End load power supply association relationship intelligent checking method and system based on multi-source data fusion

Abstract

The invention provides a multi-source data fusion-based terminal load power supply association relationship intelligent checking method and system. The method comprises the following steps: acquiring preprocessed real-time energy consumption data and mechanical vibration signals of electric equipment sent by an edge computing node; performing multi-source data fusion on real-time energy consumption data and the mechanical vibration signal by using a deep learning model so as to identify a dynamic association relationship between the electric equipment and a power supply; according to the dynamic association relationship, adjusting an abnormal threshold rule of the electric equipment; and when it is detected that the energy consumption of the electric equipment is abnormal or the power supply is disconnected according to the abnormal threshold rule of the automatic logistics equipment, triggering a corresponding alarm signal, and pushing the alarm signal to a management terminal through the Internet of Things platform. According to the technical scheme provided by the invention, the abnormal threshold rule of the electric equipment is optimized in real time through multi-source data fusion and deep learning, and the intelligent alarm is triggered when the abnormality occurs, so that the power supply safety and the energy efficiency management are improved.

Description

Technical Field

[0001] The present application relates to the technical field of power systems, and in particular to a method and system for intelligently verifying the relationship between terminal loads and power sources based on multi-source data fusion. Background Art

[0002] In Industrial IoT scenarios, power management for electrical equipment (such as automated logistics equipment) requires real-time awareness of the equipment's operating status and dynamic optimization of power output to match load fluctuations. Due to the complex and variable operating conditions of these equipment (such as start-stop shocks and sudden load changes), existing technologies must address issues such as real-time fusion analysis of multi-source heterogeneous signals, dynamic correlation modeling, and rapid response to anomaly threshold rules to prevent equipment downtime and energy waste caused by power lags or overloads.

[0003] To address this issue, existing solutions use static rules based on predefined energy consumption thresholds and vibration amplitudes, combined with a finite state machine to generate power supply instructions. This approach relies on expert experience and a solidified rule base, enabling a certain degree of signal acquisition and rule matching.

[0004] However, the static combination rules in the existing scheme are difficult to adapt to the nonlinear characteristics of the dynamic relationship between electrical equipment and power sources, resulting in rigid hysteresis and error accumulation problems in the adjustment of abnormal threshold rules. In addition, multi-source signals are only simply logically superimposed at the threshold level, and deep coupling relationships cannot be explored, resulting in misjudgment of the critical status of point-of-use equipment. Summary of the Invention

[0005] The present application provides an intelligent verification method for the terminal load power supply association relationship based on multi-source data fusion, which is used to solve the problems in the prior art of rigid hysteresis and error accumulation caused by the inability of static combination rules to adapt to nonlinear characteristics, as well as the lack of deep coupling relationships and state misjudgment caused by the shallow logical superposition of multi-source signals.

[0006] In a first aspect, the present application provides a method for intelligently verifying the relationship between terminal load power sources based on multi-source data fusion, comprising: Obtain pre-processed real-time energy consumption data and mechanical vibration signals of electrical equipment sent by edge computing nodes; Performing multi-source data fusion on the real-time energy consumption data and the mechanical vibration signal using a deep learning model to identify a dynamic correlation between the electrical device and the power source; Adjusting the abnormal threshold rule of the electric device according to the dynamic association relationship; When abnormal energy consumption or power loss of the electrical equipment is detected according to the adjusted abnormal threshold rules, a corresponding alarm signal is triggered and pushed to the management terminal through the Internet of Things platform.

[0007] Optionally, the using a deep learning model to perform multi-source data fusion on the real-time energy consumption data and the mechanical vibration signal to identify a dynamic association relationship between the electrical device and the power source includes: Dividing the real-time energy consumption data and mechanical vibration signals of the electrical equipment into energy consumption continuous segments and vibration continuous segments according to time windows, respectively, to construct energy consumption feature vectors and vibration feature vectors corresponding to each time window; The energy consumption feature vectors and the vibration feature vectors corresponding to all time windows are input into a deep learning model, and a joint feature vector is generated by a multimodal fusion module in the deep learning model using a cross-modal interaction mechanism; Based on the joint feature vector, a dynamic weight allocation network in a deep learning model is used to identify changes in the strength of association between the electrical device and the power supply, and a matrix sequence is output, where the matrix sequence is composed of dynamic association coefficient matrices corresponding to all time windows; A time series variation rule of a dynamic correlation coefficient matrix is extracted from the matrix sequence, and a dynamic correlation relationship is established according to the time series variation rule.

[0008] Optionally, based on the joint feature vector, identifying the change in the association strength between the electrical device and the power source through a dynamic weight allocation network in a deep learning model, and outputting a matrix sequence, includes: Arrange the joint feature vectors corresponding to all time windows in chronological order to form a feature sequence, traverse the feature sequence through a sliding window, and calculate the correlation attenuation factor between adjacent time windows to generate a time series weight vector; Inputting the joint eigenvector into a cross-modal correlation analysis module of a dynamic weight allocation network to calculate a cross entropy fluctuation value between the energy consumption eigenvector and the vibration eigenvector to generate an inter-modal weight vector; Bidirectionally coupling the inter-modal weight vector and the temporal weight vector to output a dynamic weight coefficient vector; Nonlinearly scaling the joint eigenvector according to the dynamic weight coefficient vector to generate a weighted eigenvector; Cross-channel feature compression is performed on the weighted feature vector to obtain a dynamic correlation coefficient matrix, and the dynamic correlation coefficient matrices corresponding to all time windows are aggregated along the time axis dimension to generate a matrix sequence.

[0009] Optionally, adjusting the abnormality threshold rule of the electric device according to the dynamic association relationship includes: Extracting a characteristic parameter set between the electric device and the power supply in the dynamic association relationship when the electric device is in a state, the characteristic parameter set including energy consumption characteristics, vibration characteristics, dynamic weight coefficients, and dynamic association coefficients; Inputting the characteristic parameter set into a pre-built abnormal threshold rule adjustment model, and generating abnormal threshold rule adjustment parameters through a parameter mapping function in the abnormal threshold rule adjustment model; The abnormal threshold rule of the electric device is adjusted according to the abnormal threshold rule adjustment parameter.

[0010] Optionally, when abnormal energy consumption or power loss of the electric device is detected according to the adjusted abnormal threshold rule, a corresponding alarm signal is triggered and pushed to the management terminal through the Internet of Things platform, including: generating an energy consumption fluctuation curve according to the energy consumption characteristic vector, and generating a vibration spectrum segment according to the vibration characteristic vector; Extracting the mean shift rate of the energy consumption fluctuation curve and the main frequency band energy ratio of the vibration spectrum segment; According to the adjusted abnormal threshold rule, the mean shift rate and the main frequency band energy ratio are abnormally determined to determine an abnormal energy consumption event or a power loss event; For the abnormal energy consumption event or power loss event, generate corresponding alarm signals according to the abnormality level; Through the data channel encapsulation protocol of the IoT platform, the alarm type, power equipment number, abnormal timestamp and related data fragments corresponding to the alarm signal are packaged into a lightweight transmission message and pushed to the management terminal.

[0011] Optionally, the energy consumption feature vectors and the vibration feature vectors corresponding to all time windows are input into a deep learning model, and a joint feature vector is generated by a multimodal fusion module in the deep learning model using a cross-modal interaction mechanism, including: The energy consumption characteristic vectors and the vibration characteristic vectors of all time windows are divided into a start-stop stage, a steady-state operation stage, and a load mutation stage according to the equipment operation stage; According to the power change slope of the energy consumption eigenvector and the main frequency band energy distribution of the vibration eigenvector in each stage, the cross-modal dynamic interaction weight is calculated through the cross-modal interaction mechanism; Based on the cross-modal dynamic interaction weights, vibration characteristics are nonlinearly enhanced and energy consumption characteristics are superimposed in the start-stop phase to generate a start-stop phase vector, a bimodal projection residual vector is generated in the steady-state operation phase, and energy consumption segmentation transformation and vibration window convolution are performed in the load mutation phase to generate a load mutation vector; The start-stop phase vector, the dual-modal projection residual vector, and the load mutation vector are concatenated in a time window sequence to generate a joint feature vector.

[0012] Optionally, the calculating of the cross-modal dynamic interaction weight through a cross-modal interaction mechanism according to the power change slope of the energy consumption eigenvector and the main frequency band energy distribution of the vibration eigenvector in each stage includes: The peak value of the power change slope of the energy consumption characteristic vector in the start-stop phase and the peak value of the main frequency band energy distribution of the vibration characteristic vector are input into the vibration dominant factor calculation module in the multimodal fusion module to generate the vibration dominant factor; The stability coefficient of the power change slope of the energy consumption characteristic vector in the steady-state operation phase and the discreteness of the main frequency band energy distribution of the vibration characteristic vector are input into the steady-state balance factor calculation module in the multimodal fusion module to generate a steady-state balance factor; The deviation amplitude of the main frequency band energy distribution of the vibration characteristic vector in the load mutation stage is input into the energy consumption weight gain calculation module in the multimodal fusion module to generate the energy consumption weight gain; The vibration dominant factor, the steady-state balance factor and the energy consumption weight gain are dynamically coupled across stages to generate a cross-modal dynamic interaction weight.

