Switch fault prediction method and early warning system based on multi-sensor fusion

Abstract

The invention relates to the technical field of switch monitoring, in particular to a switch fault prediction method and early warning system based on multi-sensor fusion, and the method comprises the steps: deploying a plurality of types of sensors at key parts of switch equipment, collecting real-time operation state data through a sensor controller, carrying out the noise reduction and time-space alignment processing of original sensing data, and carrying out the early warning of the switch fault. Constructing a multi-source sensing sample data set; acquiring inherent attribute data and historical fault records of the switch equipment, and adding fault tags for the multi-source sensing sample data set; inputting the data set with the fault label into a deep belief network based on an attention mechanism for training to obtain a switch fault prediction model; fusion reasoning is performed on multi-sensor data acquired in real time through the trained model, a fault prediction result is output, and a maintenance decision suggestion is generated in combination with inherent attribute data of the switch equipment, so that accurate prediction and intelligent maintenance decision of the switch equipment fault are realized, and the operation safety and reliability of the switch equipment are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of switch monitoring, in particular to a switch fault prediction method and early warning system based on multi-sensor fusion. BACKGROUND

[0002] With the rapid popularization of smart home, switch equipment as the core component connecting power supply and various electrical appliances in the household power supply scene, its function has gradually upgraded from traditional manual on-off control to intelligent terminal integrating power safety monitoring, equipment linkage control and intelligent state analysis. In the daily smart home scene, switch equipment (such as intelligent circuit breaker, intelligent socket, intelligent switch panel, etc.) plays a key role in controlling lighting, household appliance start-stop and ensuring household power safety, and the stability of its running state directly affects the power experience and life and property safety of family members.

[0003] However, the traditional switch equipment fault prediction method relies on a single type of sensor, and the collected data is difficult to fully reflect the complex running state of the equipment, such as mechanical wear, insulation aging, abnormal temperature rise and other multi-dimensional characteristics, and the robustness to noise data is insufficient, which limits the prediction accuracy. In addition, multi-source heterogeneous sensor data has problems such as time and space asynchronization and weak feature correlation, and traditional fusion methods are difficult to effectively mine the deep coupling relationship between data, which cannot meet the prediction needs under complex working conditions.

[0004] The above content is only used to assist in understanding the technical solutions of the present application, and does not represent the acknowledgement of the above content as prior art. SUMMARY

[0005] The main purpose of the present application is to provide a switch fault prediction method and early warning system based on multi-sensor fusion, which aims to solve the technical problem that the traditional switch equipment fault prediction method relies on a single type of sensor, and the collected data is difficult to fully reflect the complex running state of the equipment.

[0006] To achieve the above purpose, the present application provides a switch fault prediction method based on multi-sensor fusion, which comprises: Deploying multiple types of sensors at key parts of switch equipment, collecting real-time running state data of switch equipment through sensor controller to obtain original sensor data, and performing noise reduction and time-space alignment processing on the original sensor data to construct a multi-source sensor sample data set; Obtaining inherent attribute data and historical fault records of switch equipment, and adding fault labels to the multi-source sensor sample data set; Inputting the multi-source sensor sample data set with fault labels into a deep belief network based on attention mechanism for training to obtain a trained switch fault prediction model; The switch fault prediction model is used to fuse and infer real-time collected multi-sensor data, to obtain a fault prediction result, and to generate a maintenance decision suggestion in combination with inherent attribute data of the switch device.

[0007] Optionally, the denoising and spatio-temporal alignment processing of the original sensing data to construct a multi-source sensing sample data set comprises: Obtaining original sensing data and performing denoising processing, wherein the original sensing data comprises current sensor data, voltage sensor data, temperature sensor data, and mechanical characteristic sensor data; Interpolating original sensing data of different sampling rates to a unified time scale to construct a spatio-temporal aligned original data matrix; Performing wavelet packet decomposition on the original data matrix, and reconstructing each frequency band signal obtained by the decomposition into a feature component; Calculating time domain features and frequency domain features of each feature component, screening sensitive features related to switch faults through a maximum mutual information coefficient, and constructing a multi-source sensing sample data set.

[0008] Optionally, a reconstruction calculation formula of each frequency band signal is:

[0009] wherein, is a reconstructed signal of the jth layer and the kth frequency band, is a low-pass filter coefficient of wavelet packet decomposition, and is a low-pass filter coefficient used in the reconstruction of the jth layer and the kth frequency band; is a high-pass filter coefficient of wavelet packet decomposition, and is a high-pass filter coefficient used in the reconstruction of the jth layer and the kth frequency band; is a decomposition signal of the 2kth frequency band of the j-1th layer, is a low-pass component source of the reconstruction of the kth frequency band of the current layer, and t-n represents time domain translation of a signal, i.e., a time shift term in convolution operation with a filter coefficient; is a decomposition signal of the 2k+1th frequency band of the j-1th layer, and is a high-pass component source of the reconstruction of the kth frequency band of the current layer.

[0010] Optionally, the inherent attribute data and historical fault records of the switch device are obtained, and a fault label is added to the multi-source sensing sample data set, comprising: Obtaining inherent attribute data and historical fault records of the switch device, and adding a fault label to the multi-source sensing sample data set according to the historical fault records; If the historical fault records are serious faults, the fault label is set to -1, and the serious faults include switch refusal or misoperation; If it is a slight fault, the fault label is set to 1, and the slight fault includes an increase in switch contact resistance and a switch mechanical jamming warning. If normal, set the fault label to 0; add the fault label to the multi-source sensing sample data set; The intrinsic attribute data includes device model, rated voltage, rated current, operation life, and installation environment temperature and humidity.

