Intelligent comprehensive monitoring and self-adaptive compensation method and system for low-voltage switch cabinet

Abstract

The application relates to the technical field of low-voltage power distribution and intelligent power management, and discloses a low-voltage switch cabinet intelligent comprehensive monitoring and self-adaptive compensation method and system, which comprises the following steps: synchronously collecting three-phase electrical quantities, temperatures, switch states and zero-sequence currents through a multi-type sensor array, and generating standardized effective operation data through pretreatment; extracting electrical fingerprints of each power frequency cycle and identifying the load types and operation conditions of each branch; performing long short-term memory network prediction, arc slope sensing and event-triggered negative sequence compensation according to the operation conditions and clean high-frequency current waveforms, and driving a voltage source type inverter to output compensation operation records; and performing thermal resistance identification, zero-sequence spectrum decomposition and cross-validation by associating the temperatures, the zero-sequence currents and the compensation records, so as to generate health degree scores and operation logs. The application can improve the reactive power compensation timing matching capability, the branch condition identification precision and the multi-physical quantity abnormality diagnosis reliability under the impact load scene.

Description

Technical Field

[0001] This invention relates to the field of low-voltage power distribution and intelligent power management technology, and more specifically, to a method and system for intelligent integrated monitoring and adaptive compensation of low-voltage switchgear. Background Technology

[0002] In welding production workshops, low-voltage switchgear is often connected to equipment such as resistance spot welding machines, arc welding robots, cooling fans, and motors. The reactive power load changes rapidly during arc ignition, peak welding time, and no-load phases, causing three-phase imbalance and power factor fluctuations. Existing power distribution monitoring relies mainly on bus-side aggregated power metering or fixed threshold over-limit alarms. The monitoring and compensation systems are independent of each other, making it difficult to distinguish load types and process conditions for each outgoing branch, and making it difficult to provide predictable and reusable control inputs for categorized reactive power compensation.

[0003] Current reactive power compensation methods mostly employ staged switching of capacitor banks or use static var generators based on phase-locked loops for tracking compensation. Compensation commands often lag behind the arrival time of the welding machine's reactive power pulses. The rapid switching of single-phase welding machines within a single power frequency cycle makes it difficult for negative-sequence phase tracking to converge within the effective arcing period. Contact point temperature monitoring often uses absolute temperature threshold criteria, and the natural temperature rise under high current conditions can easily mask the additional temperature rise caused by contact degradation. Zero-sequence current exceeding limits alarms also struggle to distinguish between different physical causes such as three-phase imbalance, grounding leakage, and loose contact, resulting in both false alarms and missed alarms.

[0004] The aforementioned technical approaches have shortcomings in terms of monitoring granularity, compensation timing coordination, and multi-physical quantity fault identification. They are difficult to achieve an integrated closed loop of data acquisition, operating condition identification, adaptive compensation, and health diagnosis in scenarios with dense impact reactive loads. They are also difficult to meet the power distribution quality requirements of continuous production in welding workshops, thus restricting the improvement of stable power factor operation, three-phase busbar balance, and the efficiency of preventive maintenance and fault handling of switchgear equipment. Summary of the Invention

[0005] This invention provides an intelligent integrated monitoring and adaptive compensation method and system for low-voltage switchgear, which solves the technical problems in related technologies, such as the difficulty in effectively compensating for instantaneous impact reactive power in scenarios such as welding workshops within the arcing cycle, the difficulty in accurately identifying the load type and operating conditions of each branch, the lag of traditional negative sequence compensation tracking behind the rapid switching of single-phase welding machines, and the difficulty in reliably distinguishing contact point degradation and zero sequence abnormalities under multiple physical quantity conditions.

[0006] This invention provides an intelligent integrated monitoring and adaptive compensation method for low-voltage switchgear, comprising the following steps:

[0007] S1, through the multi-type sensor array of the switch cabinet, synchronously collects three-phase voltage, current, temperature, switch status and zero-sequence current, and forms a multi-dimensional synchronous raw time series dataset through multi-channel synchronous alignment;

[0008] S2 performs weighted sliding filtering, threshold verification, switch debouncing and alignment calibration on the multidimensional synchronous raw time series dataset to generate standardized operating data, including clean high-frequency current waveform sequence, bus three-phase voltage waveform sequence, temperature sequence at each point and zero-sequence current waveform sequence of each circuit.

[0009] S3 extracts electrical fingerprints for each power frequency cycle from standardized operating data, and identifies them through the encoder attention classification head to obtain the load type identifier and operating condition status identifier of each branch.

[0010] S4, based on the operating status indicators of each branch and the clean high-frequency current waveform sequence, predicts through long short-term memory network, senses arc slope and event-triggered negative sequence compensation, drives the voltage source inverter for stepless compensation through feedforward and proportional-integral feedback, and outputs the compensation operation record.

[0011] S5 associates the temperature sequence of each point, the zero-sequence current waveform sequence of each circuit with the operating condition status identifier and compensation operation record of each branch, and generates a health score, categorized abnormal alarm and operation log for each contact point through recursive least squares thermal resistance identification, zero-sequence spectrum decomposition and cross-validation.

[0012] In a preferred embodiment, in S1, a through-core broadband current transformer is independently installed for each outgoing circuit, a capacitive voltage divider is installed on each phase of the three-phase busbar, a temperature sensor is independently installed at the contact points of the busbar joint, the circuit breaker in-line and out-line terminals and the cable crimp terminals, and a dedicated zero-sequence current transformer is installed for each outgoing circuit.

[0013] Each acquisition channel shares the same clock reference, and after being timestamped and synchronized with multiple channels, a multi-dimensional synchronous original time series dataset is formed.

[0014] In a preferred embodiment, in step S2, a weighted moving average filter is applied to the three-phase current and voltage waveform sequences, and the weights of each sampling point within the time window are assigned according to the exponential decay law, with the window time span constrained by preserving the reactive pulse timing characteristics.

[0015] The upper limit of the current range is set as the product of the rated current and the first threshold. Sampling points exceeding the range are marked as abnormal and removed from the data stream. A delay confirmation mechanism is adopted for the switch status signal. The duration of the new status must exceed the second threshold to be confirmed as a switch action event. The second threshold is determined according to the upper limit of the rated bounce time of the selected circuit breaker. After completing the weighted moving average filtering, current threshold verification and switch status delay confirmation, the standardized operating data is output after sensor system error compensation.

[0016] In a preferred embodiment, in step S3, each power frequency cycle is used as the analysis window to perform a short-time Fourier transform on the clean high-frequency current waveform sequence of each outgoing circuit, extract the current amplitude of each odd harmonic and the phase angle relative to the bus voltage reference, and use the sine value of the phase difference between each harmonic component and the corresponding voltage component as the weight to calculate the reactive power contribution ratio of each harmonic component, thereby forming a harmonic reactive power distribution vector.

[0017] Extract the maximum instantaneous rate of rise of the current waveform and the area and aspect ratio of the instantaneous current-voltage trajectory within the current window; concatenate the harmonic reactive power distribution vector, the maximum instantaneous rate of rise of the current, and the trajectory shape parameters to form a multidimensional electrical fingerprint feature vector for the current power frequency cycle.

[0018] In a preferred embodiment, in S3, the operating condition recognition network adopts a three-segment structure of encoder, attention, and classification head: the encoder consists of two fully connected layers connected in series, which sequentially compress the multidimensional electrical fingerprint feature vector into intermediate representation and latent representation; the attention segment realizes feature recombination through the inner product normalization of three sets of linear transformations of query, key and value; the classification head consists of two parallel fully connected layers that output the load type probability vector and the operating condition state probability vector respectively, and the load types include welding machine type, electric motor type, frequency converter type, resistive load type and unknown type;

[0019] Semi-supervised training is adopted: tagged samples of isolated single-machine operation and untagged mixed operation samples of normal production are collected. The nameplate type and operating condition record are used to supervise the classification head. The joint constraint of the measured reactive power of each branch and the probability of load type is written into the training objective, so that the reactive power measured by the branch current transformer and the bus voltage participates in the network parameter update.

[0020] Within each power frequency cycle, for all outgoing branches that are synchronously collected with the corresponding cycle, the measured reactive power of each branch is allocated to welding machines, motors, frequency converters, and resistive loads according to the probability of each branch's load type. Unknown types are not included in the allocation. The allocation amount must be within the reasonable reactive power range determined by the branch wiring log, equipment nameplate, branch active power, and operating conditions. Any amount exceeding the range is included in the training penalty.

[0021] The same probability-based allocation and full cabinet branch accumulation method is used for online identification to obtain the total reactive power of various loads.