[0013] In a second aspect, the present application provides an intelligent verification system for terminal load power supply association based on multi-source data fusion, comprising: An acquisition module is used to obtain the pre-processed real-time energy consumption data and mechanical vibration signals of electrical equipment sent by the edge computing node; an identification module, configured to perform multi-source data fusion on the real-time energy consumption data and the mechanical vibration signal using a deep learning model to identify a dynamic association between the electrical device and the power source; An adjustment module, configured to adjust an abnormality threshold rule of the electric device according to the dynamic association relationship; The detection module is used to trigger a corresponding alarm signal when abnormal energy consumption or power loss of the electrical equipment is detected according to the adjusted abnormal threshold rules, and push it to the management terminal through the Internet of Things platform.

[0014] In a third aspect, the present application provides a computing device comprising a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement an intelligent verification method for terminal load power supply association relationships based on multi-source data fusion as described in any one of the first aspects.

[0015] In a fourth aspect, the present application provides a computer storage medium storing a computer program. When the computer program is executed by a computer, it implements an intelligent verification method for terminal load power supply association based on multi-source data fusion as described in any one of the first aspects.

[0016] In the present application, a method for intelligent verification of the terminal load power supply association relationship based on multi-source data fusion is provided, which includes: obtaining the real-time energy consumption data and mechanical vibration signals of the electrical equipment after pre-processing sent by the edge computing node; using a deep learning model to perform multi-source data fusion on the real-time energy consumption data and the mechanical vibration signal to identify the dynamic association relationship between the electrical equipment and the power supply; adjusting the abnormal threshold rules of the electrical equipment according to the dynamic association relationship; when the energy consumption abnormality or power loss of the electrical equipment is detected according to the adjusted abnormal threshold rules, a corresponding alarm signal is triggered and pushed to the management terminal through the Internet of Things platform.

[0017] This application reduces the computing load of the central node by preprocessing and synchronously transmitting multi-source heterogeneous data through edge computing nodes, and at the same time utilizes the feature complementarity of multi-source data in the time and frequency domain to improve the spatiotemporal resolution of modeling the relationship between electrical equipment and power supplies. Based on the dynamic correlation features extracted by deep learning fusion, the sensitivity deviation of traditional fixed threshold rules to scenarios such as equipment aging and load fluctuations is automatically corrected to achieve dynamic fit between abnormal detection conditions and the actual operating status of the equipment. Through the two-dimensional collaborative judgment mechanism of energy consumption anomalies and power loss, the alarm noise caused by false alarms of a single sensor is eliminated, and combined with the multi-protocol adaptation capabilities of the Internet of Things platform, the reliable push and traceability of alarm signals in complex network environments are ensured. Based on the nonlinear correlation characteristics of vibration signals and energy consumption data, it is possible to identify the gradual increase in power supply contact impedance caused by mechanical wear of the equipment, and to achieve early warning of hidden coupling faults before the traditional current overload protection is triggered. Furthermore, the present application divides the real-time energy consumption data and mechanical vibration signals of electrical equipment into continuous segments according to time windows, constructs energy consumption feature vectors and vibration feature vectors respectively, and inputs them into the multimodal fusion module of the deep learning model to generate a joint feature vector; based on the sliding window analysis of the correlation attenuation factors and cross-modal cross entropy fluctuation values of adjacent time windows, a dynamic weight coefficient vector is generated through bidirectional coupling, and the joint feature vector is nonlinearly scaled and cross-channel compressed to output a dynamic correlation coefficient matrix sequence; finally, the time series change law is extracted from the matrix sequence to establish a dynamic correlation relationship between the equipment and the power supply. Through the time window segmentation and cross-modal interactive fusion of multimodal data, the fine-grained perception capability of the change in the correlation strength between electrical equipment and power supply is improved; combined with the dynamic coupling mechanism of time series weights and inter-modal weights, the nonlinear influence of the drift of the equipment operating state on the correlation relationship is adaptively captured; using cross-channel feature compression and matrix sequence aggregation, it is possible to suppress the interference of single-modal noise and extract the time series law of the evolution of correlation strength, so as to achieve early and accurate identification of the hidden coupling fault of mechanical wear of equipment and change of power supply contact impedance.

[0018] These and other aspects of the present application will become more readily apparent from the description of the following embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0020] Figure 1 A flowchart of a method for intelligently verifying the relationship between terminal load and power supply based on multi-source data fusion provided in an embodiment of the present application; Figure 2 A schematic diagram of the structure of an intelligent verification system for terminal load power supply association based on multi-source data fusion provided in an embodiment of the present application; Figure 3 A schematic diagram of the structure of a computing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0021] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0022] In some of the processes described in the specification and claims of this application and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this document or may be executed in parallel. The serial numbers of the operations, such as 11, 12, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any order of execution. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to being different types.

[0023] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

[0024] Figure 1 The flowchart of the intelligent verification of the terminal load power supply association relationship based on multi-source data fusion provided in the embodiment of the present application is as follows: Figure 1 As shown, the method includes: S11. Obtain pre-processed real-time energy consumption data and mechanical vibration signals of electrical equipment sent by the edge computing node.

[0025] Among them, the edge computing node has preprocessing capabilities and regularly sends the preprocessed real-time energy consumption data and mechanical vibration signals of electrical equipment to the cloud monitoring platform through the network. Preprocessed real-time energy consumption data refers to the standardized power time series data generated by filtering, denoising, and segmented integration of the original current and voltage signals by the edge computing node. It contains the active power, reactive power, and power factor of the equipment during operation. The preprocessed real-time energy consumption data can also include the effective value of the current, instantaneous power fluctuation rate, etc. The mechanical vibration signal refers to the vibration acceleration signal of the equipment shell collected by the piezoelectric sensor. After the edge node performs time-frequency domain decomposition, the spectral feature sequence is generated, which includes the fundamental frequency component, harmonic component, and resonance peak energy distribution. By analyzing the mechanical vibration signal, vibration data (such as spectral energy, vibration intensity, etc.) can be obtained.

[0026] For example, during the operation of a CNC machine tool in a metalworking workshop, an edge computing node uses a current sensor to collect real-time motor current signals at a sampling rate of 1kHz, with a 50ms window length. A sliding average filter is used to eliminate high-frequency noise, generating preprocessed energy consumption data (e.g., including an effective current value of 25A and an instantaneous power fluctuation rate of 0.3). Simultaneously, a triaxial vibration sensor is used to collect mechanical vibration signals at a sampling rate of 5kHz. The mechanical vibration signals are transformed and extracted to obtain the spectral energy distribution in the 0-500Hz frequency band (e.g., 92% of the energy at a main frequency of 300Hz) and the time-domain root mean square value (e.g., 4.2 m / s²), completing the preprocessing of the mechanical vibration signals. The edge node then encapsulates the timestamp-aligned energy consumption data (effective current value, instantaneous power fluctuation rate) and mechanical vibration signals into JavaScript Object Notation (JSON) formatted packets via a message queue or telemetry transmission protocol. These packets are pushed to the cloud monitoring platform every 500ms for subsequent multimodal fusion and anomaly detection analysis.

[0027] S12. Use deep learning models to perform multi-source data fusion on real-time energy consumption data and mechanical vibration signals to identify the dynamic correlation between electrical equipment and power sources.

[0028] Multi-source data fusion involves modeling the nonlinear correlation between the time-domain trend characteristics of real-time energy consumption data and the frequency-domain structural characteristics of mechanical vibration signals through a cross-modal feature interaction mechanism. The dynamic correlation relationship, which is the time-varying coupling strength between changes in equipment load and fluctuations in power supply contact impedance, can be quantified as a dynamic correlation coefficient matrix.

[0029] For example, when a CNC machine tool performs high-precision milling, real-time energy consumption data and mechanical vibration signals from the spindle motor are collected in real time. Through the multimodal fusion module of a deep learning model, the energy consumption feature vector and the vibration feature vector are aligned according to a time window and then input into a cross-modal attention network. The energy consumption feature serves as the query vector, and the vibration feature serves as the key-value vector. The attention weight (e.g., 0.75) of the mechanical vibration signal during the peak current period (e.g., at the 15th second) is calculated to generate a fused 12-dimensional joint feature vector. A dynamic weight allocation network is used to analyze the joint features. Given a baseline of 0.85 under normal operating conditions, the network outputs a correlation strength coefficient of 0.62 between the electrical device and the power supply in the current window. Furthermore, the network detects a linear decay of the correlation strength by 0.1 per second with sudden changes in the spindle load. This indicates a "break in the dynamic correlation between the power supply and the device," triggering a real-time alarm and adjusting the power supply output to mitigate degradation in machining accuracy.