[0011] Optionally, the multi-source sensing sample data set with the fault label is input into a deep belief network based on an attention mechanism for training to obtain a trained switch fault prediction model, including: The multi-source sensing sample data set with the fault label is divided into a training set and a test set and input into a deep belief network; deep layer nonlinear features of the multi-source sensing sample data set are extracted through multi-layer nonlinear mapping of the deep belief network, and a calculation formula is as follows:

[0012] wherein, is an activation output of the lth layer, and are a weight matrix and a bias vector, respectively, is an activation function; An attention mechanism is introduced at a feature output layer of the deep belief network, and dynamic weights are allocated to different sensor features, wherein a dynamic weight allocation calculation formula is as follows:

[0013] wherein, is an attention weight of the ith sensor feature, is a total number of sensor features, is an attention score of the ith sensor feature, is a high layer feature of the ith sensor, U is a feature mapping matrix, v is an attention weight vector, c is an attention bias parameter, U, v, and c are all learnable parameters; is an exponential function, is a hyperbolic tangent activation function; Different sensor features after weight allocation are weighted and fused, the fused features are input into a Softmax classifier for fault type prediction, network parameters are optimized through a back propagation algorithm, and a loss function adopts cross entropy:

[0014] wherein, is a cross entropy loss value, which measures a difference between a switch fault prediction probability and a fault label, and the smaller the value is, the more accurate the prediction is, is a fault label of the e th sample, if the e th sample belongs to a d th fault, then , otherwise. ; is the predicted probability that the e-th sample belongs to the d-th failure class, N is the total number of samples, and E is the number of failure classes; stopping training when the failure prediction accuracy of the deep belief network on the test set no longer improves, saving the network parameters of the deep belief network, and taking the deep belief network as a switch failure prediction model.

[0015] Optionally, the trained switch failure prediction model is used to fuse and infer the real-time collected multi-sensor data to obtain a failure prediction result, and a maintenance decision suggestion is generated in combination with inherent attribute data of the switch device, including: The trained switch failure prediction model is used to fuse and infer the real-time collected multi-sensor data to obtain a failure prediction result, and the failure prediction result includes a failure probability corresponding to a switch failure type and a predicted switch remaining life of the switch failure type; A health degree evaluation model is constructed according to the failure probability corresponding to the switch failure type, the predicted switch remaining life of the switch failure type, and the inherent attribute data of the switch device, and a health degree index of the switch device is output through the health degree evaluation model; A maintenance decision suggestion of the switch device is generated according to the health degree index, and the maintenance decision suggestion includes: if the health degree index is less than a first preset health degree index, an emergency maintenance alarm of the switch device is triggered; if the health degree index is greater than or equal to the first preset health degree index and less than a second preset health degree index, a planned maintenance warning of the switch device is issued; and if the health degree index is greater than or equal to the second preset health degree index, the switch device is determined to be in a normal operating state, and continues to be monitored, wherein the second preset health degree index is greater than the first preset health degree index.

[0016] Optionally, the calculation formula of the health degree evaluation model is:

[0017] wherein, HI is the health degree index of the switch device, RUL is the predicted switch remaining life of the switch device, is the rated life of the switch device, is the failure probability of the switch device, is the average temperature of the switch device, is the maximum allowable temperature of the switch device, is a weight coefficient, wherein represents the contribution weight of the switch remaining life to the health degree index, represents the contribution weight of the switch failure probability to the health degree index, represents the contribution weight of the switch temperature state to the health degree index, .

[0018] In addition, to achieve the above object, the application further provides a switch fault early warning system based on multi-sensor fusion, comprising: A data acquisition module is configured to deploy multiple types of sensors at key positions of switch devices, acquire real-time running state data of the switch devices through a sensor controller to obtain original sensing data, and perform noise reduction and space-time alignment processing on the original sensing data to construct a multi-source sensing sample data set. A label labeling module is configured to acquire inherent attribute data and historical fault records of the switch devices, and add fault labels to the multi-source sensing sample data set. A model training module is configured to input the multi-source sensing sample data set with fault labels into a deep belief network based on an attention mechanism for training to obtain a trained switch fault prediction model. An inference decision module is configured to perform fusion inference on real-time acquired multi-sensor data through the trained switch fault prediction model to obtain a fault prediction result, and generate a maintenance decision suggestion in combination with inherent attribute data of the switch devices.

[0019] In addition, to achieve the above object, the application further provides a switch fault prediction device based on multi-sensor fusion, comprising a memory, a processor, and a switch fault prediction program based on multi-sensor fusion stored on the memory and executable on the processor, wherein the switch fault prediction program based on multi-sensor fusion is configured to implement the steps of the switch fault prediction method based on multi-sensor fusion according to any one of the above.

[0020] In addition, to achieve the above object, the application further provides a storage medium having a switch fault prediction program based on multi-sensor fusion stored thereon, wherein the switch fault prediction program based on multi-sensor fusion, when executed by a processor, implements the steps of the switch fault prediction method based on multi-sensor fusion according to any one of the above.

[0021] The application provides a switch fault prediction method based on multi-sensor fusion, which breaks through the monitoring limitation of a single sensor, comprehensively covers multi-dimensional characteristics such as electrical parameters, mechanical characteristics, temperature rise changes and the like of a device, effectively captures weak abnormal signals of early faults, and significantly improves the information integrity of fault prediction by deploying multiple types of sensors at key positions of a switch device, collecting multi-dimensional operating state data and performing noise reduction and space-time alignment processing; the inherent properties of the switch device and historical fault records are integrated into data set label labeling, so that model training takes into account the common law and individual characteristics of the device, avoids one-size-fits-all prediction, enhances the adaptation ability to different models and different operating life devices, and improves the pertinence and reliability of the prediction result; the deep belief network based on the attention mechanism can adaptively allocate the weight of multi-sensor data, focus on key features with high contribution to fault prediction, suppress noise interference, strengthen the capture ability of the model to early fault features, significantly improve the prediction accuracy and generalization performance, and perform better especially under complex working conditions; the model output not only contains the fault prediction result, but also generates customized maintenance decision suggestions in combination with the inherent properties of the device, extends the fault early warning to the development of maintenance strategies, promotes the operation and maintenance mode from post-processing to active prevention, effectively reduces the risk of unplanned power outage, reduces the cost of artificial inspection, and improves the intelligent level of the whole life cycle management of the device. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 is a structure schematic diagram of a switch fault prediction device based on multi-sensor fusion of a hardware operating environment related to an embodiment scheme of the application. Figure 2 is a flow schematic diagram of a switch fault prediction method based on multi-sensor fusion of an embodiment of the application. Figure 3 is a structure block diagram of a switch fault early warning system based on multi-sensor fusion of an embodiment of the application.

[0023] The implementation, functional features and advantages of the application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0024] It should be understood that the specific embodiments described herein are only used to explain the application, and are not used to limit the application.

[0025] Reference Figure 1 , Figure 1 is a structure schematic diagram of a switch fault prediction device based on multi-sensor fusion of a hardware operating environment related to an embodiment scheme of the application.

