Bus duct joint degradation early warning method based on machine learning and causal inference

Abstract

The application discloses a bus duct joint degradation early warning method based on machine learning and causal inference, in order to solve the problem of indirect estimation under the condition of no direct contact resistance measurement, the application realizes the robust estimation and interval early warning of the contact resistance and the degradation rate under the complex environment and load disturbance, improves the early warning accuracy and reliability by means of the event-driven neural tool variable modeling, the orthogonalization constraint, the multi-channel thermoelectric physical information neural network coded by the energy conservation and the environmental boundary, the contact resistance coupled to the current square heating term, the counterfactual gate generation baseline temperature rise and residual causal decomposition sharing the weight, and the risk score constructed by fusing the model parameters and the extrapolation uncertainty.

Description

Technical Field

[0001] This invention relates to the field of power system equipment condition monitoring, and in particular to a method for early warning of busbar joint deterioration based on machine learning and causal inference. Background Technology

[0002] Busbar trunking, as a high-current power supply channel in low-voltage power distribution systems, is susceptible to increased contact resistance at its joints due to factors such as loose bolts, oxidation corrosion, and contact surface contamination. This leads to localized heating, abnormal temperature rise, and consequently, insulation aging and safety risks. Current online monitoring methods often combine current sensing with infrared temperature sensing, supplemented by environmental parameter acquisition, and provide early warnings through threshold alarms or empirical models. Offline maintenance commonly uses micro-ohmmeters or DC injection methods to measure contact resistance, but these require power outages or changes in operating conditions, making them unsuitable for continuous online assessment. In recent years, thermoelectric digital twins, finite element thermal analysis, thermal circuit models, and data-driven regression and deep learning methods have also been used for temperature rise prediction and health assessment, but they still face challenges in terms of reliability and interpretability under complex field conditions.

[0003] The main shortcomings of existing technologies are:

[0004] 1. It is impossible to achieve indirect and robust identification of contact resistance under conditions of live load and inability to directly measure contact resistance. The temperature-current relationship is easily affected by load step, cyclic load and environmental disturbance, and it is difficult to distinguish the temperature rise caused by contact deterioration and the temperature rise caused by changes in operating conditions.

[0005] 2. Thermal models or data-driven models often rely on correlation fitting and lack causal structure and instrumental variable support. It is difficult to orthogonally decouple contact resistance from environmental boundaries and unmodeled disturbances, leading to parameter drift and misjudgment.

[0006] 3. Insufficient handling of uncertainties: Uncertainties in model parameters, emissivity calibration, and event extrapolation are not systematically quantified, making it difficult for risk scoring and alarm thresholds to be consistent, thus affecting the credibility and availability of early warnings.

[0007] Therefore, a method for early warning of busbar joint deterioration that can overcome the shortcomings of the existing technology is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0008] One objective of this invention is to propose a busbar joint degradation early warning method based on machine learning and causal inference. Addressing the problems of existing technologies, such as the inability to directly measure contact resistance under energized conditions, the difficulty in distinguishing temperature and current due to load and environmental interference, and the lack of uncertainty quantification in early warning, the following technical solution is proposed: Collecting current, surface temperature, and environmental parameters and performing emissivity calibration and time-series alignment; constructing instrumental variables through change points and periodic detection, and training a neural instrumental variable model to obtain current control residuals; constructing a multi-channel thermoelectric physical information neural network, applying energy conservation and boundary encoding, coupling contact resistance only to the current square heating term, and orthogonally injecting the control residuals into the heating and boundary terms, applying slow-varying, monotonic, and maintenance reset constraints to the contact resistance evolution; constructing a counterfactual channel under shared weights to generate baseline temperature rise, and updating contact resistance and degradation rate by combining residual causal decomposition; fusing model parameters and extrapolated uncertainties to form a risk score for early warning. This invention possesses the technical advantages of robust estimation of contact resistance and its degradation rate under conditions without direct measurement, effectively eliminating interference from operating conditions and the environment, providing interval-based risk warnings, and reducing false alarms and missed alarms.

[0009] A method for early warning of busbar joint degradation based on machine learning and causal inference according to an embodiment of the present invention includes:

[0010] S1. Collect the current at the connector, the surface temperature of the connector, and environmental parameters, and perform emissivity calibration and timing alignment.

[0011] S2. Detect time-series changes in current data, identify load step or periodic load events, and construct instrumental variables and form a sequence of instrumental variables accordingly.

[0012] S3. Train a neural instrumental variable model based on the instrumental variable sequence, environmental parameters and current data to characterize the structural influence of instrumental variables and environment on current, calculate current control residuals and obtain control residual sequence;

[0013] S4. Construct and train a multi-channel physical information neural network based on the control residual sequence and aligned data, apply energy conservation constraints and environmental boundary encoding, couple the contact resistance only to the heating term represented by the square of the current data multiplied by the contact resistance, inject the control residual into the heating term and boundary term and apply orthogonalization constraints to the loss, and apply slow variation and monotonic constraints to the time evolution of the contact resistance, and allow it to be reset to the baseline contact resistance when there is a maintenance event marker, thus obtaining the observation channel model;

[0014] S5. Construct a counterfactual channel that shares weights with the observation channel model under the condition of freezing the observation channel model weights. Set the contact resistance to the baseline value through counterfactual gating, and optionally set the environmental parameters to the baseline. Calculate the counterfactual temperature rise to obtain the counterfactual temperature rise sequence.

[0015] S6. The temperature difference sequence is obtained by subtracting the counterfactual temperature rise from the surface temperature. The residual causal decomposition is then performed to separate the residual contributed by the contact resistance and the unmodeled disturbance residual. The former is used to update the contact resistance estimate under slow variation and monotonic constraints, and reset is allowed when the maintenance event marker exists. The latter is subject to sparse and short-term constraints to absorb non-contact factors, thus obtaining the contact resistance estimation sequence and the degradation rate sequence.

[0016] S7. Based on the contact resistance estimation sequence and the degradation rate sequence, combined with the uncertainty of the observation channel model parameters and the extrapolation uncertainty of the instrumental variables, the contact resistance uncertainty interval is estimated, and a risk score is constructed accordingly. The score is compared with the preset threshold, and the degradation warning signal and uncertainty interval are output.

[0017] Optionally, step S1 includes:

[0018] At the busbar joint, current sensors, infrared temperature sensors, and environmental sensors are used to collect current data, joint surface temperature data, and environmental parameter data, respectively, and timestamps are added to the collected data.

[0019] An emissivity calibration coefficient is established by infrared emissivity calibration. The emissivity calibration coefficient is determined by comparing the joint surface temperature data with the reference temperature provided by the calibration plate or contact temperature measuring point. The emissivity calibration coefficient is applied to the joint surface temperature data to correct the radiation error and generate calibrated joint surface temperature data.

[0020] According to the time alignment rules, the current data, connector surface temperature data and environmental parameter data are aligned on the same time axis, sorted by timestamp, and resampled or interpolated for samples with unequal intervals. At the same time, duplicate or out-of-order records are removed to align the data on the same time axis. The environmental parameter data includes at least ambient temperature, relative humidity, wind speed and cabinet door status.

[0021] Output calibrated current data, calibrated connector surface temperature data, and calibrated environmental parameter data.

[0022] Terminology definition:

[0023] The busbar joint is a connection part in which adjacent conductive busbars in the busbar trunk are electrically connected and mechanically fixed by fasteners and conductive gaskets.

