connected to the inverting input terminal of U6.1 through the pin 4-5 of the dual-relay U4, forming a small feedback resistance path, providing a lower gain, which is suitable for measuring large current. When the control signal access terminal H7 controls the relay coil to be powered off, the normally closed contact of the dual-relay U4 is closed. At this time, resistor R11 is connected to the inverting input terminal of U6.1 through the pin 3-6 of the dual-relay U4, forming a larger feedback resistance path, providing a higher gain, which is suitable for measuring small current, thereby expanding the dynamic range of current measurement.

[0033] Switch Q2 is an N-channel MOSFET, and resistors R14 and R15 are used to drive dual-relay U4. The control signal access terminal H7 provides a control signal, which is connected to the gate of switch Q2 through resistor R14. Resistor R15 is a gate pull-down resistor, which ensures that switch Q2 is reliably turned off when there is no control signal. When the control signal access terminal H7 outputs a high level, switch Q2 is turned on to provide a path for the relay coil, and the relay is attracted. When H7 outputs a low level, switch Q2 is turned off, and the relay coil is powered off. Diode D4 is a freewheeling diode connected in parallel with the relay coil, which is used to release the induced electromotive force when the relay coil is powered off, protecting switch Q2 from reverse voltage impact.

[0034] H4 is a signal output interface, pin 1 of H4 is connected to the output terminal of the regulated analog voltage signal, and pin 2 is grounded. The output signal is an analog voltage signal after trans-impedance amplifier conversion, gain adjustment and level bias, whose waveform accurately corresponds to the waveform of the load current, which can be sampled, analyzed and processed by the subsequent analog-to-digital converter (ADC) or microcontroller (MCU) to realize non-intrusive load monitoring.

[0035] The embodiment combines the design of current transformer and intelligent relay dynamic switching of feedback resistance of trans-impedance amplifier, which can solve the problems of insufficient collection accuracy and distortion caused by wide dynamic range and severe fluctuation of current amplitude in household environment. Compared with the precision loss in wide dynamic range of the simple transformer scheme, the embodiment switches the relay by a low-delay microprocessor according to the real-time current amplitude, so as to quickly adjust the gain of the trans-impedance amplifier to the optimal state in a very short time, and has very high precision. This can not only ensure that the standby state of the micro-current of several milliamperes, to the peak current of several tens of amperes of the starting of high-power electrical appliances, can be accurately sampled with high linearity and high signal-to-noise ratio. The intelligent dynamic switching mechanism effectively avoids the signal distortion and sampling blind area introduced when the gain is different. This method not only improves the adaptability and measurement accuracy of the circuit, but also supports diversified power application scenarios, and provides an efficient data basis for energy management systems.

[0036] In the embodiment, the current collection device further includes a power supply system including a voltage stabilizing power supply, a voltage reduction circuit and a phase detection circuit. The voltage stabilizing power supply ensures stable power supply to the entire circuit; the voltage reduction circuit is responsible for converting 220V household alternating current to a voltage range of 20V or less suitable for chip and operational amplifier use; and the phase detection circuit is used to ensure that each sampling period starts at the zero phase point of the power frequency voltage, and to realize accurate collection of voltage phase information.

[0037] The voltage reduction circuit reduces 220V alternating current to an appropriate voltage level while retaining the power frequency characteristics in the alternating current signal. By utilizing the half-wave characteristic, the phase accuracy of the output alternating current signal is ensured, and the overall stability and reliability of the circuit are enhanced, preventing damage to the device caused by voltage overload. In addition, optical coupling is used to achieve electrical isolation to ensure safe operation of the microprocessor; in combination Figure 3 The voltage reduction circuit specifically includes a power input terminal P4, the live wire access end of the power input terminal P4 is connected in series with a resistor R16, an optical coupler U7 and an output interface H1; the input end of the optical coupler U7 is anti-parallelly connected with a diode D5, and the output end of the optical coupler U7 and the output interface H1 are connected to 5V through a pull-up resistor to achieve weak pull-up of the line. When the photosensitive transistor inside the optical coupler output side is turned on, it will shunt from the pull-up resistor. The current passing through the pull-up resistor will produce a voltage drop, so that the output end voltage is lower than 5V. The degree of conduction of the photosensitive transistor inside the optical coupler U7 (i.e. the size of the current flowing from the collector to the emitter) determines the amplitude of the output voltage drop. Therefore, the pull-up resistor converts the photosensitive signal into a measurable voltage change.