[0022] In a preferred embodiment, in step S4, for automatic spot welding machine type loads, the long short-term memory network is driven by the sequence of operating condition identifiers of each welding machine branch as input to predict the operating condition of each branch in the next power frequency cycle.

[0023] When the prediction results show that a certain branch will switch from no-load to active welding state, a feedforward compensation command is immediately issued in the current power frequency cycle. After the response delay of the compensation device, the compensation is completed synchronously with the arrival time of the predicted reactive pulse.

[0024] For arc welding robot loads, early slope characteristics are extracted from the current rise waveform during the initial short-circuit stage of arc ignition. When the difference between the measured slope and the corresponding welding machine arc ignition characteristic slope reference value in the fingerprint database falls within the preset matching window, a feedforward compensation command is immediately issued according to the typical peak reactive power amplitude of the welding machine in the fingerprint database.

[0025] In a preferred embodiment, in S4, when S3 detects a single-phase load activation event, the wiring phase and typical reactive power amplitude of the corresponding load are read from the fingerprint database, the negative sequence phase-locked loop tracking process is skipped, and a negative sequence compensation current reference based on a predefined phase and amplitude is directly generated.

[0026] The phase-locked loop is activated for fine phase calibration only after the single-phase equipment has been running continuously for more than a preset number of cycles. The feedback correction uses the difference between the real-time reactive power measurement value and the target value as the error. The feedback component is output by the proportional-integral regulator and superimposed with the feedforward command. Then, the voltage source inverter is driven by the pulse width modulation command to continuously and adjustablely output the compensation current and output the compensation operation record.

[0027] In a preferred embodiment, in step S5, for each contact point temperature monitoring point, the measured temperature is used as the observed value, and the current and ambient temperature are used as inputs. The equivalent thermal resistance estimate is updated online using the recursive least squares method. Priority is given to updating during the stage when the welding machine is continuously idle for more than a preset stable time. After convergence during the initial observation period, the stable estimate is used as the reference thermal resistance value.

[0028] The degradation progress is defined as the ratio of the current estimated thermal resistance exceeding the baseline thermal resistance to the total degradation margin of the contact point from the baseline thermal resistance to the safety limit thermal resistance. A health score ranging from 0 to 100 is generated based on the degradation progress. When the degradation progress is zero, the score is full and the corresponding contact point is in good condition. When the degradation progress reaches the upper limit of the total degradation margin, the score is zero and the corresponding current thermal resistance has reached the safety limit thermal resistance. A contact degradation warning is triggered when the temperature residual continuously exceeds the third threshold and the duration exceeds the fourth threshold. A continuous degradation alarm is triggered when the daily average increase in thermal resistance exceeds the fifth threshold.

[0029] In a preferred embodiment, in step S5, a fast Fourier transform is performed on the zero-sequence current waveform of each circuit with the power frequency period as a window to extract the fundamental zero-sequence component, the third harmonic zero-sequence component, and the high-frequency broadband zero-sequence component.

[0030] The amplitude change time of the fundamental zero-sequence component is associated with the activation time of the single-phase load detected by S3. If the fundamental zero-sequence component continues to exceed the sixth threshold under the three-phase symmetrical load condition, a leakage fault warning is triggered. The sixth threshold is the converted value of the secondary side of the zero-sequence current transformer under the rated leakage current of the selected protection level.

[0031] If the residual temperature at a certain contact point rises abnormally while the high-frequency broadband zero-sequence component of the circuit shows a pulsed increase, a contact loosening fault alarm will be generated with the highest priority. The alarm content includes the location and the current health score.

[0032] In a preferred embodiment, a low-voltage switchgear intelligent integrated monitoring and adaptive compensation system is used to perform the steps in the above-mentioned low-voltage switchgear intelligent integrated monitoring and adaptive compensation method, including:

[0033] The data acquisition module is used to synchronously acquire three-phase voltage, current, temperature, switch status and zero-sequence current through a multi-type sensor array of the switch cabinet, and form a multi-dimensional synchronous raw time series dataset through multi-channel synchronous alignment;

[0034] The data preprocessing module is used to perform weighted sliding filtering, threshold verification, switch debouncing and alignment calibration on the multidimensional synchronous raw time series dataset to generate standardized operating data, including clean high-frequency current waveform sequence, bus three-phase voltage waveform sequence, temperature sequence of each point and zero-sequence current waveform sequence of each circuit.

[0035] The operating condition identification module is used to extract electrical fingerprints for each power frequency cycle from standardized operating data, and then identify them through the encoder attention classification head to obtain the load type identifier and operating condition status identifier of each branch.

[0036] The adaptive compensation module is used to predict the operating conditions of each branch and the clean high-frequency current waveform sequence through long short-term memory network prediction, arc initiation slope sensing and event-triggered negative sequence compensation, feedforward and proportional-integral feedback to drive the voltage source inverter stepless compensation, and output the compensation operation record.

[0037] The comprehensive diagnostic module is used to associate the temperature sequence of each point, the zero-sequence current waveform sequence of each circuit with the operating condition status identifier and compensation operation record of each branch. Through recursive least squares thermal resistance identification, zero-sequence spectrum decomposition and cross-validation, it generates health scores, categorized abnormal alarms and operation logs for each contact point.

[0038] The beneficial effects of this invention are as follows:

[0039] By synchronously acquiring data through the shared clock reference of the through-core current transformers and bus voltage sensors in each outgoing circuit, and through weighted sliding filtering and threshold verification, a clean high-frequency current waveform sequence and a bus three-phase voltage waveform sequence are formed. Using the power frequency cycle as a window, a multi-dimensional electrical fingerprint is extracted and the load type and operating condition status of each branch are identified. Based on this, long short-term memory network operating condition prediction is implemented for automatic spot welding machine loads, and arc ignition slope sensing is implemented for arc welding robot loads. Combined with event-triggered negative sequence compensation and proportional-integral feedback drive voltage source inverter stepless compensation, the establishment time of compensation current and reactive pulse are connected in time sequence, improving power factor fluctuation and three-phase imbalance.

[0040] For each contact point, the equivalent thermal resistance is identified online using the recursive least squares method, and the health score is calculated based on the benchmark thermal resistance and the safety limit thermal resistance. The zero-sequence current of each circuit is decomposed into fundamental, third harmonic, and high-frequency broadband components, and cross-validated with the operating condition status indicator and the compensation operation record on the time axis. The abnormal types such as contact degradation, leakage fault, resonance risk and contact loosening are distinguished, and the abnormal alarms and structured operation logs are output, providing traceable data basis for the operation and maintenance of low-voltage switchgear throughout its entire life cycle and for post-event review. Attached Figure Description

[0041] Figure 1 This is a flowchart of an intelligent integrated monitoring and adaptive compensation method for low-voltage switchgear according to the present invention;

[0042] Figure 2 This is a flowchart of an intelligent integrated monitoring and adaptive compensation method for low-voltage switchgear according to the present invention. Figure 1 ;

[0043] Figure 3 This is a flowchart of an intelligent integrated monitoring and adaptive compensation method for low-voltage switchgear according to the present invention. Figure 2 . Detailed Implementation

[0044] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0045] At least one embodiment of the present invention discloses an intelligent integrated monitoring and adaptive compensation method for low-voltage switchgear, such as... Figures 1 to 3 As shown, it includes the following steps:

[0046] S1, through the multi-type sensor array of the switch cabinet, synchronously collects three-phase voltage, current, temperature, switch status and zero-sequence current, and forms a multi-dimensional synchronous raw time series dataset through multi-channel synchronous alignment;

[0047] In one embodiment of the present invention, the layout design of the sensors follows the principle of proximity and the principle of loop integrity, that is, each type of sensor is installed as close as possible to the physical location most directly reflected by the measurement, and the complete electrical link of each power distribution circuit from the power supply side to the load side is within the monitoring coverage.

[0048] Specifically, for voltage acquisition, high-precision capacitive voltage divider sensors are installed at the corresponding positions of phases A, B, and C of the three-phase busbar to achieve synchronous acquisition of three-phase line voltages and phase voltages. The effective sampling frequency of the voltage sensors is higher than the power grid frequency (the default value is 20 times the power grid frequency) to ensure that the phase and amplitude information of all odd harmonics (3rd, 5th, up to 15th) other than the fundamental frequency can be completely recorded. The voltage reference phase is used as a unified reference in subsequent power factor calculation and reactive power component extraction. The acquisition timing of the three-phase signals is strictly aligned, and the time deviation between different phases is kept within one sampling point.