[0024] The current at the joint is a time-series quantity of the current flowing through the phase conductor to which the busbar joint belongs. It can be an instantaneous sampled value, an effective value, or an equivalent root mean square value, and is recorded on a unified time axis with a timestamp.

[0025] The current data is a digital sampling sequence of the current at the connector, including timestamps and current amplitude information;

[0026] The joint surface temperature is the temperature of the measured surface area located at the busbar joint, which can be obtained by an infrared temperature sensor and subsequently calculated after emissivity calibration.

[0027] The joint surface temperature data is time-series data of temperature statistics within a set field of view or ROI at the joint by an infrared temperature sensor, preferably a regional average value.

[0028] The reference temperature is the actual temperature value obtained by a calibration plate or contact temperature measuring point at a location that is consistent with or corresponds to the infrared temperature measurement field of view.

[0029] The calibration plate is a standard radiation reference with known emissivity parameters and stable thermal characteristics. It can be attached or placed near the connector for emissivity calibration and comparison.

[0030] The contact temperature measuring point is a temperature measuring point installed at or near the joint using a thermocouple, platinum resistance thermometer, or other contact temperature sensor.

[0031] The time alignment rules are a set of parameters and strategies used to map multi-source data such as current, temperature and environment to a unified time axis, including target sampling period, timestamp sorting, resampling / interpolation methods and strategies for handling duplicate or out-of-order records.

[0032] The unified timeline is a common time series formed by using a unified time base and sampling step size for data from each channel, and is used to carry the synchronous index of multi-source data.

[0033] The calibrated current data is the current timing data obtained after timing alignment and outlier cleaning of the current data.

[0034] The calibrated environmental parameter data is time-series data obtained after time-series alignment and outlier cleaning of the environmental parameter data.

[0035] The cabinet door status refers to the opening status information of the busbar trunking or its associated cabinet door, which can be represented as an open / closed binary value or an opening degree level.

[0036] Optionally, step S2 includes:

[0037] Based on the calibrated current data, time-series change detection is performed on a unified time axis. First, the calibrated current data is denoised to suppress instantaneous spikes and measurement noise. Then, a sliding window is used to calculate the slope and second-order differential to locate the change point. When the current stabilizes at the new level after the change point and reaches the set dwell time and the step amplitude exceeds the preset threshold, the data segment is marked as a load step event and the event start time, end time, event type, step amplitude, and load levels before and after are recorded.

[0038] Simultaneously, autocorrelation analysis or power spectral density analysis is performed on the calibrated current data to identify periodic components. When the main frequency falls into the set frequency band and the energy proportion exceeds the preset threshold and continues to exceed the minimum number of cycles, the data segment is marked as a periodic load event and the event start time, end time, event type, main frequency, amplitude and duration are recorded.

[0039] Construct a sequence of instrumental variables based on the labeled load step events and periodic load events, so that each instrumental variable corresponds to an event record and includes the event timestamp, event type, event intensity parameter and event duration. After removing events with step amplitude below the threshold or insufficient duration, output the sequence of instrumental variables.

[0040] Terminology definition:

[0041] The time-series change detection is a process that identifies significant changes or periodic occurrences in statistical characteristics in calibrated current data on a unified time axis.

[0042] The point of change is the time index at which the current data changes significantly at the statistical level or derived characteristics, and is used as a candidate moment for the start of the event.

[0043] The load step event is the time period during which the current jumps from one stable load level to another stable load level and the residence time at the new level is not less than a set threshold, and the step amplitude exceeds the threshold.

[0044] The preceding and following load levels are representative levels of the current within the stable range before and after the load step event, preferably the mean or median of that range.

[0045] The dwell time is the minimum continuous time length used to determine whether a new load level is stable.

[0046] The step amplitude value is the absolute value of the difference in load level before and after the load step event;

[0047] The periodic load event is defined as a period in which the current data contains a significant periodic component of the main frequency within a set time period, the energy proportion of which exceeds a threshold, and the number of continuous periods is not less than the minimum number of periods.

[0048] The periodic component is an approximate sinusoidal or quasi-periodic component of the current data in the frequency domain or autocorrelation domain.

[0049] The main frequency is the frequency with the highest power or the most concentrated energy within the set frequency band;

[0050] The set frequency band is a range of frequencies of interest that are preset according to the operating conditions of the monitored system;

[0051] The energy percentage is the ratio of the spectral energy within the main frequency and its neighborhood bandwidth to the total spectrum energy.

[0052] The minimum number of cycles is the minimum number of consecutive cycles required to confirm a periodic load event;

[0053] The amplitude is the effective amplitude of the periodic component at the main frequency, which can be obtained by power spectral density or time-domain fitting.

[0054] The duration is the continuous time length of the event from the start time to the end time;

[0055] The event type is information that categorizes detected load events, including at least two types: load step and periodic load.

[0056] The autocorrelation analysis is an analytical method that estimates the autocorrelation function of current data to identify periodicity.

[0057] The power spectral density analysis is an analytical method for estimating the power spectral density of current data to identify periodicity and dominant frequencies.

[0058] The instrumental variables are constructed based on load events and reflect the intensity and timing of exogenous load disturbances. They are used to characterize the structural impact on the current and are designed to be as orthogonal as possible to the contact resistance path.

[0059] The instrumental variable sequence is a time series of instrumental variables arranged along a uniform time axis, representing the changes in the activation and intensity of each event over time;

[0060] The event intensity parameter is a quantity that characterizes the intensity of the event. For a load step event, it includes at least the step amplitude and the load levels before and after the event. For a periodic load event, it includes at least the main frequency and amplitude.

[0061] The event timestamp is a time marker used to locate key moments such as the start or end of an event;

[0062] The preset threshold is a set of parameter thresholds used for event determination and rejection, including at least the step amplitude threshold, energy percentage threshold, and dwell time threshold.

[0063] Optionally, step S3 includes:

[0064] Based on the instrumental variable sequence, calibrated environmental parameter data, and calibrated current data, a neural instrumental variable model is trained to characterize the structural influence of instrumental variables and environmental parameters on current data;

[0065] During training, the loss function is set to include an event correlation term and an orthogonalization constraint. The event correlation term is used to enhance the explanatory power of the instrumental variables on the current data, and the orthogonalization constraint is used to reduce the interference of environmental parameters on the path of the instrumental variables.

[0066] After training is completed, the current data is predicted on a unified time axis using a neural instrumental variable model, and the residual of the current control function is defined as the calibrated current data minus the model-predicted current, generating a sequence of current control function residuals.

[0067] Correlation and orthogonality tests are performed on the residual sequence of the current control function to confirm its independence from the instrumental variable sequence and its orthogonality to the environmental parameters. If the preset test threshold is not met, the weights of the loss function are adjusted and training is repeated until the test threshold is met.

[0068] Output current control function residual sequence.

[0069] Terminology definition:

[0070] The neural instrumental variable model is a structured causal modeling framework that uses a neural network as a function approximator, takes a sequence of instrumental variables and environmental parameters (and optional historical terms) as input, and outputs current data as output. It is used to characterize the structural effects of exogenous loads and the environment on current and generate control residuals.

[0071] The structural influence is a representation of the causal relationship between instrumental variables and environmental parameters on current data through a model, including elements such as influence intensity, time delay, and nonlinear mapping.