[0038] Specifically, the power input terminal P4 is an alternating current power input terminal, and the pin 1 and the pin 2 of the P4 are connected to an alternating current zero line (N) and a live wire (L) respectively, which is a high-voltage dangerous area. The resistor R16 is a current limiting resistor, and the main role of the resistor R16 is to limit the current flowing through the inside of the optocoupler U7 and the protection diode D5; for a 220V alternating current input, the peak voltage is about 220V ≈311V. When the alternating current voltage is at the peak, a large current will burn the optocoupler and the protection diode D5, and the resistor R16 can ensure that the current is within the safe working range of these devices.

[0039] The protection diode D5 is a fast switching diode, and is reversely connected in parallel with the input (between the pin 1 and the pin 2) of the optocoupler U7 in the circuit. In the positive half cycle of the alternating current, the voltage of the pin 1 of the power input terminal P4 is higher than that of the pin 2, at this time, the optocoupler diode of the optocoupler U7 is turned on, and the protection diode D5 is in a reverse blocking state. In the negative half cycle of the alternating current, the voltage of the pin 1 of the power input terminal P4 is lower than that of the pin 2, at this time, the D5 is forward biased, and the reverse voltage between the two ends of the internal diode of the optocoupler U7 is clamped at the forward conduction voltage (about 0.7V) of the D5. The optocoupler U7 realizes the electrical isolation between the high-voltage alternating current side and the low-voltage direct current digital side (usually capable of bearing thousands of volts of isolation voltage), and is composed of an infrared light emitting diode (LED) and a phototransistor, which are packaged together but electrically isolated. Effectively protects the ADC chip in the rear stage from high-voltage impact, and eliminates the problem of ground loop. The core role of the protection diode D5 is to protect the LED diode of the optocoupler U7; the internal LED of the optocoupler U7 usually has a very low reverse breakdown voltage (such as about 5V). In the negative half cycle of the alternating current, if there is no protection diode D5, the LED will be damaged by a reverse voltage far exceeding its withstand capability; the D5 provides a path for the current in the negative half cycle and protects the LED.

[0040] The output interface H1 provides an interface for the processor chip in the rear stage, the pin 1 outputs +5V power supply (which can supply power to the MCU or other external devices, or only serves as a reference), and the pin 2 is used as a signal output, which is connected to the collector of the optocoupler U7 and is also an input pin of the processor chip in the rear stage. The signal is a square wave (or an approximate square wave): in the positive half cycle of the alternating current (and when the voltage is higher than the optocoupler opening threshold), the output is low (close to 0V); in the negative half cycle of the alternating current and when the voltage is low near zero point in the positive half cycle, the output is high (+5V).

[0041] The embodiment realizes the safe operation of the microprocessor through the electrical isolation of the optocoupler, which not only avoids the loss and error that may occur in the traditional voltage conversion, but also improves the deployment efficiency of the circuit system in the smart home, ensures that the system can quickly respond and accurately process each electrical period, and provides a stable basis for subsequent current collection and data analysis.

[0042] The aggregated current waveform is collected After that, it is digitized with a high sampling rate, which is set to 3.2 kHz to ensure that the 30th harmonic (corresponding to the 50 Hz fundamental frequency) can be accurately captured. With this high sampling rate, not only the harmonic components can be accurately captured, but also the transient characteristics when the electrical appliance is started or turned off can be effectively captured.

[0043] The digitized discrete-time current signal is represented as , where n is the sample point index. The Fourier transform is performed on this discrete signal 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: ; , where X[k] represents the complex value of the spectral component at frequency index k, and N is the number of sample points used for FFT calculation. After obtaining the spectrum X[k], the 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 represent the unique harmonic characteristic spectrum of different types of electrical appliances (such as devices using switching power supply, devices with motors, or purely resistive loads) in the working state.

[0044] S102: Divide the digitized digital signal into multiple time windows through a local window of a predetermined length.