[0049] For current acquisition, a through-core broadband current transformer is installed on each outgoing circuit. The effective bandwidth of the transformer also meets the acquisition requirements for the 15th harmonic. The through-core structure requires no modification to the circuit under test, ensuring both the physical reliability of the acquisition and the convenience of modification and installation. The branch current transformers and bus voltage sensors of each circuit share the same clock reference, and the sampling timing is strictly synchronized. This allows the active power, reactive power, and power factor of each outgoing circuit to be directly calculated at each sampling time. This independent measurement capability of each circuit is the data basis for the subsequent S3 step of extracting electrical fingerprints by circuit. The current transition time constant generated by welding machine loads during the arc ignition process is usually on the order of milliseconds. The current sampling rate design can fully capture the current waveform details during the arc ignition transition period, providing a sufficient signal basis for the early slope feature extraction of the arc ignition process.

[0050] For temperature acquisition, NTC thermistors or PT100 platinum resistance surface mount sensors are used, and are mainly deployed at high-risk contact points such as phase-to-phase and branch bifurcation connections of busbars A / B / C, incoming and outgoing terminals of each outgoing circuit breaker, and crimp terminals of busbars and cables. One sensor is independently deployed at each contact point, and thermally conductive silicone grease is used to fill the gap between the sensor and the contact metal surface to ensure good thermal contact. Since the rate of change of temperature signals is much lower than that of electrical signals, the temperature sampling frequency can be set to 1 / 10 to 1 / 50 of the electrical sampling frequency, effectively reducing data storage while meeting the needs of thermal event capture.

[0051] For switch status acquisition, the auxiliary contact signals of each circuit breaker and disconnector are connected to the input acquisition unit after being isolated by optocouplers. The optocoupler structure physically isolates the control circuit from the high-voltage circuit, effectively suppressing electromagnetic interference in the industrial environment from affecting the status signals. Each input channel corresponds to the on / off state of a switchgear. A precise timestamp is recorded when the signal changes, and the status data and electrical parameter data share the same clock reference.

[0052] For zero-sequence current acquisition, dedicated zero-sequence current transformers are installed in each outgoing circuit. The three-phase conductors pass through the transformer core simultaneously, and the current induced by the transformer is equal to the sum of the instantaneous values ​​of the three-phase currents (i.e., zero-sequence current). Its sampling frequency is consistent with that of the phase current. A set of summary-type zero-sequence current transformers is installed at the main outgoing line position of the bus to provide bus-level reference values ​​for subsequent multi-cause diagnosis of zero-sequence current.

[0053] The raw digital signals from all acquisition channels are uniformly fed into the data acquisition and processing unit, and are processed by timestamp marking and multi-channel synchronization alignment to form a multi-dimensional synchronous raw time-series dataset with the power frequency cycle as the basic time granularity. This dataset includes high-frequency current waveform sequences of each outgoing circuit, three-phase voltage waveform sequences of the bus, temperature sampling sequences of each contact point, switch status event sequences, and zero-sequence current waveform sequences of each circuit.

[0054] S2 performs weighted sliding filtering, threshold verification, switch debouncing and alignment calibration on the multidimensional synchronous raw time series dataset to generate standardized operating data, including clean high-frequency current waveform sequence, bus three-phase voltage waveform sequence, temperature sequence at each point and zero-sequence current waveform sequence of each circuit.

[0055] Although the multidimensional synchronous raw time series dataset has sufficient time resolution, it inevitably contains various interference components from the industrial field: the high-frequency electromagnetic coupling noise generated by the frequent switching action on the busbar will be superimposed on the current and voltage sampling signals; the zero drift and gain offset generated by the long-term operation of the sensor will cause systematic measurement errors; the mechanical elastic jitter of the switch auxiliary contact will generate multiple rapid jumps of false signals near the real state change event.

[0056] In one embodiment of the present invention, a weighted moving average filtering algorithm is used for basic noise reduction of the three-phase current and voltage waveform sequences. Within a time window, historical sampling points are weighted according to an exponential decay law, with samples closer to the current time assigned a larger weight coefficient. The decay rate is determined based on the ratio of the center frequency of the main noise band of the measured signal to the current sampling frequency, enabling the filter to effectively suppress high-frequency noise while retaining current jump information caused by rapid changes such as arc ignition in the welding machine. Regarding the selection of the window width, the window time span is limited to within 1 / 4 of the fundamental frequency period, serving as a trade-off between retaining the reactive pulse timing characteristics of the welding machine and filtering out high-frequency noise. For temperature signals, since their rate of change is significantly lower than that of electrical signals, a longer time span mean filtering window is used, and obvious sampling abrupt changes (such as full-range jumps caused by sensor disconnection or short circuits) are directly marked as invalid points and replaced with linear interpolation.

[0057] Specifically, in the threshold verification process, normal operating ranges for each physical quantity are established based on the rated parameters on the switchgear nameplate. The upper limit of the effective range for current parameters is the product of the rated current and a multiple of the first threshold. Sampling points exceeding this range are marked as abnormal and removed from the valid data stream, while the abnormal markers are retained for reference in the S5 diagnostic step. The first threshold is determined based on the safety margin of the circuit breaker's rated breaking capacity, referring to the provisions of GB / T14048.2 for overcurrent protection; the default value of the first threshold is 1.3 times the rated current. The effective range for voltage parameters is set according to the allowable deviation range of the power grid (rated value ±10%), and the effective range for temperature parameters is set to a reasonable physical range within the sensor's range and above the lower limit of the ambient temperature. The frequency of the removed data points constitutes an auxiliary indicator of equipment abnormality and is referenced in the S5 step.

[0058] For the anti-jitter processing of switch status signals, a delayed confirmation mechanism is adopted: after the signal enters a new state, it must remain in that state for more than the second threshold duration before the state change is confirmed as a valid switch action event and a timestamp is recorded; if the signal changes state again within the second threshold duration, the entire sequence is judged as mechanical jitter, and only the jitter event is recorded without updating the switch status. The value of the second threshold should be significantly greater than the typical jitter duration of mechanical contacts, but much smaller than the shortest time interval between normal closing and opening operations. Its determination is based on 3 times the upper limit of the rated bounce duration given in the mechanical characteristic manual of the selected circuit breaker model; the default value of the second threshold is 15ms. For switch action events confirmed by anti-jitter, their timestamps serve as a key reference for load switching events in the load condition analysis of S3, and directly participate in the correlation judgment of the cause of zero-sequence current in S5.

[0059] After processing, the data from each channel undergoes final time alignment verification and sensor system error compensation (based on offline calibration coefficients). The output is standardized and effective operating data with the power frequency cycle as the basic analysis unit, including the clean high-frequency current waveform sequence of each outgoing circuit, the three-phase voltage waveform sequence of the bus, the effective temperature sequence of each point, the confirmed switch status event sequence, and the zero-sequence current waveform sequence of each circuit.

[0060] S3 extracts electrical fingerprints for each power frequency cycle from standardized operating data, and identifies them through the encoder attention classification head to obtain the load type identifier and operating condition status identifier of each branch.

[0061] In one embodiment of the present invention, two difficulties exist in the specific implementation of power distribution in the welding workshop: While the current transformers of each outgoing circuit can directly provide the instantaneous current value of each branch and calculate the reactive power of that circuit, this value itself cannot describe the current operating condition. The reactive power of the arc welding machine differs by more than an order of magnitude between the arc ignition period, the peak welding period, and the no-load period, and predictive control relies on this operating condition sequence to determine the timing and amplitude of the next reactive pulse. If the current process stage of each welding machine cannot be accurately identified, the prediction channel loses its most crucial input basis. Furthermore, in the case of multiple small-to-medium power devices (e.g., multiple manual arc welding machines protected by the same circuit breaker) connected to a branch, the branch CT measures the superposition of currents from multiple devices; the effective current value alone cannot distinguish the state contribution of each device. These two problems necessitate multi-band feature analysis of the branch signals to extract electrical fingerprints that can distinguish equipment types and operating conditions, thus providing accurate operating condition input for subsequent categorized prediction.

[0062] The operating mechanisms of different types of loads determine that their current waveforms exhibit varying harmonic distributions in the frequency domain. Furthermore, the harmonic combinations of the same type of load under different operating conditions also show significant differences. This is the physical basis for electrical fingerprinting to distinguish load types and operating states. The current waveform of resistive heating loads is approximately sinusoidal with low harmonic content; the fundamental current amplitude of induction motors is significant, the power factor lags, and the 5th and 7th harmonics are relatively weak, with the fundamental current envelope exhibiting specific transient change patterns during start-up and shutdown; the nonlinear conduction characteristics of the rectifier bridge at the front end of the frequency converter result in currents containing prominent 5th, 7th, 11th, and 13th harmonics; the nonlinear conductivity of the arc in arc welding machines makes the 3rd harmonic current component significant, and the current rise slope is extremely high during the arc ignition transition period, forming a highly identifiable power envelope in the time domain throughout the entire welding process cycle. These differences constitute the signal basis for identifying loads by type and decomposing reactive power contributions according to operating conditions.