[0072] The loss function is the objective function used for optimization when training the neural instrumental variable model, and includes a weighted sum of fitting error terms and constraint / regulation terms;

[0073] The event correlation term is an enhancement term in the loss function, used to improve the model's explanatory power for current changes driven by instrumental variables, reflecting the goal of improving the correlation or mutual information between instrumental variables and predicted current.

[0074] The orthogonalization constraint is the decorrelation / deconfounding constraint in the loss function, which is used to reduce the interference of environmental parameters on the current along the instrument variable path, so that the control residual is approximately independent of the instrument variable and approximately orthogonal to the environmental parameters;

[0075] The instrumental variable path is a mapping pathway from instrumental variable input to current output in the neural instrumental variable model, representing the causal link of the structural influence of instrumental variables on current.

[0076] The model predicts the current as the current prediction value calculated by the neural instrumental variable model on the input data on a unified time axis.

[0077] The current control residual is the difference between the calibrated current data and the model-predicted current, used to characterize the current portion not explained by instrumental variables and environmental parameters.

[0078] The current control function residual is the difference between the calibrated current data and the model-predicted current, and is synonymous with the current control residual.

[0079] The current control function residual sequence is the time series set of the current control function residuals on a unified time axis;

[0080] The correlation test is a statistical test of the correlation between the residuals of the current control function and the instrumental variable series, used to confirm whether the correlation between the two is below a threshold.

[0081] The orthogonality test is a statistical test of the correlation between the residuals of the current control function and the environmental parameters being close to zero, in order to confirm the orthogonality of the residuals with respect to the environmental path.

[0082] The independence refers to the property that two variables or time series are statistically independent of each other, requiring at least that the correlation is not significant and that a preset significance level is met;

[0083] The preset test threshold is a set of threshold parameters used to determine whether the correlation or mutual information meets the independence / orthogonality requirements, including at least the correlation coefficient threshold and the significance level;

[0084] The weights of the loss function are the weighted coefficients corresponding to each component of the loss function (including the event correlation term and the orthogonalization constraint term), which are used to balance the optimization strength of the fitting and the constraint.

[0085] Optionally, step S4 includes:

[0086] Based on the residual sequence of the current control function, the calibrated current data, the calibrated environmental parameter data, and the calibrated joint surface temperature data, a multi-channel physical information neural network is constructed and trained. The network is configured with busbar copper bus channel, contact surface channel, bolt channel, gasket channel, and air boundary layer channel, and each channel is connected by energy conservation constraints. At the same time, the environmental boundary conditions are encoded to map the environmental temperature, relative humidity, wind speed, and cabinet door status into the boundary parameters of each channel.

[0087] In the contact surface channel, the heating term is defined by multiplying the square of the current data by the contact resistance, and the contact resistance is coupled only to this heating term and not directly coupled to other channels.

[0088] The residual sequence of the current control function is injected into the heating term and the boundary term respectively to form orthogonal coupling, and orthogonalization constraints are applied to the loss function to reduce the sensitivity of the contact resistance to the residual of the current control function.

[0089] Slow variation and monotonic constraints are applied to the time evolution of contact resistance, and when a maintenance event is triggered, the contact resistance is allowed to be reset to the baseline contact resistance, which is a pre-set health or post-maintenance state reference value and is not greater than the current contact resistance estimate at the time of maintenance triggering.

[0090] After training is completed by minimizing the difference between the calibrated joint surface temperature data and the network output temperature rise, and the preset convergence condition is met, an observation channel model is generated to characterize the actual observation process.

[0091] Terminology definition:

[0092] The multi-channel physical information neural network is a neural network structure that integrates thermal and electrical physical constraints. It models the thermal paths such as busbars, contact surfaces, bolts, gaskets and air boundary layers in the form of channels, and outputs the temperature rise response at the joint with energy conservation and environmental boundary conditions as constraints.

[0093] The busbar copper bus channel is a network channel that characterizes the internal heat conduction and heat capacity behavior of the busbar copper bus. It is used to transfer the heat generated by the contact surface and the inside of the conductor and to affect the surface temperature distribution.

[0094] The contact surface channel is a network channel that characterizes the conductive contact area of ​​the joint. It contains contact resistance parameters and bears the heating term of the square of the current multiplied by the contact resistance. It is used to describe the local heat source and its heat conduction to adjacent channels.

[0095] The bolt channels are network channels that characterize the heat conduction and heat capacity paths of fasteners, and are used to describe the influence of the heat conduction branches formed by bolt connections on the overall temperature field.

[0096] The gasket channels are network channels that characterize the thermal conductivity and thermal capacity paths of conductive or insulating gaskets, and are used to describe the influence of gasket material and geometry on heat conduction.

[0097] The air boundary layer channel is a network channel that characterizes the convective and radiative heat transfer path between the outer surface of the joint and the environment, and is used to map environmental boundary conditions into external heat dissipation capacity.

[0098] The energy conservation constraint is a physical constraint used to connect the channels, which limits the generation, storage and transfer of heat at any time and at the spatial node to satisfy the energy conservation relationship, thereby ensuring that the temperature rise of the network output is physically consistent.

[0099] The environmental boundary condition encoding is a strategy for mapping environmental parameters to the heat transfer boundary description required for the boundary layer channel, including at least converting environmental temperature, relative humidity, wind speed and cabinet door status into parameters such as convective heat transfer coefficient, radiative heat transfer coefficient and boundary temperature.

[0100] The boundary parameters are a set of parameters used to characterize the external heat transfer boundary, including at least ambient temperature, convective heat transfer coefficient, effective radiation coefficient, and ventilation / shading coefficient related to the cabinet door status.

[0101] The heating term is a local heat source term in the contact surface channel formed by multiplying the square of the current data by the contact resistance, which is used to quantify the Joule heating generated at the electrical contact.

[0102] The coupling is such that the contact resistance parameter only participates in the calculation of the heating term and does not appear directly in the parameters or outputs of other channels, so as to ensure the identifiability of the contact resistance and decoupling from other paths;

[0103] The boundary term is the heat transfer term of the air boundary layer channel to the external environment, including at least convective heat transfer term and radiative heat transfer term and their parameterized representation in the network;

[0104] The orthogonal coupling is a coupling method in which the residual sequence of the current control function is injected into the heating term and the boundary term respectively, and orthogonalization constraints are applied to reduce the sensitivity of the residual to the contact resistance parameter and decouple the contact resistance from the hybrid path.

[0105] The contact resistance time evolution is a contact resistance state sequence updated over time, used to characterize the slow changes and deterioration trend of the contact condition.

[0106] The slow-varying constraint is a constraint that limits the rate of change of the contact resistance over time, so that it does not change drastically when there is no maintenance event.

[0107] The monotonic constraint is a monotonicity requirement imposed on the time evolution of contact resistance, ensuring that it does not decrease or degrade during normal operation, in order to reflect the irreversible physical characteristics of degradation.

[0108] The maintenance event trigger is an external marker of the maintenance behavior, which allows the model to execute special update rules over time when it occurs;

[0109] The network output temperature rise is the increment of the joint surface temperature relative to the ambient temperature, output by the multi-channel physical information neural network on a unified time axis.

[0110] The preset convergence conditions are a set of stopping criteria for the training phase, including at least the validation set loss reaching a threshold or no significant improvement for several consecutive rounds and the physical constraint satisfaction reaching a set level.