[0045] In this embodiment, in order to ensure the robustness of the algorithm, a local window mechanism is introduced, which is specifically implemented as follows: when extracting features from the digitized discrete-time current signal , the system does not process the entire long time sequence at one time, but divides the data stream into a series of fixed-length overlapping time windows.

[0046] Specifically, a fixed window length L is set, which corresponds to several power frequency periods and a sliding step S (S The current sampling sequence is divided into a series of sub-sequences, i.e. time windows The kth time window contains data from sample point k S to k S+L

[0047] ; To reduce spectral leakage, each sub-sequence within a window is multiplied by a window function w[n] (Hamming window) before Fourier transform, and the processed window data is .

[0048] For each independent time window processed by the window function and the data contained therein , the system independently performs a fast Fourier transform (FFT), extracts its frequency domain harmonic features, and applies a maximum neighbor difference peak detection to obtain its time domain transient and steady state features.

[0049] This embodiment, through the design of local window mechanism, enables the system to effectively deal with local transient changes and non-steady state signals, such as current spikes when the appliance is started or turned off, local harmonic distortion caused by load type switching or short-term power grid disturbance. By continuously sliding these overlapping windows on the time axis, the system can achieve continuous monitoring and dynamic capture of signal features, and real-time update of the running state of the appliance. In addition, the overlapping design of the window (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 boundary, and facilitating real-time response of the system in dynamic power load changes, improving the flexibility and practicality of data processing, thereby ensuring that feature detection remains efficient and reliable in complex power environments.

[0050] S103: Perform a fast Fourier transform on the discrete signal in each time window to extract its frequency domain harmonic features; obtain the maximum neighbor difference peak value by calculating the difference between adjacent sampling points in the discrete signal sequence within the time window; simultaneously extract the transient auxiliary features and steady state auxiliary features within each time window.

[0051] In this embodiment, for all electrical equipment to be identified, the amplitude information of the first 7 harmonics (i.e. fundamental to 7th harmonic) is uniformly extracted. These harmonic components (including fundamental amplitude, second harmonic amplitude,..., seventh harmonic amplitude) constitute the main features of the device in the frequency domain, which are sufficient to represent the unique harmonic feature spectrum of different types of electrical equipment (e.g. devices using switching power supply, devices with motors or purely resistive loads) in the working state. Considering that 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, selecting the first 7 harmonics can maintain the discrimination while considering the computational efficiency.

[0052] At the same time, through the maximum adjacent difference peak detection, the accurate time and transient characteristics of the electric appliance switching event can be accurately captured. When an electric appliance is turned on or off, the total load current will change significantly. The maximum adjacent difference peak detection can capture the sharp points on the current waveform, and the characteristics contain the accurate time point of the electric appliance state switching, providing a reference for the subsequent neural network model to segment and identify the electric appliance event.

[0053] For the current sampling data in each local window , The maximum adjacent difference peak detection calculates the difference between adjacent sampling points in the sequence and identifies the position point with the maximum difference.

[0054] For a discrete current sequence , The difference sequence of adjacent points is defined as: ; Wherein, n is the sampling point index. The goal of the algorithm is to find the index that makes reach the maximum value: ; Wherein, is the maximum adjacent difference peak point.

[0055] In addition, in order to further enhance the recognition ability of multiple electric appliances, similar load characteristics and weak events in complex home scenarios, the embodiment also innovatively introduces a cooperative extraction mechanism of multi-dimensional transient waveform and steady-state auxiliary features.

[0056] Specifically, the transient auxiliary feature can capture and quantify the non-stationary, high-frequency or low-frequency transient waveform details at the moment of electric appliance start or stop at high resolution, effectively distinguish the start-up ripple of various electric appliance loads and the smooth changes of resistive loads, and overcome the limitations of FFT in analyzing non-stationary signals. These features can be quantified from the rapid change part of the current signal when the electric appliance changes state (such as turning on, turning off, mode switching).

[0057] The transient auxiliary feature specifically includes: (1) Maximum instantaneous current change rate .

[0058] The maximum instantaneous current change rate reflects the impact strength of the electric appliance on the grid current in a very short time. High-power inductive loads (such as motor start) or switching power supplies (such as computers, chargers) will cause rapid rise of current at the moment of starting, and the characteristic difference is significant. Different electric appliances have their specific instantaneous current change rate range and mode.