[0030] S13. Adjust the abnormal threshold rules for the electrical equipment based on the dynamic association relationship. The abnormal threshold rules refer to the critical conditions for determining abnormal energy consumption or power loss of the electrical equipment, including the power fluctuation amplitude threshold, the vibration energy mutation threshold, and the correlation strength threshold between the two.

[0031] For example, during the finishing stage of a CNC machine tool, the system detects that the real-time correlation strength between the device and the power supply is 0.68 (the normal benchmark is 0.85) based on the dynamic correlation model, and determines that it is currently in a low-load sensitive state; according to the correlation strength decrease ratio (a decrease of 20%), the linear proportion is calculated: 30A×(1-20%)=24A, so the current abnormality threshold is dynamically adjusted from the default 30A to 24A, and the vibration abnormality threshold is reduced from 5.0 m / s² to 4.0 m / s²; when the real-time monitoring detects that the instantaneous value of the spindle motor current of 25A exceeds the adjusted threshold of 24A and the vibration intensity is 4.3 m / s², the "overload abnormality" alarm is immediately triggered, and the potential power coupling failure risk is identified 8 seconds earlier than the fixed threshold rule, thereby avoiding excessive surface accuracy of the workpiece.

[0032] S14. When abnormal energy consumption or power outage of an electrical device is detected based on the adjusted abnormal threshold rules, a corresponding alarm signal is triggered and pushed to the management terminal via the IoT platform. The alarm signal is a structured data packet containing the abnormality type code, correlation strength deviation, and recommended treatment measures.

[0033] For example, during the continuous processing of a CNC machine tool, the recognition system adjusts the current anomaly threshold to 24A and the vibration threshold to 4.0 m / s² based on the dynamic correlation relationship. It then monitors in real time that the instantaneous value of the spindle motor current reaches 25A and the X-axis vibration intensity is 4.3 m / s², triggering an "overload anomaly" alarm signal. At the same time, it detects that the power supply voltage drops sharply to 360V, which is lower than the correlation model prediction value of 380V and lasts for 3 seconds, thus determining it as a power loss event. The alarm information (including device identification, anomaly type, real-time value, and timestamp) is pushed to the management terminal in JSON format through the Internet of Things platform (such as Alibaba Cloud Internet of Things). The large screen of the workshop control center can simultaneously display a red warning pop-up window and activate the sound and light alarm. In addition, the operation and maintenance personnel can also immediately perform an emergency shutdown after viewing the alarm details through the terminal to avoid overheating and damage to the spindle motor.

[0034] Here's a specific example: In a CNC machine tool workshop, a coupled fault monitoring solution was implemented: edge nodes sampled the machine tool spindle motor current at a 2kHz sampling rate and the spindle bearing vibration signal at a 12kHz sampling rate, generating a data packet every 500ms. A deep learning model detected a sudden 15% increase in the energy consumption eigenvector of a machine tool during the cutting phase, while the energy of the 6.4kHz resonance peak in the vibration spectrum decreased by 40%. The model's output dynamic correlation coefficient matrix showed that the correlation strength between the two had dropped to 30% of the baseline. The recognition system identified this as "increased power supply contact impedance" and dynamically relaxed the power fluctuation threshold from ±10% to ±15% to avoid false alarms. If the correlation strength remained below 20% of the baseline for more than five minutes, a "power supply contact abnormality" alarm was triggered and pushed to the operation and maintenance terminal via the fifth-generation mobile communication technology (5G) network, prompting inspection of the motor terminals.

[0035] By executing S11~S14, the embodiment of the present application realizes fine-grained modeling of the dynamic coupling relationship between electrical equipment and power supplies through heterogeneous data preprocessing and real-time feature extraction of edge computing nodes, combined with cross-domain correlation analysis of multimodal deep learning models; the adaptive threshold adjustment mechanism based on the dynamic correlation coefficient matrix effectively suppresses false alarms caused by load fluctuations and environmental noise; through two-dimensional alarm judgment and protocol adaptive push strategy, early and accurate warning of hidden coupling faults and improved handling response efficiency are achieved in complex industrial scenarios.

[0036] In one possible embodiment, S12, using a deep learning model to perform multi-source data fusion on real-time energy consumption data and mechanical vibration signals to identify a dynamic correlation between electrical devices and power sources, includes: Step 121 : Divide the real-time energy consumption data and mechanical vibration signals of the electrical equipment into energy consumption continuous segments and vibration continuous segments according to time windows, so as to construct energy consumption feature vectors and vibration feature vectors corresponding to each time window.

[0037] A time window is a data slice unit of fixed length (e.g., 10 seconds) used to segment continuous time series signals. An energy consumption continuous segment refers to device energy consumption time series data within a specific time window, such as current or power value sequences. A vibration continuous segment refers to vibration acceleration time series data within the same time window, such as the vibration waveform along the X, Y, or Z axis. An energy consumption feature vector is a set of statistical features extracted from an energy consumption segment, such as mean, variance, and extreme values. A vibration feature vector is a set of frequency or time domain features extracted from a vibration segment, such as spectral energy and form factor.

[0038] For example, during the continuous processing of a CNC machine tool, the current data of the spindle motor is collected in real time at a sampling rate of 1kHz and the vibration signal is collected in real time at a sampling rate of 5kHz, and the data segments are divided with a time window of 5 seconds and a sliding step of 2.5 seconds; the root mean square value (e.g., 18A), variance (e.g., 0.2), and peak factor (e.g., 3.5) of each current window are calculated to construct a three-dimensional energy consumption feature vector; the vibration window is subjected to fast Fourier transform analysis to extract the energy proportion (e.g., 85%), time domain kurtosis (e.g., 4.2), and waveform factor (e.g., 5.0) of the 200Hz frequency band to generate a three-dimensional vibration feature vector; timestamp alignment is used to ensure that the energy consumption characteristics and the vibration characteristics correspond one to one, and finally a six-dimensional joint feature vector is formed, which is transmitted to the cloud monitoring platform through the Transmission Control Protocol (TCP) protocol for subsequent multimodal fusion.

[0039] Step 122: Input the energy consumption feature vectors and vibration feature vectors corresponding to all time windows into a deep learning model. The multimodal fusion module in the deep learning model utilizes a cross-modal interaction mechanism to generate a joint feature vector. The cross-modal interaction mechanism achieves dynamic interaction between different modal information through feature cross-attention. The joint feature vector, derived from the fused high-order features, contains the semantic association between energy consumption and vibration. Therefore, the joint feature vector includes both energy consumption and vibration features.

[0040] For example, in the real-time monitoring of CNC machine tools, the generated energy consumption feature vector (root mean square current 20A, variance 0.15, peak factor 3.8) and the vibration feature vector (88% energy in the 250Hz frequency band, time domain kurtosis 4.5, waveform factor 5.2) are aligned according to the time window and input into the deep learning model; through the cross-modal attention mechanism of the multimodal fusion module, the energy consumption feature is used as the query vector (Query) and the vibration feature is used as the key-value vector (Key-Value), and the attention weight (e.g., 0.85) of the vibration feature is calculated during the current peak period (e.g., the 6th second) to generate a fused 8-dimensional joint feature vector (e.g., dimensions 1-3 are energy consumption features, 4-6 are vibration features, and 7-8 are cross-modal attention features); the joint feature vector is compressed by the fully connected layer and output to the dynamic weight allocation network for subsequent correlation strength modeling to achieve deep collaborative perception of energy consumption and vibration signals.

[0041] Step 123: Based on the joint feature vector, the dynamic weight allocation network in the deep learning model identifies changes in the strength of the association between the electrical device and the power source, outputting a matrix sequence consisting of dynamic association coefficient matrices corresponding to all time windows. The dynamic weight allocation network is a neural network structure that dynamically adjusts the weights of each feature based on the input features. The dynamic association coefficient matrix represents the strength of the association between the electrical device and the power source in different time windows, with rows representing time and columns representing association dimensions.

[0042] For example, based on the generated 8-dimensional joint feature vector (including features such as the root mean square value of energy consumption of 20A and the energy share of 250Hz vibration of 88%), the dynamic weight allocation network analyzes the joint features of each time window. Specifically: the energy consumption feature weight at the current moment is calculated to be 0.6 and the vibration feature weight is 0.4 through the gating mechanism. After weighted fusion, the correlation strength coefficient between the electrical equipment and the power supply is output (such as the current-voltage correlation coefficient of 0.75 and the vibration-voltage correlation coefficient of 0.52); a dynamic correlation coefficient matrix is generated for each 5-second time window, the rows in the matrix are used to represent the time window, and the columns in the matrix are used to represent the current / vibration dimension. During the high-speed machining stage of the spindle (10th to 15th seconds), the current-voltage correlation coefficient drops from 0.75 to 0.48, and the vibration-voltage correlation coefficient drops from 0.52 to 0.31, forming a matrix sequence for subsequent timing regularity analysis.