[0111] The observation channel model is a network instance obtained after training and meeting the convergence condition, used to represent the actual observation process. Its parameters are fixed for subsequent counterfactual calculations and online inferences.

[0112] Optionally, step S5 includes:

[0113] Based on the observation channel model, a counterfactual channel is constructed without changing the weights of the observation channel model, so that the counterfactual channel shares all weights with the observation channel model;

[0114] Based on the calibrated current data and calibrated environmental parameter data, the contact resistance is set to a preset baseline contact resistance through counterfactual gating, and the environmental parameters are selectively set to preset baseline environmental parameters. This allows the counterfactual channel to calculate the temperature rise response under the baseline contact resistance and the selected environmental conditions on a unified time axis with a time step consistent with the observation channel model, thereby generating a counterfactual temperature rise sequence.

[0115] Terminology definition:

[0116] The counterfactual channel is a computational channel that shares the structure and parameters with the observation channel model without changing the weights of the observation channel model. It is used to generate a comparative temperature rise output under a set baseline condition.

[0117] The constraint that the observation channel model weights are not changed is to keep the observation channel model parameters fixed and not train or update them during the construction and operation of the counterfactual channel.

[0118] The shared weights refer to the set of parameters used by the counterfactual channel in each layer and each hot channel that are completely consistent with the observation channel model, so that the mapping between the two is consistent except for the input gating.

[0119] The counterfact gating is a mechanism that controls the input or internal state of the counterfact channel, used to replace contact resistance and optional environmental parameters with baseline values ​​and shield them from the influence of changes in measured data.

[0120] The baseline contact resistance is a reference contact resistance value used for counterfactual calculations, representing a typical level of health or post-maintenance condition, and can be a fixed setting or estimated from historical data and manufacturing specifications.

[0121] The baseline environmental parameters are a set of reference environmental boundary parameters used for counterfactual calculations, representing standard or neutral environmental conditions, and include at least baseline environmental temperature, wind speed, relative humidity, and cabinet door status.

[0122] The time step is the discrete time interval when the counterfactual channel advances along a unified time axis, and it is consistent with the inference step of the observation channel model.

[0123] The temperature rise response is a sequence of incremental joint surface temperatures relative to ambient temperature output by the counterfactual channel under given current input and baseline conditions.

[0124] The counterfactual temperature rise sequence is a temperature rise response time sequence arranged along a uniform time axis, used for differential and causal analysis with the measured or observed channel output.

[0125] Optionally, step S6 includes:

[0126] Based on the counterfactual temperature rise sequence and the calibrated joint surface temperature data, the temperature difference sequence is calculated on a unified time axis as the calibrated joint surface temperature data minus the counterfactual temperature rise sequence.

[0127] The temperature difference sequence is input into the residual causal decomposition structure, which decomposes it into a residual sequence contributed by contact resistance and an unmodeled perturbation residual sequence.

[0128] The contact resistance estimation update is driven by the residual sequence contributed by the contact resistance. During the update process, the contact resistance estimation is subject to slow variation and monotonic constraints. When there is a maintenance event marker, the contact resistance estimation is allowed to be reset to the baseline contact resistance, which is a pre-set health or post-maintenance state reference value and is not greater than the current contact resistance estimation when the maintenance event is triggered.

[0129] Sparse and short-time constraints are applied to the unmodeled perturbation residual sequence to absorb non-contact resistance factors and avoid erroneous updates;

[0130] After the preset convergence condition is met, a contact resistance estimation sequence is formed, and the deterioration rate sequence is obtained by time difference on a unified time axis.

[0131] Output contact resistance estimation sequence and degradation rate sequence.

[0132] Terminology definition:

[0133] The calibrated connector surface temperature data is the connector surface temperature time series data after infrared emissivity calibration, outlier processing, and timing alignment with other channels.

[0134] The temperature difference sequence is a time series obtained by subtracting the counterfactual temperature rise sequence from the calibrated joint surface temperature data point by point on a unified time axis, and is used as the input for residual causal decomposition.

[0135] The residual causal decomposition structure is a model structure that decomposes the temperature difference sequence into two components: contact resistance factor and unmodeled disturbance factor. It includes at least the branch that generates contact resistance-related residuals and the branch that generates unmodeled disturbance residuals, as well as their constraints.

[0136] The residual sequence contributed by the contact resistance is a time series in the output of the residual causal decomposition structure that is directly related to the change in contact resistance and is used to drive the update of the contact resistance estimation.

[0137] The unmodeled perturbation residual sequence is a time series in the output of the residual causal decomposition structure used to absorb non-contact factors such as environmental transients, sensor noise, and model mismatch.

[0138] The contact resistance estimate is an estimate of the contact resistance state at a given time step, used to construct a single point of the contact resistance estimation sequence.

[0139] The contact resistance estimation update is a process of state recursion of the contact resistance estimation based on the residual sequence contributed by the contact resistance, which is subject to slow variation and monotonicity constraints during execution and allows reset when maintenance event markers exist;

[0140] The maintenance event marker is an external time stamp or binary time series indicating the occurrence of maintenance activities such as maintenance or tightening, which is used to trigger a reset strategy for contact resistance estimation.

[0141] The reset is an update operation that lowers the current contact resistance estimate to a preset or estimated lower baseline range when a maintenance event flag is triggered;

[0142] The sparsity and short-term constraints are regularization and duration restrictions applied to the unmodeled perturbation residual sequence, so that the residual appears intermittently in time and the duration of each occurrence is limited, in order to avoid misoccupation of the contact resistance channel.

[0143] The non-contact factors are those factors that cause temperature differences other than changes in contact resistance, including at least sudden changes in ambient temperature / wind speed / cabinet door status, sensor noise, and unmodeled thermal path disturbances.

[0144] The contact resistance estimation sequence is a time series of contact resistance estimates updated over time, used to characterize the evolution of joint degradation level.

[0145] The time difference is divided into operations that calculate adjacent differences or numerical derivatives of the contact resistance estimation sequence according to the time step, in order to approximate its rate of change;

[0146] The degradation rate sequence is a time series of the rate of change of contact resistance over time obtained by time difference, used to characterize the degradation speed and trend.

[0147] Optionally, step S7 includes:

[0148] Based on the contact resistance estimation sequence and the degradation rate sequence, combined with the parameter uncertainty of the observation channel model and the variance of the current control function residual sequence, the contribution of model parameter uncertainty and the contribution of instrumental variable extrapolation uncertainty are calculated respectively. The two types of uncertainty contributions are weighted and synthesized into the contact resistance estimation sequence through sensitivity propagation to form the upper and lower boundaries of the contact resistance uncertainty interval.

[0149] A risk score is constructed on a unified time axis based on the current value of the contact resistance estimation sequence, the upper boundary of the contact resistance uncertainty interval, and the non-negative part of the deterioration rate sequence. The risk score is compared with a preset alarm threshold. When the risk score reaches or exceeds the preset alarm threshold, a bus trunking joint deterioration early warning signal is generated, and the contact resistance uncertainty interval is updated and provided.

[0150] The output busbar trunking joint deterioration warning signal and the uncertainty range of contact resistance.