[0063] Multi-band feature extraction uses the clean high-frequency current waveform sequence of each outgoing circuit and the three-phase voltage waveform sequence of the bus output from S2 as input. Each power frequency cycle (20ms) is used as a basic analysis window. A short-time Fourier transform is performed on the circuit current waveform within this window to extract the fundamental frequency (50Hz) and the phase angles of the current corresponding to the 3rd to 15th odd harmonics, totaling 8 harmonic components. The contribution ratio of each harmonic component to the total reactive power of the circuit is calculated using the sinusoidal value of the phase difference between each harmonic current component and its corresponding voltage component, forming the harmonic reactive power distribution vector of the circuit. Simultaneously, the maximum instantaneous rate of rise of the current waveform within the current analysis window is extracted, which is the maximum value of the ratio of the current difference between adjacent sampling points to the sampling time interval within the 20ms window. This feature has a low value during normal operating power frequency cycles, but a significantly higher value during the initial short-circuit stage of the arc ignition transition period of the arc welding machine. Combined with the harmonic distribution characteristics, it constitutes an effective basis for identifying the arc ignition condition, and shares the same calculation definition with the detection object of the S4 arc ignition early sensing link, ensuring that the description of the same physical phenomenon by the two steps is self-consistent. In addition, the area and aspect ratio of the VI feature trajectory obtained by plotting the instantaneous sampling values ​​of current and voltage within the window as a planar trajectory are extracted. The harmonic reactive power distribution vector, the maximum instantaneous rate of rise of the current, and the VI trajectory shape parameters are concatenated to form a multi-dimensional electrical fingerprint feature vector corresponding to the power frequency cycle, denoted as x. Its dimensions are the amplitude and phase of each of the 8 harmonics (16 dimensions) plus the maximum instantaneous rate of rise of the current, the VI area, and the aspect ratio (3 dimensions), for a total of 19 dimensions.

[0064] In one embodiment of the present invention, the working condition recognition network adopts a three-segment structure of encoder, attention, and classification head. The input and output dimensions and data transformation methods of each segment are as follows: The encoder segment consists of two fully connected layers connected in series: the first fully connected layer receives a 19-dimensional feature vector x, which is activated by a modified linear unit after a linear transformation of the weight matrix and the bias vector, and outputs a 32-dimensional intermediate representation; the second fully connected layer further compresses the 32-dimensional intermediate representation into a 16-dimensional latent representation z. The default value of 16 dimensions is determined according to the number of field equipment types. This representation performs dimensionality reduction and compression on the original features in the latent space, retaining the main change direction of the features in each frequency band. The attention segment treats z as a sequence of 16 scalar features, and generates a query matrix, a key matrix, and a value matrix through three sets of independent linear transformations. The attention weight of each position to the other positions is obtained by normalizing the inner product of the query vector and all key vectors at that position. The value matrix is ​​weighted and summed with this weight to obtain the 16-dimensional representation z_attn after attention reorganization. The physical meaning of the attention mechanism here is to identify which combinations of frequency components in the current feature vector best distinguish between load type and operating condition. For example, the joint activation mode of the third harmonic amplitude and the maximum instantaneous current rise rate during the arc ignition period of an arc welding machine is fundamentally different from that during the peak welding period. The attention weights will automatically amplify the contribution of this discriminative combination to subsequent classification. The classification header consists of two parallel fully connected layers: one maps z_attn to a K-dimensional load type probability vector (normalized by softmax, where K is the predefined number of load types), and the other independently maps it to an M-dimensional operating state probability vector (also normalized by softmax). For the welding workshop scenario, K is 5 (welding machine type, motor type, frequency converter type, resistive load type, unknown type), and M is 4 (no load, arc ignition / start-up transition, full load operation, cooling / deceleration). The specific number is set according to the on-site equipment configuration.

[0065] In one embodiment of the present invention, model construction is divided into an offline training phase and an online operation phase. The offline training phase uses known type information of factory equipment as the supervision source. Labeled electrical fingerprint samples are collected for each piece of equipment operating independently under different load rates (no load, half load, full load) and different process parameter settings, forming a labeled sample set. Simultaneously, an unlabeled sample set is collected when multiple pieces of equipment are running in combination, used for on-site constraint enhancement. The type and operating condition labels of each piece of equipment in the labeled samples are manually labeled based on equipment nameplates and operating condition records. The unlabeled samples are directly collected continuously during normal production, without the need for production shutdown or isolation.

[0066] The training strategy incorporates the measured reactive power of branches, which is readily available online in S2, into the training loop of the identification network. This ensures that the classification probability not only fits a small number of isolated labels on a single machine but also aligns with the reactive power feasible region derived from the nameplate. Therefore, it avoids using a general classification loss decoupled from the measured reactive power of branches as the sole training basis and utilizes unlabeled mixed operation data from normal production for parameter updates. Labeled samples are supervised by annotations on the type and operating condition classification heads under isolated operating conditions. Regardless of whether they are labeled or unlabeled, all outgoing branches participating in the current parameter update simultaneously apply the following physical coupling term.

[0067] The time granularity of the physical coupling term is consistent with the power frequency cycle analysis unit used in S1 and S2 and the electrical fingerprint extraction window in this step: taking a single power frequency cycle as the time unit, and taking all outgoing branches after multi-channel synchronous alignment within that cycle to form a set. This aligns with the online phase's cycle-by-cycle process of extracting and identifying fingerprints from each outgoing branch, then accumulating reactive power by category. (Branch recording) The measured reactive and active power are respectively , (Calculated from the branch current transformer and bus voltage waveforms within this cycle), the classification head outputs via softmax. In dimensional type probability, take The components (welding machines, electric motors, frequency converters, resistive loads, excluding unknown categories) are denoted as... The reactive power corresponding to the probability of the unknown class is not included in the allocation. The core constraint is: to softly allocate the reactive power of the branch to each category according to the identification probability, and to require that the allocation result falls into the category of equipment in the current... Operating conditions The following is the unfeasible feasible region:

[0068]

[0069]

[0070] in, It is the set of all outgoing branches after multi-channel synchronous alignment within a single power frequency cycle; For set Branch numbering in the middle; For predefined load category numbers, These correspond to welding machines, electric motors, frequency converters, and resistive loads, respectively. The total number of load types output by the classification header (default value is 1). (including unknown classes) , Branch roads The measured reactive power and measured active power during this power frequency cycle; branch road The load type probability vector of the first Class components; To be according to from Soft allocation to the first The reactive power of the class; For the first Classless feasible region; , These are the lower and upper bounds of the feasible region, respectively, and are determined by the branch wiring log, equipment nameplates, and working conditions Sure; branch road Identify or label the operating conditions (no load, arc ignition / start-up transition, full load operation, cooling / deceleration); For the whole cabinet Reactive power cumulative value of different load types, total reactive power of welding machines, motors, and frequency converters. , , They are respectively When retrieving the corresponding category .

[0071] when At that time, the out-of-bounds amount is included in the training objective function through a differentiable penalty function; the above soft allocation relationship is used for the reactive power of the branches in both the training and online inference phases, and the entire cabinet's first... Class load reactive power Summary. , The determination rule is: only when the branch connection ledger records the branch Possibly accessing the first When dealing with this type of equipment, the parameters on the equipment's nameplate and the measured active power of the branch within that power frequency cycle should be considered. and branch identification or labeling conditions Calculate and form Categories not recorded in the ledger are not subject to the aforementioned coupling constraints on that branch. Welding machine category... Take when unloaded or cooling ;exist For arc initiation / start-up transition or full-load operation, take , ,in The rated welding performance of the welding machine nameplate for this branch in the ledger is... The typical default value for arc power factor (taken as follows) The upper limit of peak reactive power in fusion welding is defined by the capacity on the welding machine's nameplate and not by this cycle. To avoid light load and no-load cycles Approaching zero leads to distortion of the upper bound. For motors, the load is calculated based on the lagging power factor, taking... , , , The default values ​​for the lower and upper limits of the power factor on the nameplate of the motor in this branch circuit are (typically taken as follows). , ). Variable frequency type , , The upper limit of the reactive power to active power ratio of the front-end rectifier bridge is the typical default value (taken as...). Resistive loads are classified as... The above. , , , During the system installation and commissioning phase, adjustments can be made based on the measured waveforms of the branch circuits. During arc initiation / start-up transition, reactive power may be briefly higher than steady-state levels. The network output is constrained solely by training soft constraints, without triggering protection or fault alarms. When a branch ledger corresponds to only a single type of device, This refers to the reactive power measurement value of this type, and the soft allocation result is consistent with... One-to-one correspondence; when a branch line connects to multiple types of devices, For branch circuit reactive power synthesis, In the various categories involved in the ledger The decomposition is completed under the joint constraints, which is suitable for the field conditions where only one set of current transformers is configured for the outgoing branch and only the combined electrical quantities of the branch can be obtained.