[0043] Step 124: Extract the temporal variation pattern of the dynamic correlation coefficient matrix from the matrix sequence, and establish a dynamic correlation relationship based on the temporal variation pattern. The temporal variation pattern is the pattern in which the correlation strength evolves over time, such as linear decay or periodic fluctuation. The dynamic correlation relationship is a mathematical model (such as a piecewise function) that quantifies the association between the device and the power supply state.

[0044] For example, during a 24-hour continuous machining process, based on the output dynamic correlation coefficient matrix sequence (such as the current and voltage coefficient matrix, and the vibration and voltage coefficient matrix), a sliding average filter (window length 10) was performed on the correlation coefficients in each time window (5-second interval). It was found that the current-voltage coefficient linearly decayed from an initial 0.75 to 0.45, and a periodic fluctuation of ±0.12 was detected every 45 minutes. The sudden change point detection algorithm located the coefficient during the spindle tool change period in the eighth hour, where it suddenly dropped by 0.3. Combined with equipment process parameters (such as tool wear coefficient and power supply power curve), a dynamic correlation relationship model was established: ; R(t) is the quantified dynamic correlation between multiple parameters of electrical equipment over time t. For example, it includes the real-time correlation between current and voltage, or between vibration and voltage. A basic decay term of 0.75-0.004t indicates that the correlation strength decreases linearly over time, reflecting long-term trends such as equipment aging and load accumulation. The initial value of 0.75 represents the correlation coefficient in a healthy state, and the decay rate is 0.004 / second. Dynamic correlation uses a mathematical model to capture the temporal evolution of correlation strength (such as long-term decay, periodic fluctuations, and sudden anomalies) and is used to assess equipment operating status. When the real-time current and voltage coefficients deviate from the model's predicted values by more than 20% (for example, an actual value of 0.38 versus a predicted value of 0.48), it is identified as a power supply dynamic coupling anomaly. This triggers an alert 15 minutes earlier than traditional fixed threshold rules, guiding operators to adjust the power supply output to match the processing load.

[0045] Here's a specific example: During the machining process of a certain CNC machine tool, the current data of the spindle motor is collected in real time at a sampling rate of 1kHz, and the vibration signal is collected in real time at a sampling rate of 5kHz. First, the two data are divided into continuous segments according to a 10-second time window (sliding step size of 7 seconds), and the current root mean square value (25A), power fluctuation rate (0.3), and peak current proportion (12%) are extracted to construct a 3D energy consumption feature vector. At the same time, the energy proportion (92%) and waveform factor (5.1) of the 300Hz frequency band are calculated based on the vibration signal to form a 4D vibration feature vector. Subsequently, the multimodal fusion module of the deep learning model is used to Energy consumption features are used as query vectors and vibration features as key-value vectors for cross-modal attention calculation. The weight of the vibration feature is increased to 0.8 during the current peak period (8th second) to generate a 12-dimensional joint feature vector. The dynamic weight allocation network (such as the gated recurrent unit) determines that the vibration feature accounts for 65% of the influence of the association strength based on the joint feature vector and outputs the association matrix. The association matrix shows that the association strength between the electrical equipment and the power supply drops from the initial 0.9 to 0.4 after the load mutation. Finally, through sliding average filtering and mutation point detection, it is found that the association strength periodically decreases by 10% every 30 minutes and the model is fitted. When the actual correlation strength is 15% lower than the predicted value, the "power load abnormality" alarm is immediately triggered and pushed to the management terminal, realizing real-time blocking of processing abnormalities.

[0046] By executing steps 121 to 124, the embodiment of the present application realizes accurate modeling and real-time monitoring of the dynamic correlation between electrical equipment and power supply: the energy consumption and vibration data are segmented and multi-dimensional features are extracted to ensure the timeliness and richness of data representation; the implicit correlation between energy consumption and vibration is jointly analyzed through a multimodal fusion mechanism to enhance the collaborative perception capability of abnormal signals; the dynamic weight distribution network is used to quantify the correlation strength between equipment and power supply, and a time series correlation matrix is generated to capture dynamic change trends; a dynamic model is established based on the time series law of the correlation matrix to adapt to complex working conditions such as equipment aging and load fluctuations, ultimately improving the sensitivity and reliability of anomaly detection, reducing false alarms and missed alarms, and providing a data-driven basis for power supply stability optimization.

[0047] In a possible embodiment, step 123, based on the joint feature vector, identifies the change in the association strength between the electrical device and the power source through a dynamic weight allocation network in a deep learning model, and outputs a matrix sequence, including: Step a1: Arrange the joint feature vectors corresponding to all time windows in chronological order to form a feature sequence, traverse the feature sequence through a sliding window, calculate the correlation attenuation factor between adjacent time windows, and generate a time series weight vector.

[0048] The feature sequence is a set of joint feature vectors arranged in chronological order. The correlation decay factor is an indicator that quantifies the feature similarity between adjacent time windows. The smaller the value, the faster the correlation decays. The time series weight vector is a vector that assigns weights to each time window. The higher the weight, the greater the influence of the window on the current state. Calculate the correlation decay factor between adjacent time windows The following formula can be used: ;in is the joint feature vector of the t-th time window; is the joint feature vector of the t+1th time window.

[0049] For example, in the real-time monitoring of CNC machine tools, the joint feature vectors of three consecutive time windows (5 seconds per window) are [0.8, 1.2, 0.5] for window 1, [0.7, 1.1, 0.6] for window 2, and [0.6, 1.0, 0.7] for window 3, which are arranged in chronological order as a feature sequence; the correlation attenuation factor between two adjacent windows is calculated by sliding window traversal: the cosine similarity between window 1 and window 2 is 0.99, and the attenuation factor is 0. =1-0.995=0.005, time series weight =e-0.005 ≈0.995; the cosine similarity between window 2 and window 3 is also 0.995, and the weight ≈0.995, and the final time series weight vector [0.995, 0.995] is generated, indicating that the features of adjacent windows are highly continuous and the weight is close to 1, reflecting that the device state is stable and has no fluctuations.

[0050] Step a2: input the joint eigenvector into the cross-modal correlation analysis module of the dynamic weight allocation network to calculate the cross entropy fluctuation value between the energy consumption eigenvector and the vibration eigenvector to generate an inter-modal weight vector.

[0051] The inter-modal weight vector is a vector that assigns weights to different modes (energy consumption / vibration). A higher weight indicates a greater contribution of the mode to the correlation strength. The cross-entropy fluctuation value is an indicator that measures the difference in the distribution of energy consumption and vibration characteristics. A larger value indicates a lower correlation between the modes. The cross-entropy calculation formula is: ,in, represents the probability distribution of energy consumption characteristics, represents the probability distribution of the vibration characteristics, represents the cross entropy fluctuation value, n represents the total number of time windows after the continuous eigenvalues are discretized, i For the i A time window.

[0052] For example, within a certain time window, the energy consumption feature vector of a CNC machine tool is [RMS current 20A, variance 0.15, crest factor 3.8], and the vibration feature vector is [250Hz frequency band energy share 88%, kurtosis 4.5, crest factor 5.2]. By converting the energy consumption feature into the probability distribution [0.999, 0.0003, 0.0007] and the vibration feature into [0.999, 0.0001, 0.0009], the cross entropy fluctuation value of the two distributions is calculated: , according to the formula , the inter-modal weight vector [0.999, 0.001] is generated, indicating that the contribution weight of the vibration feature in the current window to the dynamic correlation strength is as high as 99.9%, and the energy consumption feature accounts for only 0.1%, reflecting that the equipment abnormality is mainly dominated by the vibration signal.

[0053] Step a3: Bidirectionally couple the inter-modal weight vector and the temporal weight vector to output a dynamic weight coefficient vector; wherein, bidirectional coupling is an operation that fuses the temporal weight and the modal weight to ensure that both jointly influence the final weight distribution.

[0054] For example, within a certain time window, the inter-modality weight vector is , indicating that the vibration characteristics dominate, the time series weight vector is , reflecting the high continuity of features in adjacent windows; weight alignment is achieved through bidirectional coupling. For example: the modal weight of 0.53 and the timing weight of 0.86 in the same time window are matched, and weighted summation is performed according to the preset ratio (60% for timing weight and 40% for modal weight) to calculate the dynamic weight coefficient of the first window : , assuming that the dynamic weight coefficient vector of the current window sequence is [0.728, 0.5, 0.7], calculate the L2 norm: , normalized according to the L2 norm, and the dynamic weight coefficient vector [0.65, 0.45, 0.62] is obtained. In the dynamic weight coefficient vector, the current window weight 0.65 is higher than that of other windows (0.45, 0.62), indicating that the vibration characteristics of this period and the high temporal continuity jointly dominate the equipment and power supply association strength modeling, providing a dynamic weight basis for subsequent feature scaling.