[0026] As Figure 1As shown in the figure, the switch fault prediction device based on multi-sensor fusion can include a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to realize the connection communication between the components. The user interface 1003 can include a display screen. The optional user interface 1003 can also include a standard wired interface and a wireless interface. The wired interface of the user interface 1003 can be a USB interface in the present application. The network interface 1004 can optionally include a standard wired interface and a wireless interface (such as a wireless fidelity (WIreless-FIdelity, WI-FI) interface). The memory 1005 can be a high-speed random access memory (RAM) or a stable memory (Non-volatile Memory, NVM), such as a disk memory. The memory 1005 can also be an independent storage device from the aforementioned processor 1001.

[0027] Those skilled in the art can understand that Figure 1 The structure shown in the figure does not constitute a limitation on the switch fault prediction device based on multi-sensor fusion, and can include more or fewer components than the figure, or combine certain components, or different component arrangements.

[0028] As Figure 1 As shown in the figure, the memory 1005, as a computer storage medium, can include an operating system, a network communication module, a user interface module, and a switch fault prediction program based on multi-sensor fusion.

[0029] In Figure 1 In the switch fault prediction device based on multi-sensor fusion shown in the figure, the network interface 1004 is mainly used to connect a background server and communicate data with the background server. The user interface 1003 is mainly used to connect peripheral devices. The switch fault prediction device based on multi-sensor fusion calls the switch fault prediction program based on multi-sensor fusion stored in the memory 1005 through the processor 1001, and executes the switch fault prediction method based on multi-sensor fusion provided by the embodiment of the present application.

[0030] Based on the above hardware structure, an embodiment of the switch fault prediction method based on multi-sensor fusion is proposed.

[0031] Referring to Figure 2 , Figure 2 The flowchart of an embodiment of the switch fault prediction method based on multi-sensor fusion is shown. An embodiment of the switch fault prediction method based on multi-sensor fusion is proposed.

[0032] In an embodiment, the switch fault prediction method based on multi-sensor fusion comprises the following steps: Step S100, deploying multiple types of sensors at key positions of the switch device, collecting real-time operation state data of the switch device through a sensor controller to obtain original sensor data, and performing noise reduction and space-time alignment processing on the original sensor data to construct a multi-source sensor sample data set.

[0033] It should be noted that the key positions of the switch device refer to the core functional areas of the switch device (such as circuit breakers, contactors, intelligent sockets, etc.) that have the most significant impact on operation safety and fault risk, including contact points / contact positions, wiring terminals, conductive / insulating components, and operating mechanisms. The contact points / contact positions are current on-off contact points that are prone to poor contact or abnormal temperature rise due to oxidation, wear, and arc burning. The wiring terminals are the connection points of wires and devices, which may cause increased contact resistance due to loosening or corrosion. The conductive / insulating components are the conductive loop and insulating materials, which may cause overheating due to overload for the former and leakage or short circuit due to aging failure for the latter. The operating mechanism is a mechanical linkage component, such as a spring or gear, which may cause action failure due to wear or jamming. Multiple types of sensors are deployed on the switch device to collect different dimension operation state parameters, and common types include electrical parameter sensors, physical state sensors, and environmental parameter sensors. The electrical parameter sensors can be current sensors, voltage sensors, and leakage sensors. The physical state sensors can be temperature sensors, vibration sensors, and displacement sensors. The environmental parameter sensors can be humidity sensors and air pressure sensors. The sensor controller is a hardware or software system for unified management of the sensor array, which can be used for data acquisition, signal conditioning, clock synchronization, and signal conditioning. The data acquisition synchronously reads each sensor signal at a fixed frequency. The signal conditioning performs pre-processing such as amplification, filtering, and analog-to-digital conversion (A / D conversion) on the original analog signal. The communication control transmits sensor data to the central processing unit through a bus or wireless protocol. The clock synchronization ensures the consistency of the timestamps of multi-sensor data, providing a basis for subsequent space-time alignment.

[0034] The real-time running state data refers to the original signals collected by sensors in real time, reflecting the current running state of the switch device, such as the real-time current waveform output by the current sensor, the contact temperature value and temperature rise rate output by the temperature sensor, the mechanical vibration acceleration signal output by the vibration sensor, and the voltage effective value and harmonic component output by the voltage sensor. The original sensor data is the initial output signal of the sensor without any processing, which may contain noise interference, time and space deviation, and invalid data. The noise reduction processing refers to the process of removing noise components in the original data through algorithms or hardware means, which can include filtering algorithms, wavelet transform, and outlier detection. The time and space alignment processing refers to the unified calibration of multi-sensor data in time and space dimensions to ensure data fusion analysis; among them, the time alignment refers to synchronizing sensor data of different sampling frequencies to a unified time sequence through timestamp interpolation; the space alignment refers to labeling the corresponding monitoring points of the data combined with the sensor installation position and the physical structure of the device, and establishing the mapping relationship between the data and the device components. The multi-source sensor sample data set is a standardized data set formed by integrating the processed data of various sensors, and the features include multi-dimensional features, sample labels, and time sequence attributes; among them, the multi-dimensional features include electrical parameters, physical states, environmental parameters, and other multi-dimensional indicators; the sample label is the label of "normal state", "early abnormality", "fault warning", etc. which needs to be labeled in the subsequent steps, and is used for model training; the time sequence attribute is to retain the time order of the data, which is used to capture the dynamic change trend of the device state, such as the temperature rise curve and load fluctuation law.

[0035] Specifically, the noise reduction and time-space alignment processing of the original sensor data to construct a multi-source sensor sample data set comprises: Obtaining original sensor data and performing noise reduction processing, the original sensor data including current sensor data, voltage sensor data, temperature sensor data, and mechanical characteristic sensor data; Interpolating the original sensor data of different sampling rates to a unified time scale to construct a time-space aligned original data matrix; Performing wavelet packet decomposition on the original data matrix, and reconstructing each frequency band signal obtained by decomposition into a feature component; Calculating the time domain features and frequency domain features of each feature component, selecting sensitive features related to switch faults through the maximum mutual information coefficient, and constructing a multi-source sensor sample data set.