[0059] The maximum instantaneous current change rate is specifically the maximum change of current in a unit time at the detected maximum adjacent difference peak point wherein the maximum change of current in a unit time is calculated: ; wherein, represents the current sample value at the maximum adjacent difference peak point , represents the current sample value at the previous sampling period of the maximum adjacent difference peak point, and T is the sampling period.

[0060] (2) Absolute amplitude of transient peak current .

[0061] The absolute amplitude of transient peak current is used to capture the inrush current when the electrical appliance is started. The specific calculation method is that the absolute maximum value of the current in the local window where the maximum adjacent difference peak point is located: ; wherein, represents the current value in the local window.

[0062] The steady-state auxiliary features reflect the electrical characteristics of the electrical appliance in the stable running state, and are crucial for identifying the type and mode of the continuously running electrical appliance. These features are usually extracted in a period of stable running of an electrical appliance; the steady-state auxiliary features specifically include the following features: (1) Current effective value .

[0063] The current effective value is the root mean square value of the current signal calculated in one or more power frequency periods: ; wherein, represents the number of sampling points collected in a complete power frequency period. If multiple power frequency periods are used, the corresponding more smoothed current effective value results can be obtained.

[0064] (2) Active power .

[0065] The active power is specifically the active power of the bus in one or more power frequency periods: ; wherein, , , are the number of power frequency period sampling points, voltage sample value and current sample value, respectively.

[0066] (3) Reactive power .

[0067] The reactive power is calculated by the 90 degree phase difference component of the current and voltage, the reactive component of the bus power. The commonly used method is to use the imaginary part of the Fourier transform or by calculating the apparent power and active power.

[0068] ; wherein, , , , , respectively represent the amplitude and phase angle of the hth harmonic voltage and current.

[0069] (4) The multiple harmonic characteristics are calculated based on the aforementioned FFT algorithm.

[0070] (5) The total harmonic current distortion rate .

[0071] The total harmonic current distortion rate is the ratio of the effective value of the high-order harmonic current to the effective value of the fundamental current in the local time window: ; wherein, is the effective value of the fundamental current, is the effective value of the hth harmonic current.

[0072] The cooperative extraction mechanism of the multi-dimensional transient waveform and steady-state auxiliary characteristics of the embodiment ensures that the system not only can identify the occurrence of the event, but also can accurately depict the unique "electrical fingerprint" of the electrical appliance at the moment of switching and during steady-state operation from multiple dimensions, greatly improving the recognition accuracy when the load characteristics are highly similar or the power consumption behavior frequently fluctuates, especially when distinguishing complex loads such as motor-type electrical appliances and switching power supply-type electrical appliances. The fast FFT algorithm efficiently reduces the number of multiplication operations, especially when the number of sampling points increases, the amount of calculation saved also increases significantly, so that the system can quickly analyze complex frequency domain signals, ensuring the integrity and accuracy of the signals. At the same time, the maximum adjacent difference peak detection method quickly identifies the position of the maximum difference value by calculating the difference between adjacent points in the sequence, providing accurate feature points for electrical signal monitoring and identification.

[0073] S104: input the frequency domain harmonic characteristics, maximum adjacent difference peak, transient auxiliary characteristics and steady-state auxiliary characteristics extracted in all time windows into the trained electrical appliance recognition model, to obtain the working state of each household electrical appliance on the total patch panel.

[0074] The embodiment extracts the frequency domain harmonic features, the maximum neighbor difference peak value, the transient auxiliary features and the steady-state auxiliary features in each time window to form a multi-dimensional feature vector, and the multi-dimensional feature vectors of all windows are used as input features of the appliance recognition model. The model learns the complex nonlinear mapping relationship between the combined features and the working state of the appliance, so as to realize accurate recognition and monitoring of the appliance.

[0075] In view of the fact that the existing traditional time sequence model has limited ability in processing complex household electricity scene, multi-appliance concurrency and long-range dependence, and is difficult to effectively capture subtle and key transient features in the current signal, resulting in a fundamental defect of reduced recognition accuracy, the embodiment proposes an innovative, highly adaptive and deeply integrated time sequence-scale information based on the Transformer recognition model. The designed Transformer model is deeply customized and optimized for electrical features, wide dynamic range signals and multi-appliance concurrency detection problems in the non-intrusive load monitoring (NILM) task.