[0151] Terminology definition:

[0152] The uncertainty of the observation channel model parameters is a measure of the parameter estimation error and training statistical fluctuation within the observation channel model, reflecting the range of influence of parameter perturbation on the contact resistance estimation results;

[0153] The instrumental variable extrapolation uncertainty is the uncertainty introduced when instrumental variable modeling is applied outside the training distribution or outside the event intensity range, and can be characterized at least by the variance of the residual sequence of the current control function or its segment extrapolation error.

[0154] The contribution of the model parameter uncertainty is the variance or interval share obtained by mapping the uncertainty of the observation channel model parameters to the contact resistance estimation through sensitivity propagation.

[0155] The instrumental variable extrapolation uncertainty contribution is the variance or interval share obtained by mapping the instrumental variable extrapolation uncertainty to the contact resistance estimate through sensitivity propagation.

[0156] The sensitivity propagation is a calculation method that translates output uncertainty by linearizing the local sensitivity (such as Jacobian or gradient) of the contact resistance estimate with respect to model parameters and extrapolated noise.

[0157] The weighted synthesis is a process of combining uncertainties from different sources based on the weights or variance decomposition results obtained from sensitivity propagation to form total uncertainty.

[0158] The contact resistance uncertainty range is an interval defined by upper and lower boundaries and covering the possible range of values ​​for the actual contact resistance at a given confidence level.

[0159] The upper and lower boundaries are the upper and lower limits of the uncertainty range of contact resistance, respectively corresponding to the positive and negative boundaries of the uncertainty after synthesis.

[0160] The risk score is a monotonically increasing scalar function with the current value of the contact resistance estimation sequence, the upper boundary of the contact resistance uncertainty interval, and the non-negative part of the degradation rate sequence as inputs, used to measure the current degree of degradation risk.

[0161] The preset alarm threshold is a threshold parameter used to determine whether an alarm is triggered. It is set offline or online adaptively to balance false alarms and missed alarms.

[0162] The non-negative part is a time series obtained by taking the largest value of zero point by point of the degradation rate sequence, which is used to suppress the influence of the "improvement / reduction" direction on the risk score;

[0163] The busbar joint deterioration early warning signal is an alarm indication message output when the risk score reaches or exceeds the preset alarm threshold, used to indicate that the joint deterioration risk has reached the alarm condition.

[0164] The beneficial effects of this invention are:

[0165] 1. Under the condition that the contact resistance cannot be directly measured, the instrumental variables are constructed by load step and periodic events, the neural instrumental variables are modeled and orthogonalized constraints are applied, and the multi-channel physical information neural network with energy conservation and environmental boundary encoding is combined to couple the contact resistance only to the heating term of the current square, so as to effectively eliminate the mixed factors of environment and load, obtain stable and interpretable contact resistance and degradation rate estimates, and significantly improve the identification accuracy.

[0166] 2. By generating baseline temperature rise through counterfactual channels and counterfactual gating that share weights with observation channels, and inputting the difference between measured and counterfactual data into residual causal decomposition, and applying slow variation, monotonicity and maintenance reset constraints to the evolution of contact resistance, contact factors and non-contact short-term disturbances can be separated, reducing parameter false updates and false alarms, and providing early degradation identification and online tracking capabilities.

[0167] 3. By combining the uncertainty of the integrated model parameters and the uncertainty of the instrumental variable extrapolation, sensitivity propagation is used to form the uncertainty range of the contact resistance. A risk score is constructed based on the current value, the upper bound of the range, and the non-negative degradation rate. The output is a warning result with a range, which improves the credibility and decision-making ability of the warning and adapts to complex working conditions and changing working conditions. Attached Figure Description

[0168] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0169] Figure 1 This is a flowchart of a busbar joint degradation early warning method based on machine learning and causal inference proposed in this invention;

[0170] Figure 2 This is a flowchart of the multi-channel physical information neural network and observation channel model of the present invention;

[0171] Figure 3This is a flowchart of the counterfactual channel and gating with shared weights in this invention. Detailed Implementation

[0172] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0173] refer to Figure 1 A method for early warning of busbar joint degradation based on machine learning and causal inference, comprising:

[0174] S1. Collect the current at the connector, the surface temperature of the connector, and environmental parameters, and perform emissivity calibration and timing alignment.

[0175] S2. Detect time-series changes in current data, identify load step or periodic load events, and construct instrumental variables and form a sequence of instrumental variables accordingly.

[0176] S3. Train a neural instrumental variable model based on the instrumental variable sequence, environmental parameters and current data to characterize the structural influence of instrumental variables and environment on current, calculate current control residuals and obtain control residual sequence;

[0177] S4. Construct and train a multi-channel physical information neural network based on the control residual sequence and aligned data, apply energy conservation constraints and environmental boundary encoding, couple the contact resistance only to the heating term represented by the square of the current data multiplied by the contact resistance, inject the control residual into the heating term and boundary term and apply orthogonalization constraints to the loss, and apply slow variation and monotonic constraints to the time evolution of the contact resistance, and allow it to be reset to the baseline contact resistance when there is a maintenance event marker, thus obtaining the observation channel model;

[0178] S5. Construct a counterfactual channel that shares weights with the observation channel model under the condition of freezing the observation channel model weights. Set the contact resistance to the baseline value through counterfactual gating, and optionally set the environmental parameters to the baseline. Calculate the counterfactual temperature rise to obtain the counterfactual temperature rise sequence.

[0179] S6. The temperature difference sequence is obtained by subtracting the counterfactual temperature rise from the surface temperature. The residual causal decomposition is then performed to separate the residual contributed by the contact resistance and the unmodeled disturbance residual. The former is used to update the contact resistance estimate under slow variation and monotonic constraints, and reset is allowed when the maintenance event marker exists. The latter is subject to sparse and short-term constraints to absorb non-contact factors, thus obtaining the contact resistance estimation sequence and the degradation rate sequence.

[0180] S7. Based on the contact resistance estimation sequence and the degradation rate sequence, combined with the uncertainty of the observation channel model parameters and the extrapolation uncertainty of the instrumental variables, the contact resistance uncertainty interval is estimated, and a risk score is constructed accordingly. The score is compared with the preset threshold, and the degradation warning signal and uncertainty interval are output.

[0181] In this specific embodiment, S1 includes:

[0182] Current sensors, infrared temperature sensors, and environmental sensors are installed at the busbar joints to ensure that current data, joint surface temperature data, and environmental parameter data are timestamped and the sampling source and channel identifier are recorded at the acquisition end.

[0183] To perform emissivity calibration for infrared temperature measurement errors: First, calibrate plates or contact temperature measuring points are placed near the connector to obtain a reference temperature. The reference temperature is then compared with the infrared-measured connector surface temperature to determine the emissivity calibration coefficient. This coefficient is then applied to the infrared temperature measurement data to correct radiation deviations. The emissivity calibration can be performed using the following relationship:

[0184] ;

[0185] in, Indicates time index The calibrated connector surface temperature data, Indicates the infrared temperature sensor in time index The surface temperature data of the joint. Indicates the time index of calibration plates or contact temperature measuring points. The provided reference temperature Indicates basis and Compare with the determined emissivity calibration coefficients, Represents a time index on a unified timeline;

[0186] After completing the emissivity calibration, the current data, connector surface temperature data and environmental parameter data are aligned on the same time axis according to the time alignment rules, sorted by timestamp, and resampled or interpolated for samples with unequal intervals. At the same time, duplicate or out-of-order records are removed so that the multi-source data are aligned point by point on the same time axis.