[0036] It should be noted that wavelet packet decomposition is a signal multi-resolution analysis method based on wavelet transform, which can decompose the original signal into multiple different frequency bands, and realize more fine time-frequency domain analysis of the signal. For signals containing rich frequency components such as current, vibration, etc., such as high-frequency vibration noise generated when the contact is not in good contact, wavelet packet decomposition can decompose the signal into low-frequency (trend item), medium-frequency (normal operation characteristics), high-frequency (fault sensitive characteristics) and different frequency band components. For example, when the mechanical jamming of the switch device occurs, the high-frequency component (>1000Hz) of the vibration signal will be significantly enhanced, and such fault sensitive frequency band can be accurately extracted by wavelet packet decomposition. The band signal is the sub-signal of the original signal in a specific frequency range after wavelet packet decomposition, and each frequency band corresponds to different characteristics of the device operation. In the switch device scenario, the low-frequency band (<100Hz) reflects the trend characteristics of the overall operation of the device, such as long-term temperature rise change and periodic load fluctuation; the medium-frequency band (100-1000Hz) contains stable characteristics such as mechanical vibration and electromagnetic noise during normal operation; the high-frequency band (>1000Hz) is sensitive to transient fault signals such as high-frequency electromagnetic interference generated by arc discharge and impact vibration caused by contact collision. The characteristic component is an independent signal component obtained by reconstructing each frequency band signal, and each component represents the characteristic information of the original signal in a specific time-frequency domain. After wavelet packet decomposition, each frequency band signal can be restored to a time series by inverse transformation, forming independent "characteristic components" that do not overlap each other, such as component 1 corresponding to 10-100Hz vibration signal and component 2 corresponding to 100-500Hz current harmonic signal, which are used for subsequent feature extraction. The time domain feature is a statistical or morphological feature directly extracted from the signal time series, which describes the change rule of the signal in the time dimension. Among them, the commonly used time domain features of the switch device include mean / variance, peak / kurtosis, and rise time / pulse width; the mean / variance reflects the average level and fluctuation degree of the signal, such as abnormal increase of temperature mean and sudden increase of current variance, which may indicate poor contact; the peak reflects the instantaneous extreme value, such as the instantaneous arc current peak when the contact is broken; the kurtosis measures the steepness of the signal waveform, such as the significant increase of vibration signal kurtosis when the mechanical fault occurs; the rise time / pulse width is for switch action signals, such as displacement curve during contact attraction process, and abnormal prolongation of rise time may indicate mechanism jamming.

[0037] The frequency domain feature is a feature extracted after converting the signal from the time domain to the frequency domain by Fourier transform or the like, and describes the frequency component distribution of the signal. The commonly used frequency domain features of the switching device include the main frequency / secondary main frequency, the harmonic amplitude / total harmonic distortion, and the frequency band energy proportion; wherein the main frequency / secondary main frequency refers to the main frequency at which the signal energy is concentrated, for example, the main frequency of the vibration signal in normal operation corresponds to the inherent frequency of the operating mechanism, and the main frequency deviates in the fault; the harmonic amplitude / total harmonic distortion (THD) refers to the harmonic content of the current and voltage signals, and the excessive THD reflects the abnormal power quality, which may be caused by the increase of the contact resistance; the frequency band energy proportion is the proportion of the signal energy in a specific frequency band to the total energy, for example, a sudden increase in the high-frequency band energy proportion may indicate partial discharge. The maximum mutual information coefficient is a statistical index for measuring the correlation between the feature and the fault label, and the feature with the highest contribution to the fault prediction is selected by calculating the mutual information value of the feature variable and the label variable. The greater the mutual information value, the stronger the dependence of the feature on the fault label, for example, the mutual information value of the “temperature rise rate” of the temperature sensor and the “contact oxidation fault” is much higher than that of the “environmental humidity”. By setting a threshold, for example, retaining the top 30% of the features with the highest mutual information, redundant and irrelevant features, such as secondary environmental parameters that have no significant correlation with the fault, can be removed, and the model complexity is reduced. The sensitive feature is a key feature highly related to the switching fault after being screened by the maximum mutual information coefficient, and is the core input for constructing the sample data set. The typical sensitive features of the switching device include electrical, mechanical and thermal characteristics; wherein the electrical includes the contact resistance and the third harmonic amplitude; the mechanical includes the kurtosis of the vibration signal and the contact opening and closing time deviation; and the thermal characteristics include the contact temperature gradient and the temperature rise rate.

[0038] wherein the reconstruction calculation formula of each frequency band signal is:

[0039] wherein, is the reconstructed signal of the jth layer and the kth frequency band, is the low-pass filter coefficient of the wavelet packet decomposition, and is the low-pass filter coefficient used in the reconstruction of the jth layer and the kth frequency band; is the high-pass filter coefficient of the wavelet packet decomposition, and is the high-pass filter coefficient used in the reconstruction of the jth layer and the kth frequency band; is the decomposition signal of the 2kth frequency band of the j-1th layer, and is the low-pass component source for the reconstruction of the kth frequency band of the current layer, and t-n represents the time domain translation of the signal, that is, the time shift term in the convolution operation with the filter coefficient; is the decomposition signal of the 2k+1th frequency band of the j-1th layer, and is the high-pass component source for the reconstruction of the kth frequency band of the current layer.

[0040] It should be noted that the reconstructed signal Reconstructed signal (time-domain waveform) of the jth layer (decomposition scale layer), kth frequency band, which represents the target frequency band signal recovered by combining the low / high frequency components of the (j-1)th layer through a filter. Wherein, j is the decomposition layer number, the higher the layer number, the lower the signal resolution, such as j=1 for the first layer decomposition, j=2 for the second layer; t is the time variable, i.e. the time-domain coordinate of the reconstructed signal. Synthesis low-pass filter coefficient used in the reconstruction of the jth layer, kth frequency band, which corresponds to the dual of the analysis low-pass filter in decomposition, used to extract the low-frequency component of the target frequency band from the low-frequency component of the upper layer. It should be understood that, Finite-length sequence, such as the commonly used Daubechies wavelet filter coefficient length of 4 / 6 / 8, etc.; determines the low-frequency characteristics of the reconstructed signal, such as smoothness, trend item retention ability, for example, the low-frequency trend item of the switch device temperature signal can be reconstructed through the low-pass filter.