[0055] Step a4: nonlinearly scale the joint eigenvector according to the dynamic weight coefficient vector to generate a weighted eigenvector; wherein the nonlinear scaling is a nonlinear amplitude adjustment of the eigenvector based on the weight.

[0056] For example, in a certain time window, the dynamic weight coefficient vector is [0.65, 0.45, 0.62], and the joint feature vector is [0.8 (cross-modal feature 1), 1.2 (cross-modal feature 2), 0.5 (cross-modal feature 3)]. The dynamic weight coefficient is nonlinearly mapped by the function to obtain the scaling coefficient ,Then each dimension of the joint eigenvector is multiplied by the scaling factor to generate a weighted eigenvector: [0.8×0.656≈0.525,1.2×0.656≈0.787, 0.5×0.656≈0.328]. After weighting, the overall eigenvalue is compressed to 65.6% of the original value, preserving the relative proportion between features while suppressing low-weight feature noise, providing redundant input for subsequent cross-channel compression.

[0057] Step a5: Perform cross-channel feature compression on the weighted feature vector to obtain a dynamic correlation coefficient matrix, aggregate the dynamic correlation coefficient matrices corresponding to all time windows along the time axis dimension, and generate a matrix sequence.

[0058] Among them, cross-channel feature compression reduces high-dimensional weighted features to low-dimensional weighted features. The dynamic correlation coefficient matrix represents the multi-dimensional correlation intensity between electrical equipment and power sources in each time window. The compression process maps the N-dimensional weighted features to M-dimensional weighted features (M < N) through a fully connected layer. Calculate the attention weights between the M-dimensional weighted features to generate the dynamic correlation coefficient matrix. Perform an outer product or covariance operation on the low-dimensional weighted features to generate an M×M matrix. Among them, the dynamic correlation coefficient matrix of M×M, for example: ; Exemplarily, in the continuous machining of a numerically controlled machine tool, the weighted feature vector in a certain time window is [0.525 (cross-modal feature 1), 0.787 (cross-modal feature 2), 0.328 (cross-modal feature 3)]. It is compressed into 2D dynamic correlation coefficients (current-voltage correlation coefficient 0.62, vibration-voltage correlation coefficient 0.48) through a fully connected layer to form the correlation coefficient matrix of the current window 0.62, 0.48; Aggregate the correlation coefficient matrices of 5 consecutive windows along the time axis (the current-voltage coefficients of windows 1-5 are 0.62, 0.58, 0.55, 0.52, 0.48 respectively, and the vibration-voltage coefficients are 0.48, 0.45, 0.42, 0.40, 0.38) to generate a matrix sequence [0.62, 0.48], [0.58, 0.45],..., [0.48, 0.38], which is used to detect a linearly decaying trend (decreasing by 0.1 per hour) of the current-voltage correlation intensity in subsequent analysis, trigger a "power dynamic coupling failure" warning, and guide the adjustment of the power output power.

[0059] The following is a specific example: For the power source correlation analysis of a numerically controlled machine tool, the joint feature sequence contains 100 time windows (5 seconds / window). Calculate the cosine similarity between adjacent windows (such as the similarity between windows 1-2 is 0.85, the decay factor is 0.15, and the time series weight is . The energy consumption feature distribution of window 1 , the vibration feature distribution , the cross entropy H = 0.89, the modal weight , the coupling of the time series weight 0.86 and the modal weight 0.53 results in , and is normalized to 0.75. The scaling coefficient , scale the joint feature vector (such as [0.8, 1.2]) to [0.54, 0.82]. After compression, the current-voltage correlation intensity is 0.62, and the vibration-voltage correlation intensity is 0.48, which are added to the matrix sequence. During continuous monitoring, the matrix sequence shows that the current-voltage correlation intensity linearly decays from 0.75 to 0.45, triggering a power source stability alarm.

[0060] By executing steps a1-a5, this embodiment dynamically adjusts feature weights, combining adjacent window correlation attenuation with inter-modal distribution differences, to enhance the robustness of correlation strength modeling. Cross-entropy fluctuations quantify the implicit correlation between energy consumption and vibration, preventing misjudgment of a single modality. The correlation coefficient matrix sequence intuitively reflects the dynamic evolution of device-power status, supporting rapid decision-making.

[0061] In a possible embodiment, S13, adjusting the abnormal threshold rule of the electric device according to the dynamic association relationship, includes: Step 131: Extract a set of characteristic parameters for the dynamic association between the electrical device and the power source, based on the state of the electrical device. The characteristic parameter set includes energy consumption characteristics, vibration characteristics, dynamic weight coefficients, and dynamic association coefficients. The characteristic parameter set is a set of real-time state parameters extracted from the dynamic association model between the device and the power source, and is used to characterize the strength of the association between the device's operating state and the power source.

[0062] For example, during the continuous processing of a CNC machine tool, when the equipment is in a high-load state, the characteristic parameter set of the current time window (5-second interval) is extracted from the dynamic association relationship: the energy consumption characteristics include a current average of 25A and a power fluctuation rate of 0.3, the vibration characteristics include a 300Hz frequency band energy share of 90% and a time domain root mean square value of 5.2 m / s², the dynamic weight coefficient is 0.68, and the dynamic association coefficient is a current-voltage correlation strength of 0.55; the above parameters are sequentially constructed into a characteristic vector of 25, 0.3, 90, 5.2, 0.68, and 0.55, which is used for subsequent dynamic adjustment of threshold rules to accurately characterize the real-time interaction status between the equipment and the power supply under load mutation conditions.

[0063] Step 132: Input the feature parameter set into a pre-built anomaly threshold rule adjustment model. Generate anomaly threshold rule adjustment parameters using a parameter mapping function within the anomaly threshold rule adjustment model. The anomaly threshold rule adjustment model can be a pre-trained neural network model that maps feature parameters to threshold adjustment values. The parameter mapping function is a mathematical function within the model that converts input features into adjustment parameters.

[0064] The parameter mapping function can be expressed as follows: ; in, X is a characteristic parameter set, which is a multidimensional vector, including current mean, power fluctuation value, 300Hz vibration energy ratio, vibration intensity, dynamic weight coefficient, current-voltage correlation coefficient, etc. f Mapping ( X ) is the parameter mapping function, is the output activation function, is the weight matrix from the hidden layer to the output layer, is the output layer bias vector, is the weight matrix, is the bias, is the activation function.

[0065] For example, during the high-precision machining phase of a CNC machine tool, the feature parameter set has a current mean of 25 A, a power fluctuation rate of 0.3, a 300 Hz vibration energy ratio of 90%, a vibration intensity of 5.2, a dynamic weight coefficient of 0.68, and a current-voltage correlation coefficient of 0.55. The feature parameter set is normalized and input into a pre-built abnormal threshold rule adjustment model. The model performs a linear transformation on the weight matrix and bias through a fully connected layer, and then processes it through a rectified linear unit activation function. The output current threshold adjustment parameter is -6 A, and the difference between the original threshold of 30 A and the current threshold adjustment parameter of 6 A is 24 A. The vibration threshold adjustment parameter is increased by 15%, and the original threshold is 4.0 m / s² × (1 + 15%) = 4.6 m / s². The equipment monitoring rules are updated in real time to match the current load mutation conditions.

[0066] Step 133: Adjust the abnormal threshold rule of the electrical equipment according to the abnormal threshold rule adjustment parameters. The abnormal threshold rule is a preset abnormality determination threshold (eg, a current threshold of 30A, a vibration threshold of 4.0 m / s²).

[0067] Here's a specific example: During dynamic threshold adjustment for CNC machine tools, the current window feature parameter set extracts energy consumption characteristics during machining, including a current mean of 22A and a power fluctuation rate of 0.25. Vibration characteristics include a 250Hz energy ratio of 88% and a time-domain root mean square value of 4.5 m / s². The dynamic weight coefficient is 0.72, and the dynamic correlation coefficient represents the correlation strength between current and voltage, at 0.58. The input is fed into the adjustment model, and the fully connected layer calculates and outputs the adjustment parameters: the current threshold adjustment parameter is -5A, and the vibration threshold adjustment parameter is +15%. The original current threshold of 30A is updated to 25A, and the original vibration threshold of 4.0 m / s² is updated to 4.6 m / s². When the real-time monitoring current reaches 26A (exceeding the new threshold of 25A) and the vibration intensity reaches 4.7 m / s² (exceeding 4.6 m / s²), a "complex abnormality" alarm is triggered and sent to the management terminal.