[0187] During the collection and alignment process, environmental parameter data includes, but is not less than, four types of information: ambient temperature, relative humidity, wind speed, and cabinet door status, and is matched point by point with the time axis before output.

[0188] The output calibrated current data is denoted as The calibrated joint surface temperature data is recorded as follows: The calibrated environmental parameter data are recorded as follows: ,in At least include ambient temperature relative humidity Wind speed Cabinet door status and , Align on a unified timeline.

[0189] In this specific embodiment, S2 includes:

[0190] Based on calibrated current data on a unified time axis To perform time-series change detection, firstly... Denoising was performed to suppress transient spikes and measurement noise. Subsequently, a sliding window slope and second-order difference were used to jointly locate candidate change points. And set a dwell time criterion after the candidate point to confirm whether the new load level is stable;

[0191] For each candidate point, calculate the representative load level of the windows before and after the candidate point, denoted as . and and with step amplitude Determine the load step event, where Indicates the candidate change point The step amplitude at that point, Indicates length as Sliding window Previous current data The mean, Indicates length as Sliding window Subsequent current data The mean, and These represent the lengths of the front and rear windows, respectively, and are consistent with the sampling step size of the unified time axis. This indicates the candidate change time obtained by locating the change point;

[0192] when And the new level residence time is not less than the threshold. When this occurs, the segment is marked as a load step event, and the start time, end time, event type, step amplitude, and load levels before and after the event are recorded. Indicates the threshold value of the step amplitude. Indicates the threshold for dwell time;

[0193] At the same time, Perform autocorrelation analysis or power spectral density analysis to identify periodic load events, at the dominant frequency. Located in the set frequency band And energy percentage Not less than the threshold And the number of consecutive cycles is not less than At that time, the segment is marked as a periodic load event and the event start time, end time, event type, main frequency, amplitude, and duration are recorded. Indicates the main frequency, Indicates the set frequency band. This represents the proportion of the spectral energy in the main frequency neighborhood to the total spectral energy. Indicates the energy percentage threshold. Indicates the minimum number of cycles;

[0194] After determining the load step event and the periodic load event, a sequence of instrumental variables is constructed based on the event records, denoted as... Each instrumental variable corresponds to an event record and includes the event timestamp, event type, event intensity parameter, and event duration. Events with step amplitude below a threshold or insufficient duration are discarded. Output the data along a unified timeline for use in subsequent training of neural instrumental variable models.

[0195] In this specific embodiment, S3 includes:

[0196] Based on instrumental variable sequence calibrated environmental parameter data and calibrated current data Constructing a neural instrumental variable model To characterize the structural effects of instrumental variables and environmental parameters on current;

[0197] Training proceeds along a unified timeline, with the model input being... and And optional historical data, the output is the predicted current. Among them, the improvement is achieved by constraining the fitting error. right The approximation is enhanced by event relevance terms. right The explanatory power is weakened by orthogonalization constraints. The interference along the instrumental variable path affects the output and causes the residuals to... Approximate independence and orthogonality;

[0198] After training is complete, the model predictions are calculated on a unified time axis and the residuals of the current control function are obtained. Its definition is:

[0199] ;

[0200] in Indicates time index The residual of the current control function, Indicates time index The calibrated current data, Represents the parameter vector The determined neural instrumental variable model at time index The predicted output current, Represents a time index on a unified timeline;

[0201] Then on Perform correlation and orthogonality tests, using correlation coefficient thresholds. and and significance level Determine its relationship with Independence and If the orthogonality of the loss function is not met, the event relevance weights and orthogonality constraint weights in the loss function are adjusted and the training is repeated until the test conditions are met.

[0202] Output the current control function residual sequence arranged on a uniform time axis .

[0203] In this specific embodiment, S4 includes:

[0204] Based on the residual sequence of the current control function and calibrated current data calibrated environmental parameter data Compared with the calibrated connector surface temperature data Construct and train a multi-channel physical information neural network The network is configured with busbar copper busbar channels, contact surface channels, bolt channels, gasket channels, and air boundary layer channels, and connected to each channel by energy conservation constraints. Simultaneously, environmental boundary conditions are encoded to... ambient temperature relative humidity Wind speed Cabinet door status Mapped to boundary parameters to drive convective and radiative heat transfer capabilities;

[0205] In the contact surface channel, a heating term is defined by multiplying the square of the current data by the contact resistance. The contact resistance is coupled only to this heating term and not directly coupled to other channels. Simultaneously, the residual of the current control function is... The injected heat term and the boundary term form an orthogonal coupling, and an orthogonalization constraint is applied to the loss to reduce the sensitivity of contact resistance to the residual and the influence of hybrid paths;

[0206] Contact resistance using time evolution parameters Characterize and impose slow-varying and monotonic constraints on it, allowing it to reset to the baseline contact resistance when a maintenance event flag exists, so that the network outputs the temperature rise response on a uniform time axis, denoted as . During the training process, and The difference is used as the main loss, and combined with energy conservation and orthogonalization constraints as physical and causal regularization, after reaching the preset convergence condition, an observation channel model is obtained to characterize the actual observation process. The heating term of the contact surface channel can be expressed as:

[0207] ;

[0208] in Indicates time index The power of the local heat source in the contact surface channel. Indicates time index The calibrated current data, Indicates time index The contact resistance time evolution parameter, This represents the weighting coefficient from the residual of the injected current control function to the heating term. Indicates time index The residual of the current control function, This represents a time index on a unified timeline.

[0209] In this specific embodiment, S5 includes:

[0210] A counterfactual channel is constructed without changing the weights of the observation channel model, so that the counterfactual channel shares all weights with the observation channel model and progresses on a unified time axis with the same time step as the observation channel model; specifically, the observation channel model trained and with its parameters frozen in step S4 is denoted as... Its parameter vector Keep it unchanged, and connect the calibrated current data to the input terminal. Compared with calibrated environmental parameter data By using counterfactual gating to force the contact resistance path to be set to the baseline contact resistance, and selecting the baseline environmental parameters or maintaining the measured environmental parameters as needed for the environmental parameter path, the model calculates the temperature rise response under baseline conditions on a unified time axis and generates a counterfactual temperature rise sequence. The counterfactual temperature rise can be expressed as:

[0211] ;

[0212] in Indicates time index Counterfactual temperature rise output, This indicates that the model shares all weights with the observation channel model and freezes the parameter vector. Function mapping, Indicates time index The calibrated current data, This indicates the output of the environmental parameter data channel obtained through counterfactual gating. This indicates the contact resistance channel output obtained through counterfactual gating and set as the baseline contact resistance. This indicates that the preset baseline contact resistance value is used to represent the health or post-maintenance status. Represents a time index on a unified timeline;

[0213] Among them, the gating of environmental parameters is controlled by binary indicator variables, which are set when the replacement environment path is selected. When choosing to keep the actual test environment path This represents a preset set of baseline environmental parameters, including at least baseline ambient temperature, baseline relative humidity, baseline wind speed, and baseline cabinet door status. Maintain isomorphism and Align point-by-point along a unified timeline;

[0214] Will Output as a counterfactual temperature rise sequence.

[0215] In this specific embodiment, S6 includes:

[0216] Based on the counterfactual temperature rise sequence and the calibrated joint surface temperature data, the temperature difference sequence is calculated on a unified time axis and input into the residual causal decomposition structure to separate the contact resistance factor from the unmodeled disturbance factor, thereby updating the contact resistance estimate and degradation rate.