[0076] Specifically, in combination with Figure 4 , the 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 in sequence. The residual connection is introduced in the self-attention module, and the feature fusion layer splices and compresses the multi-scale features extracted by the feature extraction layer into a unified feature representation, which is used as the input of the self-attention module. The self-attention module calculates the dependence between features in different dimensions, and the residual connection adds the output of the self-attention module to the original feature representation of the input to retain the original feature information and stabilize the training process, avoid information loss and gradient problems; the final output is used as the input of the Transformer encoding module, and the Transformer encoding module can capture the complex relationship and context information between the features and weight them according to their correlation, so as to learn the higher-level relevance between different appliance behavior patterns.

[0077] Specifically, the input features are input from the input layer, first preprocessed by a linear preprocessing layer (Linear) to map the feature dimension from the original dimension to 64 (feature_dim, 64), then nonlinearly transformed by BatchNorm1d and ReLU activation function, and finally prevented from overfitting by Dropout layer. The preprocessing step aims to standardize the input data and increase the expression ability of the model.

[0078] The feature extraction layer uses four convolutional layers convld with different dilation rates (dilation rates of 1, 2, 4, and 8, respectively) to extract multi-scale features. The output of each convolutional layer is reduced in dimension by an average pooling operation, and finally the features of different scales are concatenated. This multi-dilation convolution structure (or called as dilated convolution) allows the model to increase the receptive field without increasing the number of parameters and computational complexity, capture the dependence and local patterns at different time scales, and effectively extract both transient and steady-state features.

[0079] The original features are concatenated with the outputs of the four dilated convolutions, and then a fusion layer (including a linear layer, batch normalization, ReLU activation, and Dropout) is used for feature fusion, finally outputting 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 convolution.

[0080] In this embodiment, the combination of the self-attention module, the Transformer encoding module, and the residual connection network is the key to realizing long-range dependence capture and high-level semantic understanding by the model. The 64-dimensional feature vector output by the feature fusion layer is input to the self-attention module. The 64-dimensional feature vector (where B is the batch size, =64 is the feature dimension) is first linearly transformed to generate the query Q, the key K, and the value V, i.e. , , where , , are learnable weight matrices. These parameters are then sent to the multi-head self-attention mechanism.

[0081] To alleviate the gradient influence in the training of deep networks and allow the network to learn the identity mapping, the model innovatively introduces residual connection in the self-attention module. As Figure 4 shown, the input features of the self-attention module (i.e., the output of the fusion layer) are directly added to the output of the multi-head self-attention mechanism at the element level through residual connection. The mathematical expression is: ; The residual connection design of the embodiment ensures that the original input information can be directly transmitted to the 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 layers, thereby significantly improving the stability and convergence speed of the training while maintaining the model's expression ability, which is particularly important for network structures that capture long-range dependencies. The output after the residual connection is further enhanced by layer normalization and a feedforward network, which further enhances the expression ability of the features, and together form a sub-layer of a Transformer encoder.

[0082] The core structure of the Transformer encoding module is composed of a customized adaptive scale-time Transformer encoder and a NILM (Non-Intrusive Load Monitoring) specific enhanced position encoder, which works with a multi-label classification head. The Transformer encoding module uses four layers of Transformer encoders, each containing at least four attention heads, aiming to gradually abstract and refine more discriminative appliance behavior patterns through a stacked encoder structure.

[0083] In the Transformer encoding module, the position encoder generates a fixed or time-based encoding, which is then added to the input sequence's word embeddings (or feature embeddings), and the result is fed into the Transformer encoder. The position encoder acts as a pre-processing for the Transformer encoder.