[0072] During the online operation phase, the model receives the 19-dimensional electrical fingerprint feature vectors of each outgoing branch updated every power frequency cycle. It completes forward inference within milliseconds, obtaining the load type probability vector and operating condition probability vector for each branch from the classification head. The load type identifier and operating condition identifier are the category and condition with the highest probability, respectively. The reactive power of each branch is calculated from the standardized current output from S2 and the bus voltage. Then according to the type probability vector The total reactive power of the welding machine is obtained by soft allocation and summation across the entire cabinet. Total reactive power of electric motors Variable frequency drive total reactive power (Right now (values ​​in the corresponding category) and total active power on the bus side No results in sight With system power factor (Obtained from the sum of active and reactive power measurement results of all branches in the cabinet), for direct use in the predictive modeling of S4 and the cross-diagnosis of S5. When the feature vector of a certain branch continuously falls outside the reasonable feature distribution range of the known type and the classification confidence is lower than the set lower limit for several consecutive periods, it is determined that the branch may have been connected to a new type of device, triggering the online incremental update process: the fingerprint sample library is locally expanded with the current observation data, and the classification head weight of the network is fine-tuned, while the encoder weight is kept frozen to prevent the feature representation of the existing category from being destroyed.

[0073] S4, based on the operating status indicators of each branch and the clean high-frequency current waveform sequence, predicts through long short-term memory network, senses arc slope and event-triggered negative sequence compensation, drives the voltage source inverter for stepless compensation through feedforward and proportional-integral feedback, and outputs the compensation operation record.

[0074] The output of S3 has determined the load type and current operating condition for each outgoing branch, and is continuously updated in cycles of the power frequency. This step uses these results for two levels of control objectives. On the one hand, for the operating condition evolution patterns of different types of loads, appropriate short-term reactive power demand prediction channels are established to generate feedforward compensation references that are ahead of the current moment. On the other hand, feedback correction is driven by real-time measured reactive power deviation, and after being superimposed with the feedforward command, it is output to the continuously variable compensation device, so that the compensation action is synchronized with the reactive power pulse of the load in time.

[0075] Short-term reactive power demand forecasting employs fundamentally different predictive triggering mechanisms tailored to the operational characteristics of two types of welding machines: automatic spot welding machines and arc welding robots. For automatic spot welding machine loads, each station is controlled by a PLC program, and there is a fixed phase relationship between the welding cycles of adjacent stations (for example, stations 1 and 3 alternate arc ignition according to the process design by staggering several power frequency cycles to balance peak current). The combined timing pattern of this multi-station operating condition contains far more predictable information than a single current amplitude sequence.

[0076] In one embodiment of the present invention, the prediction channel adopts a Long Short-Term Memory (LSTM) network structure, taking the working condition status identifier sequence of each welding machine branch as input. The input dimension of each time step is equal to the total number of welding machine branches, and each dimension is a code of the current working condition of the branch (no load = 0, arc ignition transition = 1, welding peak = 2, cooling = 3). The LSTM maintains the internal memory unit through the coordinated control of forget gates, input gates and output gates, learns the timing law of multi-station state transition, and predicts the working condition of each branch in the next power frequency cycle (after 20ms). When the prediction result shows that a certain welding machine branch will transition from no load state to welding active state in the next power frequency cycle, the system immediately issues a feedforward compensation command for that branch in the current power frequency cycle. After the command is delayed by Δτ by the compensation device, the compensation current is established.

[0077] The LSTM lead prediction time window is one power frequency cycle (20ms), and the default value of Δτ is on the same order of magnitude. The two are naturally aligned in time, making the establishment time of the compensation current basically synchronized with the actual time when the welding machine enters the fusion welding state. This is the core reason why this method chooses to predict the operating condition sequence rather than the reactive power value as the prediction object. The discrete transition time of the operating condition determines the compensation triggering time, and its amplitude information is supplemented by the typical reactive power amplitude of each welding machine under the corresponding operating condition in the fingerprint database (obtained from the statistical data of single-machine isolated operation during the offline calibration stage), thereby transforming the state prediction result into a feedforward compensation reference quantity. .

[0078] For arc welding robot loads, there is a no-load positioning phase between two weld seams. During this phase, the electrical signal is close to zero. Relying solely on historical electrical timing predictions during the no-load phase lacks signal support and cannot predict the arc ignition time in advance like an automatic spot welding machine. To address this specific challenge, this method utilizes the high-sampling-rate branch current waveform acquired by S1 to extract early slope features from the current rise waveform during the initial short-circuit phase of the arc transition period (i.e., the early stage after the welding torch electrode contacts the workpiece but before an arc column is formed, lasting approximately 3 to 8 ms). This enables early arc ignition sensing approximately 10 to 15 ms before the stable arc current is established.

[0079] The current rise rate during the initial short-circuit stage of arc ignition is determined by the short-circuit open-circuit voltage of the welding transformer and the total inductance of the welding circuit, and has a deterministic relationship with the rated parameters of the welding machine and the welding current setting. This characteristic slope value is statistically derived from the measured arc ignition waveforms of each arc welding robot under different welding parameter settings during offline operation and stored in the fingerprint database as a reference value for the early characteristic slope of arc ignition under the corresponding welding parameters of that robot. During online operation, when the current in a branch of a certain arc welding robot rises rapidly from near zero, and the difference between the rise rate and the reference value of the arc ignition characteristic slope of that welding machine under the current welding parameters in the fingerprint database falls within the set matching window, the system determines that the branch has entered the initial short-circuit stage of the arc ignition transition period and immediately issues a feedforward compensation command targeting the typical peak reactive power amplitude of that welding machine under the current welding parameters in the fingerprint database.

[0080] The matching window is determined as follows: During the offline calibration phase, at least 50 samples of the initial short-circuit current slope of each arc welding robot under each welding parameter setting are collected. The mean and standard deviation are statistically analyzed. The absolute value of the difference between the measured slope and the reference mean is no more than twice the sample standard deviation as the criterion for successful matching. The default value of the matching window multiple is twice the standard deviation, which can be adjusted to three times for environments with strong electromagnetic interference. This triggering time is about 10 to 15 ms earlier than the traditional triggering method based on stable arc current detection, and is on the same order of magnitude as the response delay Δτ of the compensation device. This allows the establishment time of the compensation current of the arc welding robot to be earlier than or basically synchronized with the establishment time of the stable arc column, thus realizing the advanced compensation of the reactive pulse of arc welding.

[0081] The feedforward compensation command generation denotes the complete response delay from the controller command output to the stable establishment of the compensation current on the VSI AC side as Δτ. This value is determined by actual measurement through applying a step current reference to the compensation device and measuring the time required for the response output to reach 90% of the reference value. It includes the sum of the controller calculation delay and the IGBT switching transition delay. The measured result is written into the control parameter configuration after debugging. For a typical VSI compensation device using industrial-grade IGBT modules, the default value of Δτ is 18ms. For automatic spot welding machine type loads, the feedforward reference is taken as the predicted reactive power value of the welding machine at the time (t+Δτ) predicted by LSTM. That is, the command is issued at time t, and the compensation current is established after Δτ, which is exactly aligned with the arrival time of the predicted welding machine reactive power pulse. For arc welding robot loads, the feedforward command is issued immediately at the identification time t_arc_early during the initial short-circuit stage of arc ignition. The compensation current is established at time (t_arc_early + Δτ), while the stable arc burning of the arc column is usually completed 10 to 15 ms after t_arc_early. As long as Δτ does not exceed the total duration of the arc ignition transition period, the establishment time of the compensation current is earlier than or basically synchronized with the arrival of the stable arc current, thus achieving advance compensation.