[0217] Specifically, the temperature difference sequence is first calculated point-by-point along a unified time axis, defined as:

[0218] ;

[0219] in, Indicates time index Temperature difference sequence, Indicates time index The calibrated connector surface temperature data, Indicates time index Counterfactual temperature rise output, Represents a time index on a unified timeline;

[0220] Then Input the residual causal decomposition structure and decompose it into a residual sequence contributed by contact resistance. Compared with unmodeled perturbation residual sequences ,in Used to drive the estimation and updating of contact resistance time evolution parameters. By applying sparse and short-term constraints, environmental transients, sensor noise, and unmodeled thermal path disturbances are absorbed to avoid erroneous updates.

[0221] Contact resistance estimation is characterized using the same time evolution parameters as in step S4, denoted as... During the update process Slow-varying and monotonic constraints are applied to reflect the irreversibility of contact degradation, and maintenance gating is used to reset to a lower baseline range when maintenance event markers are present to reflect maintenance improvements. The maintenance event markers are recorded as binary time series and aligned with a unified time axis.

[0222] Once the residual causal decomposition and state update meet the preset convergence condition, a contact resistance estimation sequence is formed. and on a unified timeline The degradation rate sequence is obtained by performing time difference analysis to characterize the rate and trend of contact resistance change over time;

[0223] Output contact resistance estimation sequence and degradation rate sequence.

[0224] In this specific embodiment, S7 includes:

[0225] On a unified time axis, an uncertainty interval and risk score are constructed by combining the contact resistance estimation sequence and the degradation rate sequence, and a degradation early warning signal is output: the contact resistance estimation sequence is denoted as... The degradation rate sequence is denoted as ,in Indicates time index The estimated contact resistance value, Indicates time index Depend on The degradation rate is obtained by time difference and aligned with a unified time axis;

[0226] Secondly, to characterize the parameter uncertainty of the observation channel model, let the parameter vector of the observation channel model be... Its parameter covariance is and to about Jacobian sensitivity is denoted as Based on this, the variance contribution of the parameter uncertainty mapped to the contact resistance estimate via sensitivity propagation is denoted as... ;

[0227] Simultaneously characterizing the extrapolation uncertainty of instrumental variables, using the residual sequence of the current control function. The uncertainty intensity of the time-variable variance estimation extrapolation is denoted as . and order The Jacobian sensitivity of the instrumental variable extrapolation noise path is denoted as . Therefore, the contribution of the extrapolation uncertainty to the variance of the contact resistance estimate is denoted as . ;

[0228] The two types of variance contributions are weighted by sensitivity and combined to form the total standard deviation of the contact resistance estimate, denoted as . quantiles at a given confidence level The following forms the uncertainty range for contact resistance, with the upper and lower boundaries denoted as . and , its origin Subtract and add respectively get;

[0229] Finally, a risk score is constructed and compared with a preset alarm threshold to determine whether a degradation warning should be issued. The risk score can be in a weighted linear form.

[0230] ;

[0231] in Indicates time index Risk score, This indicates that the non-negative weighting coefficients are used to balance the three contributions. Indicates time index The upper boundary of the uncertainty range of contact resistance, Indicates time index The non-negative part of the degradation rate sequence is... Take the larger value than zero. Represents a time index on a unified timeline;

[0232] When the risk score reaches or exceeds the preset alarm threshold Output busbar joint deterioration early warning signal and simultaneously output uncertainty range. .

[0233] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0234] This invention focuses on indirect robust estimation under conditions where contact resistance cannot be directly measured. It uses event-driven neural instrumental variables to capture exogenous disturbances from load steps and cyclic loads, and obtains current control residuals through orthogonalization constraints to eliminate confounding. Subsequently, a multi-channel thermoelectric physical information neural network is used to couple only the contact resistance to the current square heating term under energy conservation and environmental boundary encoding to ensure identifiability. Baseline surface temperature rise is generated in a counterfactual gating system with shared weights, and the difference between the baseline and observed temperature rise is separated from contact factors and unmodeled disturbances through residual causal decomposition. Then, the contact resistance and degradation rate are updated under slow variation, monotonicity, and maintenance reset constraints. Reliable early warning is achieved by combining uncertainty intervals and risk scores.

[0235] This invention proposes four improvements to address the technical problems: orthogonal coupling of neural instrument variables and physical information neural networks reduces the sensitivity of parameters to load and environmental paths; multi-channel thermoelectric modeling and boundary coding refine heat conduction, convection, and radiation, improving the accuracy of temperature rise simulation and identification; shared-weight counterfactual channels and gating ensure consistency between observations and counterfactual data, enhancing the stability of baseline generation; residual causal decomposition combined with sparsity, short-time constraints, and slow-varying, monotonic contact resistance constraints, along with maintenance reset constraints, suppresses erroneous updates and enhances early degradation identification. Combined with uncertainty sensitivity propagation and interval-based early warning, it significantly reduces false alarms and missed alarms, improving robustness and decision-making ability under complex operating conditions.

Claims

1. A method for early warning of busbar joint deterioration based on machine learning and causal inference, characterized in that, include: S1. Collect the current at the connector, the surface temperature of the connector, and environmental parameters, and perform emissivity calibration and timing alignment. S2. Detect time-series changes in current data, identify load step or periodic load events, and construct instrumental variables accordingly to form a sequence of instrumental variables. S3. Train a neural instrumental variable model based on the instrumental variable sequence, environmental parameters and current data to characterize the structural influence of instrumental variables and environment on current, calculate current control residuals and obtain control residual sequence; S4. Construct and train a multi-channel physical information neural network based on the control residual sequence and aligned data, apply energy conservation constraints and environmental boundary encoding, couple the contact resistance only to the heating term represented by the square of the current data multiplied by the contact resistance, inject the control residual into the heating term and boundary term and apply orthogonalization constraints to the loss, and apply slow variation and monotonic constraints to the time evolution of the contact resistance, and allow it to be reset to the baseline contact resistance when there is a maintenance event marker, thus obtaining the observation channel model; S5. Construct a counterfactual channel that shares weights with the observation channel model under the condition of freezing the observation channel model weights. Set the contact resistance to the baseline value through counterfactual gating, and optionally set the environmental parameters to the baseline. Calculate the counterfactual temperature rise to obtain the counterfactual temperature rise sequence. S6. The temperature difference sequence is obtained by subtracting the counterfactual temperature rise from the surface temperature. The residual causal decomposition is then performed to split the residual sequence into the contact resistance contribution residual sequence and the unmodeled disturbance residual sequence. The contact resistance estimate is updated with the contact resistance contribution residual sequence under slow variation and monotonic constraints, and reset is allowed when the maintenance event marker exists. Sparse and short-term constraints are applied to the unmodeled disturbance residual sequence to absorb non-contact factors, resulting in the contact resistance estimation sequence and the degradation rate sequence. S7. Based on the contact resistance estimation sequence and the degradation rate sequence, combined with the uncertainty of the observation channel model parameters and the extrapolation uncertainty of the instrumental variables, the contact resistance uncertainty interval is estimated, and a risk score is constructed accordingly. The score is compared with the preset threshold, and the degradation warning signal and uncertainty interval are output.