[0041] High-pass filter coefficient Synthesis high-pass filter coefficient used in the reconstruction of the jth layer, kth frequency band, which corresponds to the dual of the analysis high-pass filter in decomposition, used to extract the high-frequency component of the target frequency band from the high-frequency component of the upper layer. High-pass filter coefficient The low-pass filter coefficient satisfies the orthogonal relationship, which ensures the conservation of signal energy; sensitive to the high-frequency details of the signal, such as the transient vibration impact signal generated by the collision of the switch contact, which needs to be reconstructed through the high-pass filter. Upper layer decomposition signal Decomposition signal of the (j-1)th layer, 2kth frequency band, low-frequency parent node signal, which is the source of the low-pass component of the current layer k frequency band; wavelet packet decomposition follows a binary tree structure, each frequency band in each layer is decomposed into 2 sub-frequency bands, low frequency 2k and high frequency 2k+1, and the sub-node signal is synthesized by the parent node signal during reconstruction, such as the jth layer k frequency band is composed of 2k and 2k+1 frequency bands of the (j-1)th layer. The time-domain translation term t-n represents the translation of the upper layer signal Or Shifted along the time axis by n units, the essence is the convolution operation of the filter coefficient and the signal, and the weighted superposition is realized through time shift. The time-domain translation term t-n ensures that the filter coefficient and the signal are aligned in the time domain, realizing frequency-selective filtering, such as extracting different features of the signal at different positions through different n weights.

[0042] It should be understood that the low-frequency component in the target frequency band is extracted by weighting and summing the upper layer low-frequency signal Through the low-pass filter The high-frequency component in the target frequency band is extracted by weighting and summing the upper layer high-frequency signal Through the high-pass filter The complete reconstructed signal of the current layer k frequency band is obtained by adding the low-frequency and high-frequency components , to realize the separation and recovery of the specific frequency band characteristics of the original signal.

[0043] By adjusting j (the number of decomposition layers) and k (the frequency band number), the key frequency band related to the fault can be reconstructed directionally: if the contact resistance increases, the mean value of the low-frequency component (<100 Hz) of the current signal will rise, and the trend feature can be extracted by low-pass reconstruction; if the operating mechanism is stuck, the energy of the high-frequency component (>500 Hz) of the vibration signal will surge, and the impact feature can be captured by high-pass reconstruction. The low-pass filter suppresses high-frequency noise such as environmental electromagnetic interference, and retains the trend anomaly of the device state such as gradual temperature rise; the high-pass filter eliminates low-frequency drift such as sensor zero drift, and highlights the instantaneous fault signal such as the instantaneous vibration peak value when the contact is opened or closed.

[0044] In step S200, the inherent attribute data and historical fault record of the switch device are obtained, and a fault label is added to the multi-source sensing sample data set.

[0045] It should be noted that the inherent attribute data reflects the basic data of the physical characteristics, design parameters and static attributes of the switch device, and does not change with the running state, which is the core identifier of the individual difference of the device. The historical fault record records the structured data of the time, type, cause, influence and treatment measures of the past faults of the switch device, which is the core basis for fault label annotation. The fault label is a category identifier given to the multi-source sensing sample data set according to the inherent attribute data and the historical fault record, which is used to indicate the running state or fault type of the device at the corresponding time point, and is the core supervision signal for the training of the supervised learning model.

[0046] Specifically, the inherent attribute data and historical fault record of the switch device are obtained, and a fault label is added to the multi-source sensing sample data set, including: The inherent attribute data and historical fault record of the switch device are obtained, and a fault label is added to the multi-source sensing sample data set according to the historical fault record; If the historical fault record is a serious fault, the fault label is set to -1, and the serious fault includes switch failure or misoperation; If it is a slight fault, the fault label is set to 1, and the slight fault includes an increase in switch contact resistance and a switch mechanical sticking warning; If it is a normal state, the fault label is set to 0; and the fault label is added to the multi-source sensing sample data set; The inherent attribute data includes device model, rated voltage, rated current, running life, and installation environment temperature and humidity.

[0047] It should be noted that the label-1 (serious fault state) marking condition is that when the sample timestamp is before T minutes (such as T=30 minutes) before the occurrence of a historical serious fault to before the fault is repaired, and the associated fault type is a misoperation, the SCADA system action record and sensor mutation data need to be verified. As a strong supervision signal, the training model identifies the misoperation feature combination of the sudden change of the closing coil current and the missing vibration signal, and outputs a high confidence emergency shutdown warning. The label 1 (minor fault state) marking logic is based on the gradual change features before the occurrence of a minor fault in the historical record, such as a continuous 3-day daily increase of the contact resistance of >1%, and when similar features appear, it is marked as an early anomaly, guiding the model to learn the trend correlation of the temperature rise rate and the harmonic amplitude. The label 0 (normal state) is defined as a period without historical fault records, and all sensor data is within the rated threshold, such as current <0.9 times the rated value, temperature <ambient temperature+30℃, and vibration signal main frequency within the design inherent frequency ±5%. As a negative sample subject, it needs to be divided by a time window to avoid class imbalance, such as generating 1 normal sample per hour.

[0048] Step S300, inputting the multi-source sensing sample data set with fault labels into a deep belief network based on an attention mechanism for training to obtain a trained switch fault prediction model.

[0049] It should be noted that the deep belief network is a deep neural network based on a probabilistic graphical model, which is stacked by multiple restricted Boltzmann machines (RBM), has layered learning ability of unsupervised pre-training and supervised fine-tuning, and is suitable for processing high-dimensional, nonlinear multi-source sensing data. The attention mechanism can simulate human visual attention, so that the model dynamically focuses on key time points or key sensor features when processing multi-source data, and suppresses irrelevant information interference, such as reducing the weight of the environmental temperature and humidity sensor during normal operation, and increasing the attention of the vibration sensor before the fault.

[0050] Specifically, the inputting the multi-source sensing sample data set with fault labels into a deep belief network based on an attention mechanism for training to obtain a trained switch fault prediction model comprises: dividing the multi-source sensing sample data set with fault labels into a training set and a test set, and inputting the training set and the test set into the deep belief network; extracting deep nonlinear features of the multi-source sensing sample data set through multi-layer nonlinear mapping of the deep belief network, and the calculation formula is:

[0051] wherein, is the activation output of the lth layer, and are a weight matrix and a bias vector, respectively, is an activation function; It should be noted that, is the activation output vector of the l-th layer, which is the processing result of the input signal by the neurons of the l-th layer, representing the feature vector extracted by the current layer, and the dimension varies with the network level. When l = 1, may include the ratio of temperature to rated current, which reflects the load rate, and the proportion of high-frequency components of the vibration signal, which reflects the degree of mechanical wear; when l = 2, may include the change gradient of the vibration amplitude when the load rate exceeds 80%, which identifies the mechanical stress anomaly caused by overload. Each layer of activation values corresponds to an upgrade of a feature language, from raw data such as sensor voltage values to fault-sensitive features such as contact contact resistance anomaly indicators. is used to describe the connection strength between the features of the previous layer and the current layer, and the essence is to quantify the feature correlation relationship. The row vector represents the attention degree of the i-th neuron of the l-th layer to all features of the upper layer; the element represents the contribution weight of the j-th feature of the upper layer to the i-th feature of the current layer, which can be positive or negative. A positive value indicates a positive correlation, and a negative value indicates an inhibition relationship. For example, if corresponds to the correlation weight of the contactor coil current mutation and the missing vibration signal, indicating that the contribution of these two features to the refusal failure is very high. Through training, the weight matrix will automatically strengthen the causal path weight from coil current anomaly to control loop failure and from control loop failure to refusal, and suppress the weak correlation weight from environmental temperature and humidity fluctuations to irrelevant features. Through unsupervised pre-training initialization and supervised fine-tuning optimization, the weight matrix adapts to the fault feature mode of the switchgear. is the bias vector of the l-th layer, which is equivalent to the threshold offset of the feature activation, allowing the activation function to trigger a response even when the input signal is weak, enhancing the flexibility of the model. For example, a positive bias reduces the activation difficulty, such as for devices with long service life, reducing the activation threshold of the contact wear feature; a negative bias increases the activation threshold, such as for newly commissioned devices, suppressing the false activation of aging-related features. is the activation function, which can be a sigmoid function, a ReLU function, or a hyperbolic tangent function.

[0052] An attention mechanism is introduced to the feature output layer of the deep belief network to assign dynamic weights to different sensor features, where the dynamic weight allocation calculation formula is:

[0053] where, is the attention weight of the i-th sensor feature, is the total number of sensor features, The attention score of the i-th sensor feature, The high-level feature of the i-th sensor, U is the feature mapping matrix, v is the attention weight vector, and c is the attention bias parameter. U, v, and c are all learnable parameters. The exponential function, The hyperbolic tangent activation function; It should be noted that, The high-level feature vector of the i-th sensor, the sensor abstract feature output by the DBN after multiple nonlinear mappings, represents the key information related to the fault in the sensor data. For vibration sensors, It can be the high-frequency vibration energy proportion extracted by the DBN, the time domain entropy value of the closing and opening impact signal, and other fault-sensitive features; for current sensors, It can be the harmonic distortion rate, the current rising slope, and other features reflecting the contact state of the contact. As the input of the attention mechanism, its quality directly depends on the feature extraction ability of the DBN, such as the DBN needing to separate the sharp current features caused by poor contact of the contact in advance from the original current waveform. U is the feature mapping matrix, with dimensions n x d, where d is the dimension of the high-level feature , and n is the dimension of the attention space. U maps the sensor high-level feature from the original feature space to the relevance measurement space dedicated to attention, which essentially encodes the features for fault relevance. Each row of the matrix corresponds to a fault-related feature query mode, such as a mechanical fault-sensitive mode and an electrical fault-sensitive mode. Through training, U will strengthen the feature dimensions related to the fault, such as mapping the kurtosis value of the vibration signal and the effective value of the current to the high activation area. c is the attention bias parameter, which introduces a baseline offset for attention score calculation, similar to a threshold regulator for feature activation: a positive bias reduces the difficulty of attention activation, allowing the model to maintain a certain degree of attention to weakly related features, such as retaining the basic weight of the environmental sensor for newly commissioned equipment; a negative bias increases the attention activation threshold, suppressing noise features, such as reducing the weight of the occasional fluctuation sensor for long-term stable equipment. The hyperbolic tangent activation function outputs in the range [-1, 1], which is suitable for preserving the positive and negative relevance of features in the attention space. Positive values indicate promoting faults, and negative values indicate suppressing faults; the nonlinear characteristic allows the model to capture complex interactions between features, such as the synergistic fault effect of current overload and contact oxidation. The exponential function is used to amplify the differences in attention scores, making the weights of key features significantly higher than those of secondary features.

[0054] The weighted and fused features of different sensors are input into the Softmax classifier for fault type prediction. The network parameters are optimized through the backpropagation algorithm, and the loss function uses cross-entropy:

[0055] wherein, is a cross-entropy loss value, measuring the difference between the switch fault prediction probability and the fault label, the smaller the value, the more accurate the prediction, is the fault label of the e-th sample, if the e-th sample belongs to the d-th fault, then , otherwise ; is the prediction probability of the e-th sample belonging to the d-th fault, N is the total number of samples, and E is the number of fault categories; When the fault prediction accuracy of the deep belief network on the test set no longer improves, stop training, save the network parameters of the deep belief network, and use the deep belief network as a switch fault prediction model.

[0056] Need to be explained, provides a clear supervision signal for multi-classification problems, for example, in switch devices, E=5 typical faults can be defined, and each sample corresponds to only one clear fault label. The prediction probability directly depends on the quality of the weighted fusion features, and is dynamically adjusted by the attention mechanism. If a certain type of fault (such as mechanical jamming) is mainly determined by the vibration sensor feature, the attention weight of the vibration sensor will be automatically increased during training , thereby improving the confidence of the fault. is a cross-entropy loss value, measuring the dissimilarity between the model prediction distribution and the true distribution, the smaller the value, the closer the prediction to the true fault category; when , the loss is 0; when tends to 0, the loss tends to infinity, forming a strong gradient to drive model correction. The total number of samples N represents the total amount of switch device fault samples used for training, including normal state and various fault scenarios such as closing and opening coil fault, contact wear, etc. The samples need to cover different operating conditions and different device models to ensure the model generalization ability. The number of fault categories E is the total number of fault types to be identified, which needs to be defined in advance according to the switch device maintenance regulations.

[0057] Step S400, fuse and infer the real-time collected multi-sensor data through the trained switch fault prediction model to obtain a fault prediction result, and generate a maintenance decision suggestion in combination with inherent attribute data of the switch device.

[0058] It should be noted that the maintenance decision suggestion refers to a specific operation scheme generated by a rule engine or a decision algorithm based on fault prediction results and device inherent attribute data. Specifically, the trained switch fault prediction model is used to fuse and reason the real-time collected multi-sensor data to obtain fault prediction results, and the inherent attribute data of the switch device is combined to generate a maintenance decision suggestion, which includes: The trained switch fault prediction model is used to fuse and reason the real-time collected multi-sensor data to obtain fault prediction results, and the inherent attribute data of the switch device is combined to generate a maintenance decision suggestion, which includes: A health degree evaluation model is constructed according to the fault probability corresponding to the switch fault type, the predicted switch residual life of the switch fault type, and the inherent attribute data of the switch device, and the health degree index of the switch device is output by the health degree evaluation model. According to the health degree index, a maintenance decision suggestion for the switch device is generated, which includes: if the health degree index is less than a first preset health degree index, an emergency maintenance alarm of the switch device is triggered; if the health degree index is greater than or equal to the first preset health degree index and less than a second preset health degree index, a planned maintenance warning of the switch device is issued; if the health degree index is greater than or equal to the second preset health degree index, the switch device is determined to be in a normal operating state, and continues to be monitored, wherein the second preset health degree index is greater than the first preset health degree index.