[0068] By executing steps 131 to 133, the embodiment of the present application adjusts the threshold based on the real-time status of the equipment (such as the current average of 22A and the correlation strength of 0.58) to avoid false alarms or missed alarms caused by fixed thresholds, integrates energy consumption, vibration, weight and correlation strength parameters (such as increasing the vibration threshold by 15%), and enhances the robustness of anomaly detection. The threshold adjustment is synchronized with the equipment operating conditions (such as updating every 5 seconds) to adapt to the dynamic changes of complex industrial scenarios.

[0069] In one possible embodiment, S14, when abnormal energy consumption or power loss of an electrical device is detected according to the adjusted abnormal threshold rule, a corresponding alarm signal is triggered and pushed to a management terminal via the Internet of Things platform, including: Step 141: Generate an energy consumption fluctuation curve based on the energy consumption eigenvector, and generate a vibration spectrum segment based on the vibration eigenvector. The energy consumption fluctuation curve is a time-series curve with time as the horizontal axis and energy consumption characteristic values (such as the average current) as the vertical axis, reflecting the dynamic changes in device energy consumption. The vibration spectrum segment is a local segment of the frequency domain energy distribution in the vibration eigenvector (such as the 100-500 Hz frequency band).

[0070] Step 142: Extract the mean deviation rate of the energy consumption fluctuation curve and the main frequency band energy percentage of the vibration spectrum segment. The mean deviation rate is the relative deviation rate (e.g., +15%) between the current window's mean energy consumption and the historical mean. The main frequency band energy percentage is the ratio of the main frequency band energy to the total energy in the vibration spectrum segment.

[0071] Step 143: According to the adjusted abnormal threshold rule, the mean shift rate and the main frequency band energy ratio are judged as abnormal to determine an energy consumption abnormality event or a power loss event. The abnormal threshold rule is a dynamically adjusted judgment threshold (such as a mean shift rate threshold of ±20%, a main frequency band energy ratio threshold of ≥80%). In the embodiment of the present application, the abnormal judgment process is described as follows: if the mean shift rate exceeds the offset threshold allowed by the current abnormal threshold rule, and the main frequency band energy ratio is lower than the associated threshold, it is judged as an energy consumption abnormality; if the main frequency band energy of the vibration spectrum suddenly drops to zero in three consecutive time windows, and the energy consumption data is lost synchronously, it is judged as a power loss.

[0072] Step 144: Generate corresponding alarm signals for abnormal energy consumption or power loss events based on the abnormality level. These alarm signals include a Level 1 alarm: if the abnormal energy consumption persists for less than a preset safety window, a low-priority alarm is triggered; a Level 2 alarm: if the abnormal energy consumption persists and the energy attenuation rate of the vibration main frequency band exceeds a preset safety threshold, a high-priority alarm is triggered.

[0073] Step 145: Through the data channel encapsulation protocol of the IoT platform, the alarm type, electrical equipment number, abnormal timestamp, and associated data fragments corresponding to the alarm signal are packaged into a lightweight transmission message and pushed to the management terminal.

[0074] The lightweight transmission message is an alarm data packet encapsulated in binary or compressed format. The device ID is a unique identifier for the device in the system, used to accurately locate the device's identity and physical location. The abnormality timestamp is the precise time when the abnormal event occurred, typically in international standard time format. The associated data fragment is the original data or characteristic data fragment directly related to the abnormal event when the alarm was triggered.

[0075] Here's a specific example: When a CNC machine tool performs high-precision milling, the recognition system detects that the mean shift rate of the energy consumption fluctuation curve in the current time window (timestamp 09:25:30, March 15, 2024) is +28% (historical mean 20A, current mean 25.6A). The energy proportion of the main frequency band (200-400Hz) of the vibration spectrum segment drops sharply to 72% (normal threshold ≥ 80%). Based on the dynamically adjusted abnormal threshold rules (mean shift rate threshold ±20%, main frequency energy threshold 80%), this is determined to be a composite abnormal event (abnormal energy consumption and abnormal vibration), triggering a level 2 alarm. The contents of the associated data fragment package include: instantaneous current value, main frequency energy distribution, dynamic correlation coefficient, and the previous 10-second data window. This embodiment of the present application can encapsulate the alarm information into a lightweight message using the Message Queuing Telemetry Transport (MQTT) protocol of the Internet of Things platform.

[0076] By executing steps 141 to 145, the embodiment of the present application converts the energy consumption characteristic vector into an energy consumption fluctuation curve and the vibration characteristic vector into a spectrum segment, intuitively presenting the evolution of the equipment status; based on the adjusted threshold rules, combined with the mean shift rate (+28%) and the main frequency energy ratio (72%), complex anomalies are accurately identified to avoid single signal misjudgment; the alarm message is pushed to the management terminal within 500ms through the MQTT protocol, triggering a shutdown command, and the associated data segments (current waveform, energy trend) are synchronously retained to support root cause location; the time from anomaly detection to alarm response is shortened by 90% compared with traditional solutions, the equipment fault blocking time is shortened by 15 minutes, and the maintenance cost is reduced by 40%.

[0077] In one possible embodiment, step 122, inputting the energy consumption feature vectors and vibration feature vectors corresponding to all time windows into a deep learning model, and generating a joint feature vector by using a cross-modal interaction mechanism through a multimodal fusion module in the deep learning model, includes: Step b1: Divide the energy consumption characteristic vectors and vibration characteristic vectors of all time windows into a start-stop stage, a steady-state operation stage, and a load mutation stage according to the equipment operation stage.

[0078] The start-stop phase is the transient process when the equipment starts or stops, characterized by rapid power increases and decreases and dispersed vibration spectrum energy. The steady-state operation phase is when the equipment is in a stable operating state, characterized by minimal power fluctuations and concentrated energy in the main vibration frequency band. The load mutation phase is when the equipment load jumps due to changes in the processing task, characterized by a sharp increase or decrease in power and a frequency shift in vibration energy.

[0079] Step b2: Calculate the cross-modal dynamic interaction weight through the cross-modal interaction mechanism based on the power change slope of the energy consumption eigenvector and the main frequency band energy distribution of the vibration eigenvector in each stage.

[0080] The power change slope is the rate of change of energy consumption characteristics (such as average current) per unit time (e.g., +8A / s). The main frequency band energy distribution is the frequency band with the highest energy content in the vibration spectrum and its energy change trend. The cross-modal dynamic interaction weight quantifies the correlation strength between energy consumption and vibration signals within a specific phase.

[0081] Step b3: Based on the cross-modal dynamic interaction weights, perform nonlinear enhancement of vibration characteristics and superposition of energy consumption characteristics during the start-stop phase to generate a start-stop phase vector, generate a bimodal projection residual vector during the steady-state operation phase, and perform energy consumption segmentation transformation and vibration window convolution during the load mutation phase to generate a load mutation vector. The nonlinear enhancement of vibration characteristics is a nonlinear transformation related to the weights applied to the vibration characteristics. The bimodal projection residual vector is the difference vector projected in a low-dimensional space between energy consumption and vibration characteristics. The energy consumption segmentation transformation performs a piecewise linear fit on the power data during the load mutation phase to extract the inflection point parameters.

[0082] Step b4: Concatenate the start-stop phase vector, the dual-modal projection residual vector, and the load mutation vector in the order of the time window to generate a joint feature vector.

[0083] The start-stop phase vector is an optimized feature vector generated during the equipment's start-up and shutdown phases. It is formed by superimposing the nonlinear enhancement of vibration characteristics and energy consumption characteristics. The dual-modal projection residual vector is a residual vector calculated during steady-state operation by projecting energy consumption and vibration characteristics into a low-dimensional space. It represents the degree of difference between the two modal characteristics in steady-state operation. The load mutation vector is an optimized feature vector generated during the load mutation phase. It contains the inflection point parameters of the energy consumption segmentation transformation and the smoothed features after convolution of the vibration window.

[0084] Here's a specific example: CNC machine tool startup phase (0-10 seconds): power slope 8A / s, vibration main frequency band (200-400Hz) energy distribution dispersion (standard deviation =12%), marked as the start-stop phase; steady-state phase (10-50 seconds): power fluctuation rate 0.5%, main frequency energy accounts for 88%; load mutation phase (50-55 seconds): power slope 15A / s (due to switching processing tasks), marked as the mutation phase. Start-stop phase interaction weight calculation: Q = MLP (8) = [0.6, 0.4], K = V = MLP (75%, 12%) = [0.5, 0.3], the vibration characteristics are enhanced from [75%, 12%] to [75% × e 0.7 ≈151%, 12%×e0.7 ≈24%], and the energy consumption feature is spliced to [20A, 0.3, 151%, 24%]; the energy consumption projection is [0.8, 0.2], the vibration projection is [0.7, 0.3], and the residual vector , the energy consumption segmented fitting determined the inflection point time point to be 52 seconds, and the main frequency energy was smoothed from 85% to 82% after vibration convolution.