2. The busbar joint degradation early warning method based on machine learning and causal inference according to claim 1, characterized in that, S1 includes: At the busbar joint, current sensors, infrared temperature sensors, and environmental sensors are used to collect current data, joint surface temperature data, and environmental parameter data, respectively, and timestamps are added to the collected data. An emissivity calibration coefficient is established by infrared emissivity calibration. The emissivity calibration coefficient is determined by comparing the joint surface temperature data with the reference temperature provided by the calibration plate or contact temperature measuring point. The emissivity calibration coefficient is applied to the joint surface temperature data to correct the radiation error and generate calibrated joint surface temperature data. According to the time alignment rules, the current data, connector surface temperature data and environmental parameter data are aligned on the same time axis, sorted by timestamp, and resampled or interpolated for samples with unequal intervals. At the same time, duplicate or out-of-order records are removed to align the data on the same time axis. The environmental parameter data includes at least ambient temperature, relative humidity, wind speed and cabinet door status. Output calibrated current data, calibrated connector surface temperature data, and calibrated environmental parameter data.

3. The method for early warning of busbar joint deterioration based on machine learning and causal inference according to claim 1, characterized in that, S2 includes: Based on the calibrated current data, time-series change detection is performed on a unified time axis. First, the calibrated current data is denoised to suppress instantaneous spikes and measurement noise. Then, a sliding window is used to calculate the slope and second-order differential to locate the change point. When the current stabilizes at the new level after the change point and reaches the set dwell time and the step amplitude exceeds the preset threshold, the data segment is marked as a load step event and the event start time, end time, event type, step amplitude, and load levels before and after are recorded. Simultaneously, autocorrelation analysis or power spectral density analysis is performed on the calibrated current data to identify periodic components. When the main frequency falls into the set frequency band and the energy proportion exceeds the preset threshold and continues to exceed the minimum number of cycles, the data segment is marked as a periodic load event and the event start time, end time, event type, main frequency, amplitude and duration are recorded. Construct a sequence of instrumental variables based on the labeled load step events and periodic load events, so that each instrumental variable corresponds to an event record and includes the event timestamp, event type, event intensity parameter and event duration. After removing events with step amplitude below the threshold or insufficient duration, output the sequence of instrumental variables.

4. The busbar joint degradation early warning method based on machine learning and causal inference according to claim 1, characterized in that, S3 includes: Based on the instrumental variable sequence, calibrated environmental parameter data, and calibrated current data, a neural instrumental variable model is trained to characterize the structural influence of instrumental variables and environmental parameters on current data; During training, the loss function is set to include an event correlation term and an orthogonalization constraint. The event correlation term is used to enhance the explanatory power of the instrumental variables on the current data, and the orthogonalization constraint is used to reduce the interference of environmental parameters on the path of the instrumental variables. After training is completed, the current data is predicted on a unified time axis using a neural instrumental variable model, and the residual of the current control function is defined as the calibrated current data minus the model-predicted current, generating a sequence of current control function residuals. Correlation and orthogonality tests are performed on the residual sequence of the current control function to confirm its independence from the instrumental variable sequence and its orthogonality to the environmental parameters. If the preset test threshold is not met, the weights of the loss function are adjusted and training is repeated until the test threshold is met. Output current control function residual sequence.

5. The method for early warning of busbar joint deterioration based on machine learning and causal inference according to claim 1, characterized in that, S4 includes: Based on the residual sequence of the current control function, the calibrated current data, the calibrated environmental parameter data, and the calibrated joint surface temperature data, a multi-channel physical information neural network is constructed and trained. The network is configured with busbar copper bus channel, contact surface channel, bolt channel, gasket channel, and air boundary layer channel, and each channel is connected by energy conservation constraints. At the same time, the environmental boundary conditions are encoded to map the environmental temperature, relative humidity, wind speed, and cabinet door status into the boundary parameters of each channel. In the contact surface channel, the heating term is defined by multiplying the square of the current data by the contact resistance, and the contact resistance is coupled only to this heating term and not directly coupled to other channels. The residual sequence of the current control function is injected into the heating term and the boundary term respectively to form orthogonal coupling, and orthogonalization constraints are applied to the loss function to reduce the sensitivity of the contact resistance to the residual of the current control function. Slow variation and monotonic constraints are applied to the time evolution of contact resistance, and when a maintenance event is triggered, the contact resistance is allowed to be reset to the baseline contact resistance, which is a pre-set health or post-maintenance state reference value and is not greater than the current contact resistance estimate at the time of maintenance triggering. After training is completed by minimizing the difference between the calibrated joint surface temperature data and the network output temperature rise, and the preset convergence condition is met, an observation channel model is generated to characterize the actual observation process.

6. The busbar joint degradation early warning method based on machine learning and causal inference according to claim 1, characterized in that, S5 include: Based on the observation channel model, a counterfactual channel is constructed without changing the weights of the observation channel model, so that the counterfactual channel shares all weights with the observation channel model; Based on the calibrated current data and calibrated environmental parameter data, the contact resistance is set to a preset baseline contact resistance through counterfactual gating, and the environmental parameters are selectively set to preset baseline environmental parameters. This allows the counterfactual channel to calculate the temperature rise response under the baseline contact resistance and the selected environmental conditions on a unified time axis with a time step consistent with the observation channel model, thereby generating a counterfactual temperature rise sequence.

7. The busbar joint degradation early warning method based on machine learning and causal inference according to claim 1, characterized in that, S6 include: Based on the counterfactual temperature rise sequence and the calibrated joint surface temperature data, the temperature difference sequence is calculated on a unified time axis as the calibrated joint surface temperature data minus the counterfactual temperature rise sequence. The temperature difference sequence is input into the residual causal decomposition structure, which decomposes it into a residual sequence contributed by contact resistance and an unmodeled perturbation residual sequence. The contact resistance estimation update is driven by the residual sequence contributed by the contact resistance. During the update process, the contact resistance estimation is subject to slow variation and monotonic constraints. When there is a maintenance event marker, the contact resistance estimation is allowed to be reset to the baseline contact resistance, which is a pre-set health or post-maintenance state reference value and is not greater than the current contact resistance estimation when the maintenance event is triggered. Sparse and short-time constraints are applied to the unmodeled perturbation residual sequence to absorb non-contact resistance factors and avoid erroneous updates; After the preset convergence condition is met, a contact resistance estimation sequence is formed, and the deterioration rate sequence is obtained by time difference on a unified time axis. Output contact resistance estimation sequence and degradation rate sequence.

8. The method for early warning of busbar joint deterioration based on machine learning and causal inference according to claim 1, characterized in that, S7 includes: Based on the contact resistance estimation sequence and the degradation rate sequence, combined with the parameter uncertainty of the observation channel model and the variance of the current control function residual sequence, the contribution of model parameter uncertainty and the contribution of instrumental variable extrapolation uncertainty are calculated respectively. The two types of uncertainty contributions are weighted and synthesized into the contact resistance estimation sequence through sensitivity propagation to form the upper and lower boundaries of the contact resistance uncertainty interval. A risk score is constructed on a unified time axis based on the current value of the contact resistance estimation sequence, the upper boundary of the contact resistance uncertainty interval, and the non-negative part of the deterioration rate sequence. The risk score is compared with a preset alarm threshold. When the risk score reaches or exceeds the preset alarm threshold, a bus trunking joint deterioration early warning signal is generated, and the contact resistance uncertainty interval is updated and provided. The output busbar trunking joint deterioration warning signal and the uncertainty range of contact resistance.

Images