[0084] The NILM-specific enhanced position encoder in this embodiment is a time-frequency-aware dynamic position encoder. The input of the time-frequency-aware dynamic position encoder includes: The phase information of the fundamental frequency and important harmonic components in each time window The physical feature vector composed of the maximum neighbor difference peak, transient auxiliary features, and steady-state auxiliary features And the gain switching state of the current acquisition circuit (derived from the actual gain of the circuit); these input features are deeply fused through a gating network to obtain ; The time-frequency-aware dynamic position encoding is obtained Specifically: ; Where, t represents the time step, b represents the basic time sequence position encoding, w is the weight of the basic time sequence position encoding, is the basic time sequence position encoding, which is learned through network training and provides the most basic time sequence information for each time step in the sequence. is a learnable weight matrix that linearly transforms the base positional encoding to make it more flexible to fit into the feature space of the Transformer model. And represents a high-dimensional enhanced physical feature embedding, which is a small fully connected network that maps the unique physical prior information in the current signal (such as frequency domain features, event features, and gain states) to a representation consistent with the model's feature dimension.

[0085] When the model receives the input sequence, the gating network dynamically adjusts the weights and form of the position encoding based on the frequency domain and event features within the current time window, allowing it to more strongly emphasize those that represent key transient moments and energy change periods of electrical appliances. This dynamic and physically-aware position encoding allows the Transformer encoding module not only to understand the temporal order in the sequence, but also to deeply understand the physical meaning and occurrence context carried by electrical events in the current signal, greatly improving the model's contextual understanding of electrical events and sensitivity to specific electrical behavior patterns. This design allows the model to adjust and integrate temporal position information and electrical characteristics at each calculation step of the Transformer encoding module based on the real-time physical properties of the current signal. This method not only improves the accuracy of the Transformer's long-range electrical dependence capture, but also significantly enhances the ability to recognize subtle electrical transient features.

[0086] Each Transformer encoder layer follows the basic structure of "multi-head self-attention + residual connection + layer normalization" and "feedforward network + residual connection + layer normalization". This encoder structure effectively captures complex relationships and contextual information between features, allowing the model to learn higher-level associations between different electrical behavior patterns. The Transformer encoder, through its internal self-attention mechanism, allows each input feature in the model to "focus" on all other features in the input sequence and weight them based on their relevance, thereby capturing long-range dependencies. This is particularly advantageous in household appliance recognition tasks where different electrical operating states may influence each other, or certain event features require correlation over a longer time window to be fully understood. Through the stacking of multiple encoder layers, the model can gradually abstract and refine more discriminative feature representations.

[0087] The model's classifier consists of two linear layers, with batch normalization, GELU activation function, and Dropout layer in between for non-linear transformation to enhance the model's generalization ability and prevent overfitting. The final output is the predicted probability of each electrical appliance category, activated by a Sigmoid function; this Sigmoid function maps the output to between 0 and 1.

[0088] The output layer of the model outputs a set, each element of the set corresponds to a known appliance category and indicates the current working state of the appliance. This state can be a simple binary (on / off), or a multi-value (such as gear, running mode), depending on the granularity of the training data and the classification ability of the model.

[0089] In the training of the appliance recognition model, air conditioners, heaters, hair dryers, electric pots, electric fans, table lamps and electric mosquito repellents are used as sample appliances for data sampling of single appliances and different appliance combinations to construct the training data set. The model outputs an 8-dimensional vector, each dimension corresponding to an appliance or specific state, indicating whether the appliance is currently working and the probability of the working state. For example, the model outputs an 8-dimensional vector, and indicates that the probability of air conditioner cooling and electric fan starting is 80%, indicating that the two loads are started on the current bus.

[0090] During the training process, the binary cross-entropy loss function (BCEWithLogitsLoss) is used to calculate the error between the predicted value and the true label to optimize the model parameters. For this type of task, the binary cross-entropy loss can independently evaluate the prediction accuracy of each class, avoiding the assumption of label exclusivity of traditional multi-classification loss functions (such as Softmax cross-entropy).

[0091] The learning rate scheduler ReduceLROnPlateau is used to dynamically adjust the learning rate according to the F1 score of the validation set. If the F1 score of the validation set does not improve for 10 consecutive rounds, the learning rate will be reduced to half of the original. This adaptive learning rate adjustment strategy helps the model to jump out of the local optimum, speeds up the convergence and improves the final performance of the model.

[0092] During the training process, data augmentation is performed on the categories with insufficient samples by adding Gaussian noise and random scaling to generate new samples to balance the data distribution. This data augmentation strategy can effectively alleviate the problem of data imbalance, improve the model's ability to recognize unusual patterns, and enhance the model's generalization and robustness.