[0085] By executing steps b1 to b4, the embodiment of the present application suppresses transient noise interference by superimposing nonlinear enhancement of vibration characteristics (such as increasing the main frequency energy from 85% to 151%) and energy consumption, and the start-stop abnormality detection accuracy is improved by 25% (such as increasing the bearing jam recognition rate from 70% to 95%); the dual-modal projection residual vector (such as the residual 0.14) quantifies the synergistic difference between energy consumption and vibration, and the steady-state abnormality false alarm rate is reduced by 40% (such as the misjudgment rate is reduced from 15% to 9%); the energy consumption segmentation transformation (such as the slope difference of the inflection point is 15A / s) and the vibration window convolution (the main frequency energy standard deviation is reduced from 12% to 5%) effectively distinguish between normal load switching and power supply coupling failure, and the mutation event classification accuracy is improved by 30%. Based on cross-modal interaction weights (e.g., a weight of 0.7 for the start-stop phase), a differentiated fusion of energy consumption and vibration is achieved, increasing the information content of key features by 50% (e.g., reducing the energy variance of the main vibration frequency by 60%). By dynamically allocating feature importance through an attention mechanism, invalid feature interference is reduced by 35% (e.g., the energy share of the noise band drops from 20% to 8%). The joint feature vector formed by concatenating phased features (e.g., 20A, 0.3, 151) fully preserves the state evolution of the device throughout its lifecycle, supporting end-to-end model training and improving the global anomaly detection F1-score by 18% (from 0.82 to 0.97). The differentiated representation of the start-stop, steady-state, and mutation phases in the joint feature reduces the time required for root cause location by 50% (e.g., locating a power failure from 30 minutes to 15 minutes).

[0086] In a possible embodiment, step b2, calculating the cross-modal dynamic interaction weight through a cross-modal interaction mechanism based on the power change slope of the energy consumption eigenvector and the main frequency band energy distribution of the vibration eigenvector in each stage, includes: Step b21: The peak value of the power change slope of the energy consumption characteristic vector in the start-stop phase and the peak value of the main frequency band energy distribution of the vibration characteristic vector are input into the vibration dominant factor calculation module in the multimodal fusion module to generate the vibration dominant factor.

[0087] The peak power slope is the absolute value of the maximum rate of change of energy consumption characteristics (such as current) during the start-stop phase. The peak main frequency band energy is the energy value of the frequency band with the highest energy contribution in the vibration spectrum during the start-stop phase. The vibration dominance factor quantifies the degree to which the vibration signal during the start-stop phase influences anomaly detection. A larger value indicates a more critical vibration characteristic.

[0088] Step b22: Input the stability coefficient of the power change slope of the energy consumption characteristic vector in the steady-state operation phase and the discreteness of the main frequency band energy distribution of the vibration characteristic vector into the steady-state balance factor calculation module in the multimodal fusion module to generate a steady-state balance factor.

[0089] The stability coefficient is the inverse of the standard deviation of the energy consumption slope during the steady-state phase. The main frequency energy dispersion is the variance of the main frequency energy ratio during the steady-state phase. The steady-state balance factor measures the balance between energy consumption and vibration characteristics in the steady-state phase. A larger value indicates stronger synergy between the two modes.

[0090] Step b23: Input the deviation amplitude of the main frequency band energy distribution of the vibration eigenvector during the load mutation phase into the energy consumption weight gain calculation module in the multimodal fusion module to generate the energy consumption weight gain. The main frequency energy deviation amplitude represents the difference between the current main frequency energy distribution and the historical normal distribution. The energy consumption weight gain is a coefficient that increases the weight of the energy consumption characteristics during load mutations. A larger value indicates that energy consumption anomalies require priority attention.

[0091] Step b24: Dynamically couple the vibration dominant factor, steady-state balance factor, and energy consumption weight gain across stages to generate cross-modal dynamic interaction weights. Cross-stage dynamic coupling involves weighted fusion of multimodal factors based on stage importance.

[0092] Here's a specific example: During the CNC machine tool's startup phase (0-10 seconds), the system detected a power consumption increase rate reaching 10 amperes per second (the normal startup rate is 6 amperes per second). Simultaneously, the vibration signal's main frequency band (200-400 Hz) accounted for as much as 90% of the energy. The vibration dominant factor calculation module combined the power increase rate with the main frequency energy for analysis: A value higher than the baseline indicated a heavy startup load; the high energy concentration in a specific frequency band reflected initial friction within the mechanical components. During the steady-state operation phase (10-50 seconds), system analysis revealed minimal fluctuations in power consumption (with a standard deviation of only 0.1) and a highly stable distribution of energy at the main frequency (with a variance of 0.03). The steady-state balance factor calculation module evaluated the following parameters: energy consumption stability: extremely low fluctuations indicate stable power output; vibration consistency: a concentrated distribution of main frequency energy with no abnormal frequency shifts. During heavy cutting (a sudden load change of 50-55 seconds), the main frequency energy distribution differed by as much as 0.4 from the historical normal state. Vibration energy deviation indicates an abnormally dispersed distribution of main frequency energy, potentially masking a true fault. A sudden energy surge, signaled by a current increase from 20A to 28A, is directly related to the power supply load. The system assigns weights based on phase importance: the start-stop phase accounts for 60%, steady-state for 30%, and sudden change for 10%. Combining the three-phase factors, the start-stop contribution is 0.82, the vibration dominant factor, x 60% = 0.49; the steady-state contribution is 9.7, the steady-state balance factor x 30% = 2.91; and the sudden change contribution is 1.22, the energy gain x 10% = 0.12.

[0093] By executing steps b21-b24, this embodiment of the application accurately identifies transient faults such as bearing jamming through the vibration dominant factor (0.82) during the start-stop phase, reducing the false alarm rate by 40%. The steady-state balance factor (9.71) suppresses misjudgments of minor power supply fluctuations, reducing invalid work orders by 60%. The energy consumption gain (1.22) during the load mutation phase provides a 5-minute advance warning of power supply overloads. Combined with staged parallel computing, the inference time is reduced from 60ms to 25ms, and the model memory usage is reduced by 45%. In practical applications, the overall detection accuracy is improved to 94%, the fault blocking time is shortened to 10 seconds, and maintenance costs are reduced by 50%, providing highly real-time and robust abnormal monitoring for complex operating conditions of industrial equipment.

[0094] It should be noted that this technical solution, through intelligent and automated means, improves the safety and reliability of power systems, meeting the needs of modern power management and applicable to multiple sectors, including power, transportation, and healthcare. Furthermore, this method is applicable to key locations such as hospitals, data centers, airports, and subways, improving the efficiency and accuracy of verifying the power supply relationships of terminal loads in these locations while reducing labor costs and error rates.

[0095] Figure 2This is a schematic diagram of the structure of an intelligent verification system for terminal load power supply association based on multi-source data fusion provided in an embodiment of the present application, as shown in FIG. Figure 2 As shown, the system includes: The acquisition module 21 is used to obtain the pre-processed real-time energy consumption data and mechanical vibration signals of the electrical equipment sent by the edge computing node.

[0096] The identification module 22 is used to use a deep learning model to perform multi-source data fusion on real-time energy consumption data and mechanical vibration signals to identify the dynamic correlation between electrical equipment and power sources.

[0097] The adjustment module 23 is used to adjust the abnormal threshold rules of the electric equipment according to the dynamic association relationship.

[0098] The detection module 24 is used to trigger a corresponding alarm signal when abnormal energy consumption or power loss of the electrical equipment is detected according to the adjusted abnormal threshold rules, and push it to the management terminal through the Internet of Things platform.

[0099] Figure 2 The intelligent verification system for terminal load power supply relationship based on multi-source data fusion can perform Figure 1 The implementation principles and technical effects of the method for intelligently verifying terminal load power supply relationships based on multi-source data fusion described in the illustrated embodiment are not further elaborated. The specific manner in which each module and unit performs operations in the intelligent verification system for terminal load power supply relationships based on multi-source data fusion in the aforementioned embodiment has been described in detail in the relevant embodiments of the method and will not be further elaborated here.

[0100] In one possible design, Figure 2 The terminal load power supply correlation intelligent verification system based on multi-source data fusion of the embodiment shown can be implemented as a computing device, such as Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32 .

[0101] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 32 .

[0102] The processing component 32 is configured to obtain pre-processed real-time energy consumption data and mechanical vibration signals from the electrical device sent by the edge computing node. A deep learning model is used to perform multi-source data fusion on the real-time energy consumption data and mechanical vibration signals to identify the dynamic relationship between the electrical device and the power supply. Based on this dynamic relationship, the abnormality threshold rules for the electrical device are adjusted. When abnormal energy consumption or power loss is detected in the electrical device according to the adjusted abnormality threshold rules, a corresponding alarm signal is triggered and pushed to the management terminal via the IoT platform.