[0082] The three-phase imbalance handling mechanism is specifically designed to address the unique challenge of frequent and rapid switching of single-phase welding machines in welding workshops. Traditional negative-sequence compensation control based on phase-locked loops typically requires 2 to 3 power frequency cycles (approximately 40 to 60 ms) for negative-sequence phase tracking convergence, while the arc ignition-arc termination process of a single-phase welding machine lasts only about 1 power frequency cycle (approximately 20 ms). Traditional solutions remain in the phase-locked tracking transition phase throughout the entire effective arc burning period of the welding machine, resulting in a significant lag in the actual compensation current and an inability to effectively compensate for the imbalance during the effective operation of the welding machine. To address this issue, this method, upon detecting a single-phase load activation event in S3, directly reads the wiring phase (A / B / C phases, registered during installation and commissioning) and the typical reactive power amplitude under the current welding parameters from the fingerprint database. It skips the negative-sequence phase-locked loop tracking process and directly generates a negative-sequence compensation current reference based on a predefined phase and amplitude, completing the negative-sequence compensation output within Δτ after the event occurs. The phase-locked loop is only activated for fine phase calibration when a single-phase device operates continuously for more than three power frequency cycles to eliminate phase drift caused by arc nonlinearity. The switching between the two modes is automatically determined by the duration of the S3 operating condition, requiring no manual intervention. For reactive components that have undergone positive and negative sequence decomposition, compensation current references are independently designed in their respective rotating coordinate systems. After inverse coordinate transformation and merging, a three-phase feedforward compensation current reference vector with unequal amplitude and unequal phase is generated.

[0083] The feedback correction stage uses the difference between the real-time reactive power measurement and the system's target reactive power value as the feedback error, which is input to the proportional-integral (PI) regulator. The proportional coefficient is responsible for quickly responding to residual errors, while the integral coefficient is responsible for eliminating prediction bias and slowly changing steady-state errors. The feedback correction is superimposed with the feedforward reference to obtain the final compensation current reference, which is then converted into pulse width modulation commands for the three-phase IGBT switching devices through current demodulation and coordinate transformation. The continuously variable compensation device adopts a three-phase full-bridge voltage source inverter (VSI) topology, with energy storage capacitors configured on the DC side and the AC side connected to the bus via a filter inductor. By controlling the high-frequency pulse width modulation operation of the IGBTs, the AC output of the VSI can generate a controllable AC current that leads or lags the bus voltage by 90 degrees. The compensation capacity is continuously adjustable within the rated range, without the level limitations of traditional tiered capacitor switching. The response time from the change of control command to the establishment of the compensation current can be controlled within 20ms.

[0084] The output includes the compensation control quantity (feedforward component, feedback component and total compensation quantity) for each power frequency cycle, three-phase compensation current data and real-time state sequence of system power factor.

[0085] S5 associates the temperature sequence of each point, the zero-sequence current waveform sequence of each circuit with the operating status identifier and compensation operation record of each branch, and generates health scores, categorized abnormal alarms and operation logs for each contact point through recursive least squares thermal resistance identification, zero-sequence spectrum decomposition and cross-validation.

[0086] Using standardized effective temperature and zero-sequence current data as the primary diagnostic data sources, and load condition status indicators and compensation operation records as auxiliary correlation information, a multi-dimensional cross-validation anomaly diagnosis system is constructed. A single-dimensional sensed quantity is often insufficient to support reliable anomaly judgment; therefore, information from multiple physical dimensions must be correlated and analyzed along a common time axis. The reliability of the diagnosis is improved through mutual corroboration between anomalies in different physical quantities.

[0087] In the adaptive thermal resistance identification and contact degradation detection stage, a first-order lumped-parameter thermodynamic model is independently established for each temperature monitoring contact point deployed in S1. The contact point is simplified into a thermal resistance network node with equivalent thermal resistance R and equivalent heat capacity C. The physical law described by this model is: under the quasi-steady-state condition of constant current, the steady-state value of the contact point temperature is equal to the sum of the ambient temperature and the square of the effective current multiplied by the equivalent thermal resistance; after a step change in current, the temperature exponentially approaches the new steady-state condition with the thermal time constant determined by the product of heat capacity and thermal resistance. Recursive least squares (RLS) is used for online parameter identification of this model. The measured temperature at each sampling moment is used as the observation value, and the current current and ambient temperature are used as inputs. The estimated equivalent thermal resistance R_est is continuously updated according to the recursive formula, while maintaining the covariance matrix to track the estimation uncertainty. Recursive least squares is a mature method in the field of parameter identification; only its application in this scenario is described here. Under complex load conditions with frequent switching of the welding machine, the system prioritizes data from the period when the welding machine is in a continuous no-load state (lasting more than 30 seconds) for RLS updates. During this period, the load current level is stable, making it the optimal time to obtain high-quality thermal resistance estimation samples, effectively avoiding interference from frequent transients of the welding machine on parameter identification and convergence. After a sufficiently long initial observation period (typically 1 to 2 days of normal production), R_est converges to a stable baseline value R_base, which reflects the true thermal resistance characteristics of the contact point under good contact conditions.

[0088] The health score of each contact point quantitatively describes the degree of degradation of the current contact condition relative to the good condition: the numerator is the amount by which the current thermal resistance estimate exceeds the baseline thermal resistance value, and the denominator is the total degradation margin of the contact point from the baseline thermal resistance to the safety limit thermal resistance (safety limit thermal resistance minus the baseline thermal resistance value). The ratio of the two is the degradation progress. A health score ranging from 0 to 100 is generated based on the degradation progress. When the degradation progress is zero, the score is the full score of 100. When the degradation progress reaches the upper limit of the total degradation margin, the score is 0. When the calculated result is lower than 0, it is truncated to 0, and when it is higher than 100, it is the full score of 100. The reference thermal resistance value is taken as the stable estimate of the contact point after the RLS identification converges after a sufficiently long observation period under good contact conditions. The safety limit thermal resistance is determined by dividing the maximum allowable temperature rise of the contact material (65℃ for copper / copper alloy contact materials, according to GB 7251.1) by the square of the rated load current of the contact point. This is the equivalent thermal resistance of the contact point when the temperature rise under rated current just reaches the safety limit.

[0089] When the score is below 50, the system automatically generates maintenance suggestions. When the score is below 20, an emergency alarm is triggered. The alarm content includes the deviation of the specific location from the current thermal resistance estimate and the baseline value.

[0090] The temperature residual is defined as the difference between the measured temperature and the expected temperature calculated based on R_base and the current current. A positive and continuously increasing residual indicates that the actual thermal resistance of the contact point is higher than the reference value, indicating contact degradation. If the temperature residual continuously exceeds the third threshold and the duration exceeds the fourth threshold, a contact degradation warning is triggered. The third threshold is determined by calculating the additional temperature rise caused by a 10% increase in R_base under rated current as the lower limit for triggering the warning. The default value of the third threshold is 8℃. The default value of the fourth threshold is 10 minutes to filter out short-term temperature fluctuations caused by current transients. Its value is based on no less than twice the thermal time constant (heat capacity multiplied by the reference thermal resistance) of the contact point. When R_est increases monotonically over a long period and the daily average rate of increase exceeds the fifth threshold, a contact resistance degradation alarm is generated, prompting maintenance personnel to promptly arrange for joint tightening or cleaning. The default value of the fifth threshold is 0.5% of R_base per day, corresponding to a degradation process of approximately 200 days from good condition to safety limit, which can be used to plan preventive maintenance in advance. This diagnostic mechanism based on thermal resistance estimation residuals eliminates the blind spot in the traditional absolute temperature threshold method for judging the additional temperature rise caused by natural temperature rise under high current conditions that shields contact degradation.

[0091] The zero-sequence current multi-cause spectral decomposition performs a Fast Fourier Transform on the zero-sequence current waveforms of each circuit output by S2, using the power frequency period as a window, to extract the fundamental component I0_50, the third harmonic component I0_150, and the high-frequency broadband energy I0_hf above 150Hz. Parallel classification and diagnosis are achieved by correlating the amplitude variation trends of each frequency band component with the load operating conditions output by S3 on the time axis. For I0_50 (fundamental zero-sequence component), its amplitude change time is compared with the single-phase load activation time detected by S3: if the rise of I0_50 occurs before or after the single-phase load activation time and returns to near zero after the load is disconnected, it is determined to be a normal operating condition record of three-phase imbalance, and no leakage alarm is triggered; if I0_50 persists and exceeds the sixth threshold while all loads are in a three-phase symmetrical state, it is determined to be a potential grounding leakage path and triggers a leakage fault warning, with the alarm content specifying the specific circuit involved. The sixth threshold is the secondary-side converted current value corresponding to the zero-sequence CT of each outgoing circuit under the rated leakage current of the selected protection level. The conversion method is to divide the rated leakage current of the protection level (30mA for Class A protection and 300mA for Class B protection according to GB14287) by the transformation ratio of the selected zero-sequence CT (the ratio of the primary-side rated current to the secondary-side rated current, based on the value marked on the CT nameplate). Taking a typical zero-sequence CT with a transformation ratio of 100:1 and a Class A protection level as an example, the default value of the sixth threshold is 0.3mA (secondary side of the zero-sequence CT). If I0_150 (the 3rd harmonic zero-sequence component) is still significant when all nonlinear loads are in an unloaded state as shown in S3, it may be due to the 3rd harmonic resonance between the compensation capacitor bank and the line inductance, triggering a resonance risk alarm, which is provided for maintenance personnel to check the compensation capacitor parameters.