[0093] The feature information of the training data is continuously collected, which contains the multi-label distribution of the features in the dynamic process, as the basis for optimizing the performance of the model. The system will transmit the loss and index information in the training process to the background in real time, providing data analysis and chart display for data interaction. The operator will execute the corresponding strategy based on this information to ensure the stability and convergence of the model. The validation set configured during the model training process will monitor the generalization ability of the model to ensure the robustness and accuracy of the model in actual application.

[0094] The recognition model based on the Transformer of the embodiment can intelligently learn and fuse the features of electrical appliances of different time scales, and overcome the recognition bottleneck of traditional time sequence models in the multi-appliance concurrent scene. This not only greatly improves the recognition accuracy of the system when the load features are similar or the power consumption behavior fluctuates frequently, but also enhances the dynamic adaptability and compatibility of the system when facing unknown electrical appliances and complex environments, so that the system can be widely and flexibly applied in home, commercial and even industrial environments. The real-time analysis capability of the built-in data processing module, combined with the continuous optimization of model training, further improves the practicality of energy consumption mode evaluation. This series of forward-looking innovative designs and functional integrations constitute an efficient, stable, high-precision and high-robustness household appliance recognition and monitoring solution, which provides solid technical support for smart home and fine energy management, and has undoubtedly broad application prospects and great popularization value.

[0095] Embodiment two In one or more embodiments, a non-invasive household appliance recognition and monitoring system is disclosed, comprising: A current acquisition device is installed on the power line of the total plugboard to acquire the output current signal of the total plugboard in real time, and convert the acquired current signal into a discretized digital signal; A signal processing module is used to divide the discretized digital signal into multiple time windows through a local window of a predetermined length; A feature extraction module is used to perform fast Fourier transform on the discrete signal in 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 in the time window; and the transient auxiliary features and steady-state auxiliary features in each time window are extracted; A state recognition module is used to input the frequency domain harmonic features, the maximum adjacent difference peak value, the transient auxiliary features and the steady-state auxiliary features extracted in all time windows into the trained appliance recognition model to obtain the working state of each household appliance on the total plugboard.

[0096] The specific implementation of each module is the same as that in Embodiment One, and will not be described in detail.

[0097] Although the specific embodiments of the present application have been described in combination with the drawings, it is not a limitation on the protection scope of the present application. Those skilled in the art should understand that various modifications or variations made by those skilled in the art without creative labor on the basis of the technical solutions of the present application are still within the protection scope of the present application.

Claims

1. A non-invasive household appliance identification and monitoring method, characterized by, The application relates to an electric appliance recognition method and device. The output current signal of the total plug-in board is collected in real time through a current collection device installed on the power line of the total plug-in board, and the collected current signal is converted into a discretized digital signal; The discretized digital signal is divided into multiple time windows through a local window with a predetermined length; The discrete signal in each time window is subjected to fast Fourier transform to extract the frequency domain harmonic features thereof; The maximum adjacent difference peak value is obtained by calculating the difference between adjacent sampling points in the discrete signal sequence in the time window; meanwhile, the transient auxiliary features and the steady-state auxiliary features in each time window are extracted; The frequency domain harmonic features, the maximum adjacent difference peak value, the transient auxiliary features and the steady-state auxiliary features extracted in all the time windows are input into a trained electric appliance recognition model to obtain the working states of various household electric appliances on the total plug-in board.

2. A non-intrusive household appliance identification and monitoring method as claimed in claim 1, characterized in that, The current collection device comprises a current detection circuit, the current detection circuit comprises a current transformer, a power bus passes through the circular ring of the primary coil of the current transformer, the secondary coil of the current transformer is connected with a trans-impedance amplifier, the non-inverting input end of the trans-impedance amplifier is grounded, the inverting input end of the trans-impedance amplifier is connected with the secondary coil of the current transformer and a feedback resistor network respectively, and the output end generates a voltage signal proportional to the input current. The feedback resistor network comprises a resistor R10, a resistor R11 and a double-way relay U4; the resistor R10 or the resistor R11 is connected between the reverse input end and the output end of the trans-impedance amplifier by controlling the switch of the double-way relay; the resistance value of the resistor R10 is smaller than that of the resistor R11.