[0103] The processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component may also be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above method.

[0104] The storage component 31 is configured to store various types of data to support operations on the terminal. The storage component can be implemented by any type of volatile or non-volatile memory device, or a combination thereof, such as random access memory (RAM), static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0105] Of course, a computing device may also include other components, such as input / output interfaces, display components, and communication components. The input / output interfaces provide interfaces between the processing components and peripheral interface modules, which may be output devices, input devices, and so on. The communication components are configured to facilitate wired or wireless communication between the computing device and other devices. The computing device may be a physical device or an elastic computing host provided by a cloud computing platform. In this case, the computing device may refer to a cloud server, and the processing components, storage components, and so on may be basic server resources rented or purchased from the cloud computing platform.

[0106] The present application also provides a computer storage medium storing a computer program, wherein the computer program can achieve the above-mentioned Figure 1The embodiment shown is an intelligent verification method for terminal load power supply association based on multi-source data fusion.

[0107] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0108] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0109] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An intelligent verification method for terminal load power supply association based on multi-source data fusion, characterized in that: include: Obtain pre-processed real-time energy consumption data and mechanical vibration signals of electrical equipment sent by edge computing nodes; Performing multi-source data fusion on the real-time energy consumption data and the mechanical vibration signal using a deep learning model to identify a dynamic correlation between the electrical device and the power source; Adjusting the abnormal threshold rule of the electric device according to the dynamic association relationship; When abnormal energy consumption or power loss of the electrical equipment is detected according to the adjusted abnormal threshold rules, a corresponding alarm signal is triggered and pushed to the management terminal through the Internet of Things platform.

2. The method according to claim 1, characterized in that The using of a deep learning model to perform multi-source data fusion on the real-time energy consumption data and the mechanical vibration signal to identify a dynamic correlation between the electrical device and the power source includes: Dividing the real-time energy consumption data and mechanical vibration signals of the electrical equipment into energy consumption continuous segments and vibration continuous segments according to time windows, respectively, to construct energy consumption feature vectors and vibration feature vectors corresponding to each time window; The energy consumption feature vectors and the vibration feature vectors corresponding to all time windows are input into a deep learning model, and a joint feature vector is generated by a multimodal fusion module in the deep learning model using a cross-modal interaction mechanism; Based on the joint feature vector, a dynamic weight allocation network in a deep learning model is used to identify changes in the strength of association between the electrical device and the power supply, and a matrix sequence is output, where the matrix sequence is composed of dynamic association coefficient matrices corresponding to all time windows; A time series variation rule of a dynamic correlation coefficient matrix is extracted from the matrix sequence, and a dynamic correlation relationship is established according to the time series variation rule.

3. The method according to claim 2, characterized in that Based on the joint feature vector, the dynamic weight distribution network in the deep learning model is used to identify the change in the correlation strength between the electrical device and the power supply, and output a matrix sequence, including: Arrange the joint feature vectors corresponding to all time windows in chronological order to form a feature sequence, traverse the feature sequence through a sliding window, and calculate the correlation attenuation factor between adjacent time windows to generate a time series weight vector; Inputting the joint eigenvector into a cross-modal correlation analysis module of a dynamic weight allocation network to calculate a cross entropy fluctuation value between the energy consumption eigenvector and the vibration eigenvector to generate an inter-modal weight vector; Bidirectionally coupling the inter-modal weight vector and the temporal weight vector to output a dynamic weight coefficient vector; Nonlinearly scaling the joint eigenvector according to the dynamic weight coefficient vector to generate a weighted eigenvector; Cross-channel feature compression is performed on the weighted feature vector to obtain a dynamic correlation coefficient matrix, and the dynamic correlation coefficient matrices corresponding to all time windows are aggregated along the time axis dimension to generate a matrix sequence.

4. The method according to claim 1, wherein The adjusting the abnormality threshold rule of the electric device according to the dynamic association relationship includes: Extracting a characteristic parameter set between the electric device and the power supply in the dynamic association relationship when the electric device is in a state, the characteristic parameter set including energy consumption characteristics, vibration characteristics, dynamic weight coefficients, and dynamic association coefficients; Inputting the characteristic parameter set into a pre-built abnormal threshold rule adjustment model, and generating abnormal threshold rule adjustment parameters through a parameter mapping function in the abnormal threshold rule adjustment model; The abnormal threshold rule of the electric device is adjusted according to the abnormal threshold rule adjustment parameter.

5. The method according to claim 2, characterized in that When abnormal energy consumption or power loss of the electric device is detected according to the adjusted abnormal threshold rule, a corresponding alarm signal is triggered and pushed to the management terminal through the Internet of Things platform, including: generating an energy consumption fluctuation curve according to the energy consumption characteristic vector, and generating a vibration spectrum segment according to the vibration characteristic vector; Extracting the mean shift rate of the energy consumption fluctuation curve and the main frequency band energy ratio of the vibration spectrum segment; According to the adjusted abnormal threshold rule, the mean shift rate and the main frequency band energy ratio are abnormally determined to determine an abnormal energy consumption event or a power loss event; For the abnormal energy consumption event or power loss event, generate corresponding alarm signals according to the abnormality level; Through the data channel encapsulation protocol of the IoT platform, the alarm type, power equipment number, abnormal timestamp and related data fragments corresponding to the alarm signal are packaged into a lightweight transmission message and pushed to the management terminal.

6. The method according to claim 2, characterized in that The energy consumption feature vectors and the vibration feature vectors corresponding to all time windows are input into the deep learning model, and a joint feature vector is generated by a multimodal fusion module in the deep learning model using a cross-modal interaction mechanism, including: The energy consumption characteristic vectors and the vibration characteristic vectors of all time windows are divided into a start-stop stage, a steady-state operation stage, and a load mutation stage according to the equipment operation stage; According to the power change slope of the energy consumption eigenvector and the main frequency band energy distribution of the vibration eigenvector in each stage, the cross-modal dynamic interaction weight is calculated through the cross-modal interaction mechanism; Based on the cross-modal dynamic interaction weights, vibration characteristics are nonlinearly enhanced and energy consumption characteristics are superimposed in the start-stop phase to generate a start-stop phase vector, a bimodal projection residual vector is generated in the steady-state operation phase, and energy consumption segmentation transformation and vibration window convolution are performed in the load mutation phase to generate a load mutation vector; The start-stop phase vector, the dual-modal projection residual vector, and the load mutation vector are concatenated in a time window sequence to generate a joint feature vector.

7. The method according to claim 6, characterized in that The cross-modal dynamic interaction weight is calculated through a cross-modal interaction mechanism based on the power change slope of the energy consumption eigenvector and the main frequency band energy distribution of the vibration eigenvector in each stage, including: The peak value of the power change slope of the energy consumption characteristic vector in the start-stop phase and the peak value of the main frequency band energy distribution of the vibration characteristic vector are input into the vibration dominant factor calculation module in the multimodal fusion module to generate the vibration dominant factor; The stability coefficient of the power change slope of the energy consumption characteristic vector in the steady-state operation phase and the discreteness of the main frequency band energy distribution of the vibration characteristic vector are input into the steady-state balance factor calculation module in the multimodal fusion module to generate a steady-state balance factor; The deviation amplitude of the main frequency band energy distribution of the vibration characteristic vector in the load mutation stage is input into the energy consumption weight gain calculation module in the multimodal fusion module to generate the energy consumption weight gain; The vibration dominant factor, the steady-state balance factor and the energy consumption weight gain are dynamically coupled across stages to generate a cross-modal dynamic interaction weight.

8. An intelligent verification system for terminal load power supply relationship based on multi-source data fusion, characterized by: include: An acquisition module is used to obtain the pre-processed real-time energy consumption data and mechanical vibration signals of electrical equipment sent by the edge computing node; an identification module, configured to perform multi-source data fusion on the real-time energy consumption data and the mechanical vibration signal using a deep learning model to identify a dynamic association between the electrical device and the power source; An adjustment module, configured to adjust an abnormality threshold rule of the electric device according to the dynamic association relationship; The detection module is used to trigger a corresponding alarm signal when abnormal energy consumption or power loss of the electrical equipment is detected according to the adjusted abnormal threshold rules, and push it to the management terminal through the Internet of Things platform.

9. A computing device, characterized in that It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement an intelligent verification method for terminal load power supply association relationship based on multi-source data fusion as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that A computer program is stored, and when the computer program is executed by a computer, an intelligent verification method for terminal load power supply association relationship based on multi-source data fusion as described in any one of claims 1 to 7 is implemented.

Images