[0092] Multidimensional cross-validation merges the two types of diagnostic results mentioned above along a time axis and integrates them with the compensation operation records of S4 for comprehensive judgment. If the residual temperature of a specific contact point abnormally increases during a certain period, and a pulsed increment related to the circuit where the contact point is located is observed in I0_hf (high-frequency broadband zero-sequence component), the intermittent micro-arcs generated by contact loosening when passing a large current will conduct high-frequency components through the zero-sequence path. In this case, the two types of independent evidence are merged to determine that there is a contact loosening fault at the point with higher confidence, generating a contact loosening fault alarm with the highest priority. The alarm content includes the location, diagnostic basis, and the current health score of the contact point. If the power factor is consistently lower than the target threshold and the compensation device has reached the output saturation boundary, and the load structure detected by S3 deviates significantly from the historical records, a compensation capacity configuration verification reminder is triggered. It is recommended that maintenance personnel review whether the total current connected load exceeds the design capacity limit of the compensation device.

[0093] The final diagnostic results include health scores for each contact point, categorized abnormal alarms (contact degradation / leakage fault / resonance risk / insufficient compensation capacity), S4 compensation operation summary, and all dimensions of equipment operation parameters, all recorded in chronological order in a structured operation log. The log data structure supports retrieval by time period, by location, and by fault type, providing traceable data for equipment lifecycle management and post-fault review.

[0094] In one embodiment of the present invention, the power distribution scenario of a welding production workshop in the field of automotive parts manufacturing is applied. Such workshops are usually equipped with multi-station welding equipment and other auxiliary power equipment, and are one of the industrial power distribution scenarios with the most concentrated instantaneous impact reactive load density.

[0095] Taking a car door frame welding production line as a specific application scenario: the workshop's low-voltage switchgear is a dedicated low-voltage distribution cabinet for a 630kVA transformer. The cabinet houses four three-phase resistance spot welding machines (stations 1 to 4, each with a different rated welding power), two arc welding robots, three cooling fans, and one lighting circuit. The workstations are distributed throughout the workshop according to the process flow, sharing the same 380V low-voltage busbar. The following data comes from a simulation model of this power distribution scenario based on the MATLAB / Simulink Power System Toolbox. The simulation model is configured according to the actual equipment nameplate parameters, and the default value for the VSI compensation device response delay Δτ is 18ms. Tables 1 and 2 select seven consecutive power frequency cycles that include the simultaneous operation of the welding machine starting arc and the starting of a cooling fan. Serial numbers 4 to 7 represent the continuous records of the fan's acceleration phase during startup. The complete startup process of the fan lasts approximately 400ms (about 20 power frequency cycles). This table only extracts the first seven cycles to demonstrate the dynamic response characteristics of the system under the most severe concurrent transient conditions. The power factor stability of the actual system under non-extreme conditions will be better than the results shown in this table.

[0096] According to the method of this invention, steps S1 to S5 sequentially perform full-process processing on the aforementioned power distribution system. Taking simulation data from a representative time period (including a typical concurrent operating condition of multiple welding machines operating simultaneously and wind turbines starting up) as an example, Table 1 shows the multi-dimensional sensor data for this time period:

[0097] Table 1. Examples of synchronous data acquisition from multi-dimensional sensors (representative time period of 7 consecutive power frequency cycles).

[0098]

[0099] After the load electrical fingerprinting process of S3, the load type and operating status of each branch have been identified, and the reactive power contribution of each type is determined by the measured reactive power of the branch. After soft allocation, press The summation of all cabinet values ​​is shown in Table 2; the compensation control output corresponding to S4 is shown in Table 2. The symbols in the table are consistent with those in the text. , , These are respectively for welding machines, electric motors (including cooling fans), and frequency converters. Resistive loads such as lighting have reactive power that is approximately zero after soft distribution and are not listed separately. This is the total reactive power of the three types of compensable loads. The target value for reactive power compensation in the current power frequency cycle of S4 (see above) , and and the relationship between line residual superposition; the steady-state period is usually slightly higher than Typical margin is approximately 1 to 2 kvar. This represents the actual reactive power compensation amount output by the voltage source inverter (a positive value indicates capacitive compensation). This represents the measured power factor on the bus side during the steady-state period. Approaching If during a transient concurrent cycle Higher than ,but Lagging behind actual demand, busbar A brief decline occurred:

[0100] Table 2. Results of reactive power decomposition by load type and examples of compensation control output (unit: kvar)

[0101]

[0102] In the above data example, numbers 3 and 4 correspond to two spot welding machines entering the peak fusion welding stage. (From 78.5 kvar to 124.3 kvar) and started simultaneously with one cooling fan. The concurrent operating condition (starting from 13.4kvar and increasing) during this cycle kvar. The reactive power gradually increases from 34.7 kvar to 56.4 kvar between serial numbers 4 and 7, reflecting the law that the slip decreases and the reactive power converges from transient to steady state during the induction motor startup process. At serial number 4, the reactive power increase during fan startup is faster than that identified by the S3 operating condition drive. The power frequency cycle is refreshed. kvar is lower than ,although kvar is already higher than It still cannot be fully covered in a short period of time. busbar The value dropped to 0.978; serial numbers 5 to 7 were gradually corrected as the S3 operating condition status was updated and the feedforward-feedback mechanism was implemented. , and Reaching consensus again The voltage recovered to 0.991 to 0.993 within approximately 60 ms (3 power frequency cycles), verifying the dynamic response characteristics of the feedforward-feedback coordinated control under extreme concurrent transient conditions. Simultaneously, the thermal resistance identification results of S5 showed that the estimated equivalent thermal resistance values ​​of connectors 1 and 2 were near the baseline values ​​without significant drift, the health score H remained consistently above 90, and the changes in the fundamental component of the zero-sequence current in each circuit highly matched the timing of the single-phase load connection event detected by S3. There were no independently existing leakage current anomalies, indicating that the overall equipment was operating well.

[0103] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for intelligent integrated monitoring and adaptive compensation of low-voltage switchgear, characterized in that, Includes the following steps: S1, through the multi-type sensor array of the switch cabinet, synchronously collects three-phase voltage, current, temperature, switch status and zero-sequence current, and forms a multi-dimensional synchronous raw time series dataset through multi-channel synchronous alignment; S2 performs weighted sliding filtering, threshold verification, switch debouncing and alignment calibration on the multidimensional synchronous raw time series dataset to generate standardized operating data, including clean high-frequency current waveform sequence, bus three-phase voltage waveform sequence, temperature sequence at each point and zero-sequence current waveform sequence of each circuit. S3 extracts electrical fingerprints for each power frequency cycle from standardized operating data, and identifies them through the encoder attention classification head to obtain the load type identifier and operating condition status identifier of each branch. S4, based on the operating status indicators of each branch and the clean high-frequency current waveform sequence, predicts through long short-term memory network, senses arc slope and event-triggered negative sequence compensation, drives the voltage source inverter for stepless compensation through feedforward and proportional-integral feedback, and outputs the compensation operation record. S5 associates the temperature sequence of each point, the zero-sequence current waveform sequence of each circuit with the operating condition status identifier and compensation operation record of each branch, and generates a health score, categorized abnormal alarm and operation log for each contact point through recursive least squares thermal resistance identification, zero-sequence spectrum decomposition and cross-validation.

2. The intelligent integrated monitoring and adaptive compensation method for low-voltage switchgear according to claim 1, characterized in that, In S1, a through-core broadband current transformer is independently installed for each outgoing circuit, a capacitive voltage divider is installed on each phase of the three-phase busbar, a temperature sensor is independently installed at the contact points of the busbar joint, the circuit breaker in-line and out-line terminals and the cable crimping terminals, and a dedicated zero-sequence current transformer is installed for each outgoing circuit. Each acquisition channel shares the same clock reference, and after being timestamped and synchronized with multiple channels, a multi-dimensional synchronous original time series dataset is formed.