[0059] It should be noted that the predicted switch residual life refers to the remaining available time of the device from the current state to the occurrence of the target fault based on the fault evolution trend output by the switch fault prediction model, such as the contact wear degree and the insulation aging rate, or the corresponding switch residual available time matched from historical switch use fault data. The health degree evaluation model refers to a device state quantification model that takes fault probability, residual life, and device inherent attribute data as input and is constructed by a multi-index comprehensive evaluation algorithm, which outputs a health degree index in the range of 0-100, which is used to comprehensively represent the operating state health degree of the switch device.

[0060] The calculation formula of the health degree evaluation model is:

[0061] wherein HI is the health degree index of the switch device, RUL is the predicted switch residual life of the switch device, is the rated life of the switch device, is the fault probability of the switch device, is the average temperature of the switch device, is the maximum allowable temperature of the switch device, is a weight coefficient, wherein a contribution weight of a switch residual life to a health index, a contribution weight of a switch failure probability to a health index, a contribution weight of a switch temperature state to a health index, .

[0062] Need to explain, reflects the health margin of the remaining life of the device relative to the designed life, and the higher the ratio indicates the lower the aging degree. characterizes the influence degree of the residual life on the health, and embodies the importance of the aging dimension of the device, converts the failure probability into a safety probability, directly reflecting the reliability of the current operating state, characterizes the influence of the current failure risk on the health, and embodies the priority of the real-time failure dimension of the device, quantifies the influence of temperature on the health of the device, and reflects the safety space of the current temperature from the temperature risk threshold. characterizes the influence of temperature on the health, and embodies the harmfulness of the thermal stress dimension of the device.

[0063] In addition, the embodiment of the present application also provides a storage medium, wherein the storage medium stores a switch failure prediction program based on multi-sensor fusion, and the switch failure prediction program based on multi-sensor fusion is executed by a processor to realize the steps of the switch failure prediction method based on multi-sensor fusion as described above.

[0064] In addition, with reference to Figure 3 , the embodiment of the present application also provides a switch failure early warning system based on multi-sensor fusion, which comprises: a data acquisition module 10, configured to deploy multiple types of sensors at key parts of a switch device, acquire real-time operating state data of the switch device through a sensor controller to obtain original sensing data, and perform noise reduction and space-time alignment processing on the original sensing data to construct a multi-source sensing sample data set; a label labeling module 20, configured to acquire inherent attribute data and historical failure records of the switch device, and add a failure label to the multi-source sensing sample data set; a model training module 30, configured to input the multi-source sensing sample data set with the failure label into a deep belief network based on an attention mechanism for training to obtain a trained switch failure prediction model; an inference decision module 40, configured to fuse and infer the real-time acquired multi-sensor data through the trained switch failure prediction model to obtain a failure prediction result, and generate a maintenance decision suggestion in combination with the inherent attribute data of the switch device.

[0065] Other embodiments or specific implementations of the switch fault early warning system based on multi-sensor fusion described in the present application can refer to the above-mentioned method embodiments, which will not be described here.

[0066] It should be noted that in this paper, the term "includes", "contains" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such process, method, article or device. Without more limitations, the element defined by the sentence "includes a…" does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0067] The above-mentioned embodiment numbers of the present application are only for description, not representing the advantages and disadvantages of the embodiments. In the system module claims listed in several systems, several of these systems can be embodied by the same hardware item. The use of the words first, second, and third does not represent any order, and these words can be interpreted as names.

[0068] From the above description of the embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be realized by software and the necessary general hardware platform, of course, they can also be realized by hardware, but in many cases the former is a better embodiment. Based on such understanding, the technical solutions of the present application or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as a read-only memory image (ROM) / random access memory (RAM), a magnetic disk, an optical disk), including a number of instructions to make a terminal user device (which can be a mobile phone, a computer, a server, an air conditioner, or a network user device, etc.) execute the methods described in various embodiments of the present application.

[0069] The above is only the preferred embodiment of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent process transformation using the content of the present application specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A switch failure prediction method based on multi-sensor fusion, characterized by, The method comprises: Deploying multiple types of sensors at key parts of the switch device, collecting real-time operation state data of the switch device through a sensor controller to obtain original sensing data, performing noise reduction and space-time alignment processing on the original sensing data, and constructing a multi-source sensing sample data set; Obtaining inherent attribute data and historical fault records of the switch device, and adding fault labels to the multi-source sensing sample data set; Inputting the multi-source sensing sample data set with fault labels into a deep belief network based on an attention mechanism for training to obtain a trained switch fault prediction model; Fusing and reasoning the multi-sensor data collected in real time through the trained switch fault prediction model to obtain a fault prediction result, and generating a maintenance decision suggestion in combination with the inherent attribute data of the switch device.

2. The multi-sensor fusion based switch failure prediction method as claimed in claim 1, wherein, The noise reduction and space-time alignment processing on the original sensing data and the construction of the multi-source sensing sample data set comprise: Obtaining original sensing data and performing noise reduction processing, wherein the original sensing data comprises current sensor data, voltage sensor data, temperature sensor data and mechanical characteristic sensor data; Interpolating the original sensing data of different sampling rates to a unified time scale to construct a space-time aligned original data matrix; Performing wavelet packet decomposition on the original data matrix, and reconstructing each frequency band signal obtained by the decomposition into a feature component; Calculating the time domain features and frequency domain features of each feature component, screening sensitive features related to switch faults through the maximum mutual information coefficient, and constructing a multi-source sensing sample data set.