3. A non-intrusive household appliance identification and monitoring method as claimed in claim 2, characterized in that, The coil of the double-way relay U4 is connected with the source electrode of a switch tube Q2, the gate electrode of the switch tube Q2 is connected with a control signal input end H7 through a resistor R14, the gate electrode of the switch tube Q2 is connected with a pull-down resistor R15 and then grounded, the drain electrode of the switch tube Q2 is grounded, and the coil of the double-way relay U4 is connected with a continuous diode D4 in parallel.

4. A non-intrusive household appliance identification and monitoring method as claimed in claim 1, characterized in that, The current collection device further comprises a voltage reduction circuit, the voltage reduction circuit comprises a power input terminal P4, the firewire input end of the power input terminal P4 is connected with a resistor R16, an optical coupler U7 and an output interface H1 in sequence, the input end of the optical coupler U7 is connected with a diode D5 in anti-parallel connection, the output end of the optical coupler U7 and the output interface H1 are connected to 5V through a pull-up resistor R17 to realize weak pull-up of the line, so that the output voltage is lower than 5V, and the degree of conduction of the photosensitive transistor in the optical coupler U7 determines the drop amplitude of the output voltage.

5. A non-intrusive household appliance identification and monitoring method as claimed in claim 1, characterized in that, The transient auxiliary features in each time window comprise the maximum instantaneous current change rate and the absolute amplitude of the transient peak current. The maximum instantaneous current change rate is the maximum current change amount in unit time at the position of the maximum adjacent difference peak point, and the absolute amplitude of the transient peak current is the maximum value of the current absolute value in the time window.

6. A non-intrusive household appliance identification and monitoring method as claimed in claim 1, characterized in that, The steady-state auxiliary features in each time window comprise the current effective value, the active power, the reactive power, the multiple harmonic and the total harmonic current distortion rate, and the total harmonic current distortion rate is specifically the ratio of the sum of the effective values of all high-order harmonic currents to the effective value of the fundamental wave current in the time window.

7. A non-intrusive household appliance identification and monitoring method as claimed in claim 1, characterized in that, The 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 in sequence; wherein a residual connection is introduced in the self-attention module, the feature fusion layer splices and compresses the multi-scale features extracted by the feature extraction layer into a unified feature representation as the input of the self-attention module, the self-attention module calculates the dependency between features in different dimensions, and the residual connection adds the output of the self-attention module to the feature representation of the original input to retain the original feature information; the final output is input into the Transformer encoding module, the Transformer encoding module can capture the complex relationship and context information between features and weight them according to their correlation, thereby learning the higher-level relevance between different appliance behavior patterns.

8. A non-intrusive household appliance identification and monitoring method as claimed in claim 7, characterized in that, The Transformer encoding module comprises a time-frequency perception dynamic position encoder and a Transformer encoder; the time-frequency perception dynamic position encoder generates a fixed or time-based encoding, adds the encoding to the feature embedding of the input sequence, and then sends the result to the Transformer encoder.

9. A non-intrusive household appliance identification and monitoring method as claimed in claim 8, characterized in that, The input of the time-frequency perception dynamic position encoder includes phase information of a fundamental frequency and important harmonic components in each time window The physical feature vector is composed of a maximum neighbor difference peak value, a transient auxiliary feature, and a steady-state auxiliary feature And a gain switching state 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: ； wherein, denotes a time step, denotes a base timing position encoding, is a weight for the base timing position encoding, is an enhanced physical feature embedding, obtained by physical prior information in the current signal.

10. A non-intrusive household appliance identification and monitoring system, characterized by, It comprises: A current acquisition device is installed on the power line of the total plug-in board to acquire the output current signal of the total plug-in board in real time and convert the acquired current signal into a discretized digital signal. A signal processing module is used to divide the discretized digital signal into multiple time windows through a local window of a predetermined length. A feature extraction module is used to perform fast Fourier transform on the discrete signal in 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 in the time window; at the same time, the transient auxiliary features and the steady-state auxiliary features in each time window are extracted. A state recognition module is used to input the frequency domain harmonic features, the maximum adjacent difference peak value, the transient auxiliary features and the steady-state auxiliary features extracted in all time windows into the trained appliance recognition model to obtain the working state of each household appliance on the total plug-in board.

Images