3. The intelligent integrated monitoring and adaptive compensation method for low-voltage switchgear according to claim 1, characterized in that, In S2, a weighted moving average filter is used for the three-phase current and voltage waveform sequences. The weights of each sampling point within the time window are assigned according to the exponential decay law, and the time span of the window is constrained to retain the reactive pulse timing characteristics. Set the upper limit of the current range to the product of the rated current and the first threshold. Sampling points that exceed the range are marked as abnormal and removed from the data stream. A delay confirmation mechanism is adopted for the switch status signal. The duration of the new status must exceed the second threshold to be confirmed as a switch action event. The second threshold is determined according to the upper limit of the rated bounce time of the selected circuit breaker. After completing the weighted moving average filtering, current threshold verification and switch status delay confirmation, the standardized operating data is output after sensor system error compensation.

4. The intelligent integrated monitoring and adaptive compensation method for low-voltage switchgear according to claim 1, characterized in that, In S3, each power frequency cycle is used as the analysis window. Short-time Fourier transform is performed on the clean high-frequency current waveform sequence of each outgoing circuit to extract the current amplitude of each odd harmonic and the phase angle relative to the bus voltage reference. The reactive power contribution ratio of each harmonic component is calculated by using the sine value of the phase difference between each harmonic component and the corresponding voltage component as the weight, and a harmonic reactive power distribution vector is formed. Extract the maximum instantaneous rate of rise of the current waveform and the area and aspect ratio of the instantaneous current-voltage trajectory within the current window; concatenate the harmonic reactive power distribution vector, the maximum instantaneous rate of rise of the current, and the trajectory shape parameters to form a multidimensional electrical fingerprint feature vector for the current power frequency cycle.

5. The intelligent integrated monitoring and adaptive compensation method for low-voltage switchgear according to claim 1, characterized in that, In S3, the working condition recognition network adopts a three-segment structure of encoder, attention, and classification head: the encoder consists of two fully connected layers connected in series, which compress the multidimensional electrical fingerprint feature vector into intermediate representation and latent representation in sequence; the attention segment realizes feature recombination through the inner product normalization of three sets of linear transformations of query, key and value; the classification head consists of two parallel fully connected layers that output the load type probability vector and the operating condition state probability vector respectively. The load types include welding machine, motor, frequency converter, resistive load and unknown type; Semi-supervised training is adopted: tagged samples of isolated single-machine operation and untagged mixed operation samples of normal production are collected. The nameplate type and operating condition record are used to supervise the classification head. The joint constraint of the measured reactive power of each branch and the probability of load type is written into the training objective, so that the reactive power measured by the branch current transformer and the bus voltage participates in the network parameter update. Within each power frequency cycle, for all outgoing branches that are synchronously collected with the corresponding cycle, the measured reactive power of each branch is allocated to welding machines, motors, frequency converters, and resistive loads according to the probability of each branch's load type. Unknown types are not included in the allocation. The allocation amount must be within the reasonable reactive power range determined by the branch wiring log, equipment nameplate, branch active power, and operating conditions. Any amount exceeding the range is included in the training penalty. The same probability-based allocation and full cabinet branch accumulation method is used for online identification to obtain the total reactive power of various loads.

6. The intelligent integrated monitoring and adaptive compensation method for low-voltage switchgear according to claim 1, characterized in that, In S4, for automatic spot welding machine type loads, the long short-term memory network is driven by the operating condition status identifier sequence of each welding machine branch as input to predict the operating condition status of each branch in the next power frequency cycle. When the prediction results show that a certain branch will switch from no-load to active welding state, a feedforward compensation command is immediately issued in the current power frequency cycle. After the response delay of the compensation device, the compensation is completed synchronously with the arrival time of the predicted reactive pulse. For arc welding robot loads, early slope characteristics are extracted from the current rise waveform during the initial short-circuit stage of arc ignition. When the difference between the measured slope and the corresponding welding machine arc ignition characteristic slope reference value in the fingerprint database falls within the preset matching window, a feedforward compensation command is immediately issued according to the typical peak reactive power amplitude of the welding machine in the fingerprint database.

7. The intelligent integrated monitoring and adaptive compensation method for low-voltage switchgear according to claim 1, characterized in that, In S4, when S3 detects a single-phase load activation event, it reads the wiring phase and typical reactive power amplitude of the corresponding load from the fingerprint database, skips the negative sequence phase-locked loop tracking process, and directly generates a negative sequence compensation current reference based on a predefined phase and amplitude. The phase-locked loop is activated for fine phase calibration only after a single-phase device has been running continuously for more than a preset number of cycles. The feedback correction uses the difference between the real-time reactive power measurement value and the target value as the error. The feedback component is output by the proportional-integral regulator and superimposed with the feedforward command. Then, the voltage source inverter is driven by the pulse width modulation command to continuously and adjustablely output the compensation current and output the compensation operation record.

8. The intelligent integrated monitoring and adaptive compensation method for low-voltage switchgear according to claim 1, characterized in that, In S5, for each contact point temperature monitoring point, the measured temperature is used as the observed value, and the current and ambient temperature are used as inputs. The equivalent thermal resistance estimate is updated online using the recursive least squares method. During the stage when the welding machine is continuously idle for longer than the preset stable time, priority is given to participating in the update. After convergence during the initial observation period, the stable estimate is used as the reference thermal resistance value. The degradation progress is defined as the ratio of the current estimated thermal resistance exceeding the baseline thermal resistance to the total degradation margin of the contact point from the baseline thermal resistance to the safety limit thermal resistance. A health score ranging from 0 to 100 is generated based on the degradation progress. When the degradation progress is zero, the score is full and the corresponding contact point is in good condition. When the degradation progress reaches the upper limit of the total degradation margin, the score is zero and the corresponding current thermal resistance has reached the safety limit thermal resistance. A contact degradation warning is triggered when the temperature residual continuously exceeds the third threshold and the duration exceeds the fourth threshold. A continuous degradation alarm is triggered when the daily average increase in thermal resistance exceeds the fifth threshold.

9. The intelligent integrated monitoring and adaptive compensation method for low-voltage switchgear according to claim 1, characterized in that, In S5, a fast Fourier transform is performed on the zero-sequence current waveform of each circuit with the power frequency period as the window to extract the fundamental zero-sequence component, the third harmonic zero-sequence component and the high-frequency broadband zero-sequence component. The amplitude change time of the fundamental zero-sequence component is associated with the activation time of the single-phase load detected by S3. If the fundamental zero-sequence component continues to exceed the sixth threshold under the three-phase symmetrical load condition, a leakage fault warning is triggered. The sixth threshold is the converted value of the secondary side of the zero-sequence current transformer under the rated leakage current of the selected protection level. If the residual temperature at a certain contact point rises abnormally while the high-frequency broadband zero-sequence component of the circuit shows a pulsed increase, a contact loosening fault alarm will be generated with the highest priority. The alarm content includes the location and the current health score.

10. A low-voltage switchgear intelligent integrated monitoring and adaptive compensation system, used to execute the steps in the low-voltage switchgear intelligent integrated monitoring and adaptive compensation method as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to synchronously acquire three-phase voltage, current, temperature, switch status and zero-sequence current through a multi-type sensor array of the switch cabinet, and form a multi-dimensional synchronous raw time series dataset through multi-channel synchronous alignment; The data preprocessing module is used to perform weighted sliding filtering, threshold verification, switch debouncing and alignment calibration on the multidimensional synchronous raw time series dataset to generate standardized operating data, including clean high-frequency current waveform sequence, bus three-phase voltage waveform sequence, temperature sequence of each point and zero-sequence current waveform sequence of each circuit. The operating condition identification module is used to extract electrical fingerprints for each power frequency cycle from standardized operating data, and then identify them through the encoder attention classification head to obtain the load type identifier and operating condition status identifier of each branch. The adaptive compensation module is used to predict the operating conditions of each branch and the clean high-frequency current waveform sequence through long short-term memory network prediction, arc initiation slope sensing and event-triggered negative sequence compensation, feedforward and proportional-integral feedback to drive the voltage source inverter stepless compensation, and output the compensation operation record. The comprehensive diagnostic module is used to associate the temperature sequence of each point, the zero-sequence current waveform sequence of each circuit with the operating condition status identifier and compensation operation record of each branch. Through recursive least squares thermal resistance identification, zero-sequence spectrum decomposition and cross-validation, it generates health scores, categorized abnormal alarms and operation logs for each contact point.

Images