3. The multi-sensor fusion based switch failure prediction method as claimed in claim 2, wherein, The reconstruction calculation formula of each frequency band signal is: ， wherein, is the reconstructed signal of the jth layer and the kth band, is the low-pass filter coefficient of the wavelet packet decomposition, and is the low-pass filter coefficient used in the reconstruction of the jth layer and the kth band; is the high-pass filter coefficient of the wavelet packet decomposition, and is the high-pass filter coefficient used in the reconstruction of the jth layer and the kth band; is the decomposed signal of the (j-1)th layer and the 2kth band, and is the source of the low-pass component in the reconstruction of the kth band of the current layer, represents the time domain translation of the signal, i.e., the time shift term in the convolution operation with the filter coefficient; is the decomposed signal of the (j-1)th layer and the 2k+1th band, and is the source of the high-pass component in the reconstruction of the kth band of the current layer.

4. The multi-sensor fusion based switch failure prediction method as claimed in claim 1, wherein, The obtaining of the inherent attribute data and the historical fault records of the switch device and the adding of the fault labels to the multi-source sensing sample data set comprise: Obtaining the inherent attribute data and the historical fault records of the switch device, and adding fault labels to the multi-source sensing sample data set according to the historical fault records; If the historical fault record is a serious fault, the fault label is set to -1, and the serious fault includes switch refusal or misoperation; If it is a slight fault, the fault label is set to 1, and the slight fault includes an increase in switch contact resistance and a switch mechanical jamming warning; If it is a normal state, the fault label is set to 0; and the fault label is added to the multi-source sensing sample data set; The inherent attribute data includes device model, rated voltage, rated current, operation life and installation environment temperature and humidity. 5.The switch fault prediction method based on multi-sensor fusion according to claim 1, wherein, The inputting of the multi-source sensing sample data set with fault labels into the deep belief network based on the attention mechanism for training to obtain the trained switch fault prediction model comprises: Dividing the multi-source sensing sample data set with fault labels into a training set and a test set, and inputting them into the deep belief network; extracting deep nonlinear features of the multi-source sensing sample data set through multi-layer nonlinear mapping of the deep belief network, and the calculation formula is: ， wherein, is an activation output of the lth layer, and are a weight matrix and a bias vector, respectively, is an activation function; Introducing an attention mechanism at a feature output layer of the deep belief network to assign dynamic weights to different sensor features, wherein the dynamic weight allocation calculation formula is: ， wherein, is an attention weight for the i-th sensor feature, is the total number of sensor features, is an attention score for the i-th sensor feature, is a high-level feature for the i-th sensor, is a feature mapping matrix, is an attention weight vector, is an attention bias parameter, U, v, c are all learnable parameters; is an exponential function, is a hyperbolic tangent activation function; The different sensor features after weight distribution are fused by weighting, the fused features are input into a Softmax classifier for fault type prediction, network parameters are optimized by a back propagation algorithm, and a cross entropy is used as a loss function: ， wherein, is a cross-entropy loss value, measuring the difference between the switch fault prediction probability and the fault label, the smaller the value, the more accurate the prediction, is the fault label of the e-th sample, if the e-th sample belongs to the d-th fault, then , otherwise ; is the prediction probability of the e-th sample belonging to the d-th fault, N is the total number of samples, and E is the number of fault categories; The training is stopped when the fault prediction accuracy of the deep belief network on the test set no longer improves, the network parameters of the deep belief network are saved, and the deep belief network is used as a switch fault prediction model. 6.The switch fault prediction method based on multi-sensor fusion according to claim 1, wherein, The switch fault prediction model after training is used to fuse and reason the real-time collected multi-sensor data to obtain a fault prediction result, and a maintenance decision suggestion is generated in combination with inherent attribute data of the switch device, including: The switch fault prediction model after training is used to fuse and reason the real-time collected multi-sensor data to obtain a fault prediction result, and a maintenance decision suggestion is generated in combination with inherent attribute data of the switch device, including: A health degree evaluation model is constructed according to the fault probability corresponding to the switch fault type, the predicted switch residual life of the switch fault type, and the inherent attribute data of the switch device, and the health degree index of the switch device is output through the health degree evaluation model; According to the health degree index, a maintenance decision suggestion for the switch device is generated, including: if the health degree index is less than a first preset health degree index, an emergency maintenance alarm of the switch device is triggered; if the health degree index is greater than or equal to the first preset health degree index and less than a second preset health degree index, a planned maintenance warning of the switch device is issued; and if the health degree index is greater than or equal to the second preset health degree index, the switch device is determined to be in a normal operating state, and continues to be monitored, wherein the second preset health degree index is greater than the first preset health degree index.

7. The multi-sensor fusion based switch failure prediction method as claimed in claim 6, wherein, The calculation formula of the health degree evaluation model is: ， wherein HI is a health index of the switch device, RUL is a predicted switch remaining useful life of the switch device, is a rated life of the switch device, is a failure probability of the switch device, is an average temperature of the switch device, is a maximum allowable temperature of the switch device, , , is a weight coefficient, wherein characterizes a contribution weight of the switch remaining useful life to the health index, characterizes a contribution weight of the switch failure-free probability to the health index, characterizes a contribution weight of the switch temperature state to the health index, .

8. A switch failure early warning system based on multi-sensor fusion, characterized in that, The switch fault warning system based on multi-sensor fusion includes: A data acquisition module is configured to deploy multiple types of sensors at key parts of a switch device, acquire real-time operating state data of the switch device through a sensor controller to obtain original sensor data, and perform noise reduction and space-time alignment processing on the original sensor data to construct a multi-source sensing sample data set; A label annotation module is configured to obtain inherent attribute data and historical fault records of the switch device, and add fault labels to the multi-source sensing sample data set; A model training module is configured to input the multi-source sensing sample data set with fault labels into a deep belief network based on an attention mechanism for training to obtain a trained switch fault prediction model; An inference decision module is configured to fuse and reason real-time collected multi-sensor data through the trained switch fault prediction model to obtain a fault prediction result, and generate a maintenance decision suggestion in combination with inherent attribute data of the switch device.

9. A switch failure prediction device based on multi-sensor fusion, characterized by, The switch fault prediction device based on multi-sensor fusion comprises a memory, a processor, and a switch fault prediction program based on multi-sensor fusion stored on the memory and executable on the processor, and the switch fault prediction program based on multi-sensor fusion is configured to implement the steps of the switch fault prediction method based on multi-sensor fusion as claimed in any one of claims 1 to 7.

10. A storage medium, characterized by The storage medium stores a switch fault prediction program based on multi-sensor fusion, and the switch fault prediction program based on multi-sensor fusion implements the steps of the switch fault prediction method based on multi-sensor fusion as claimed in any one of claims 1 to 7 when executed by the processor.

Images