Unmanned equipment board-level circuit state monitoring and positioning method based on multiple sensors

Abstract

The invention discloses an unmanned equipment board-level circuit state monitoring and positioning method based on multiple sensors, and relates to the technical field of unmanned equipment state monitoring and evaluation. The problems that an existing unmanned equipment state monitoring method cannot recognize early hidden faults, and an existing network model is insufficient in spatial-temporal feature joint extraction capacity are solved. According to the method, the collected original signals are preprocessed to obtain the multi-dimensional feature data, and the built CNN-LSTM model is utilized to predict the fault state of each monitoring target, so that fault monitoring and positioning of the board-level circuit are realized. The method is mainly applied to unmanned equipment board-level circuit state monitoring.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned equipment status monitoring and evaluation. Background Art

[0002] In recent years, with the development of intelligent and automated technologies, unmanned equipment has been widely used in various fields. Therefore, the status monitoring of the board-level circuits of unmanned equipment has become increasingly important. Traditional status monitoring relies on a single sensor (such as only voltage signal monitoring), which cannot capture the coupled faults of multiple physical fields. The method based on threshold warning has a high false alarm rate and cannot identify early latent faults. Existing network models have insufficient ability to jointly extract spatio-temporal features. Therefore, the above problems need to be solved urgently. Summary of the Invention

[0003] The purpose of the present invention is to solve the problems that the existing status monitoring methods of unmanned equipment cannot identify early latent faults and the existing network models have insufficient ability to jointly extract spatio-temporal features. The present invention provides a method for monitoring and locating the status of board-level circuits of unmanned equipment based on multi-sensors.

[0004] A method for monitoring and locating the status of board-level circuits of unmanned equipment based on multi-sensors, the method comprising:

[0005] S1. Conduct board-level fault injection simulation on the board-level circuit, collect the voltage signals, current signals, and temperature signals of the nodes to be measured of the board-level circuit at each sampling moment, as well as the mechanical vibration signal and motor speed signal of the board-level circuit, and use all the signals at each sampling moment as a training sample. Among them, the signals collected from each node to be measured are different, and the voltage signal, current signal, temperature signal, mechanical vibration signal, and motor speed signal are all used as monitoring targets;

[0006] The voltage signal of the node to be measured is the output voltage signal of the power supply module of the board-level circuit;

[0007] The current signal of the node to be measured is the current signal of the power loop of the board-level circuit;

[0008] The temperature signal of the node to be measured is the temperature signal of the heat dissipation surface of the power device in the board-level circuit;

[0009] Use the fault status injected into all monitoring targets at the corresponding moment of each training sample as a label, and the fault status is faulty, critical fault, or no fault;

[0010] S2. Preprocess each training sample to obtain multi-dimensional feature data;

[0011] S3. Build a CNN-LSTM model and train the CNN-LSTM model using each training sample. The CNN-LSTM model predicts the fault status according to each received training sample.

[0012] S4. Preprocess the voltage signal, current signal, temperature signal, mechanical vibration signal, and motor speed signal of the board-level circuit collected at the current moment. The obtained multi-dimensional feature data is used as the input of the trained CNN-LSTM model, so as to predict the fault status of each monitoring target and realize the fault monitoring and location of the board-level circuit.

[0013] Preferably, the multi-dimensional feature data includes the kurtosis and FFT spectral peak density SPD of the voltage signal, the total harmonic distortion rate THD and sideband energy ratio SBER of the current signal, the heating rate and fluctuation variance of the temperature signal, the vibration kurtosis and waveform factor of the mechanical vibration signal, and the volatility and pulse loss rate of the speed signal.

[0014] Preferably, the implementation method of preprocessing each training sample to obtain multi-dimensional feature data includes:

[0015] Perform median filtering and Z-score normalization on the voltage signal in sequence, and then perform kurtosis calculation and FFT calculation to obtain the kurtosis and FFT spectral peak density SPD of the voltage signal.

[0016] Perform median filtering and Z-score normalization on the current signal in sequence, and then perform total harmonic distortion rate calculation and sideband energy ratio calculation to obtain the total harmonic distortion rate THD and sideband energy ratio SBER of the current signal.

[0017] Perform moving average filtering on the temperature signal, and then perform rate calculation and variance calculation to obtain the heating rate and fluctuation variance of the temperature signal.

[0018] Perform median filtering and Z-score normalization on the mechanical vibration signal in sequence, and then perform kurtosis calculation and waveform factor calculation to obtain the vibration kurtosis and waveform factor of the mechanical vibration signal.

[0019] Perform moving average filtering on the speed signal, and then perform volatility and pulse loss rate calculation to obtain the volatility and pulse loss rate of the speed signal.

[0020] Preferably, the CNN-LSTM model is implemented by cascading a CNN and a bidirectional LSTM in sequence.

[0021] The CNN is used to perform convolution, activation, and pooling processing on the received multi-dimensional feature data in sequence, extract local features and fuse them to obtain a sequence of fused time-series feature vectors.

[0022] The bidirectional LSTM is used to perform bidirectional time series modeling and feature fusion processing on the received sequence of fused time series feature vectors, and obtain the fault state classification result.

[0023] Preferably, the CNN includes 3 branches and processes in parallel in 3 paths;

[0024] The first branch is used to sequentially perform convolution, activation, and pooling processing on the kurtosis of the voltage signal and the spectral peak density SPD of the FFT in the multi-dimensional feature data;

[0025] The second branch is used to sequentially perform convolution, activation, and pooling processing on the total harmonic distortion rate THD and the sideband energy ratio SBER of the current signal in the multi-dimensional feature data;

[0026] The third branch is used to sequentially perform convolution, activation, and pooling processing on the heating rate and fluctuation variance of the temperature signal, the vibration kurtosis and waveform factor of the mechanical vibration signal, and the volatility and pulse loss rate of the rotational speed signal in the multi-dimensional feature data;

[0027] Channel dimension splicing is performed on the sequences of time series feature vectors output by the 3 branches to obtain the fused sequence of time series feature vectors.

[0028] Preferably, in the first branch, the number of convolution kernels is 16, the size of the convolution kernel is 3, the ReLU function is used for activation, the pooling operation is max pooling, and the size of the max pooling is 2 and the stride is 1;

[0029] In the second branch, the number of convolution kernels is 16, the size of the convolution kernel is 3, the ReLU function is used for activation, the pooling operation is average pooling, and the size of the average pooling is 2 and the stride is 2;

[0030] In the third branch, the number of convolution kernels is 32, the size of the convolution kernel is 5, the ReLU function is used for activation, the pooling operation is max pooling, and the size of the max pooling is 2 and the stride is 2.

[0031] Advantages of the present invention:

[0032] The method for monitoring and locating the state of the board-level circuit of unmanned equipment based on multi-sensing according to the present invention specifically constructs a board-level circuit state monitoring system with both multi-dimensional sensing accuracy and dynamic evolution analysis ability through deep fusion and intelligent modeling of multi-source heterogeneous sensing data. It innovatively integrates electrical parameters such as voltage, current, and temperature with mechanical signals such as vibration and rotational speed, not only breaking through the perception blind area of traditional single-dimensional monitoring, but also generating a refined data set covering critical states through fault injection simulation, providing a training environment for the model to approximate the real fault evolution law; on this basis, a CNN-LSTM model is constructed to capture local features such as voltage transient distortion and current high-frequency ripple through convolution, and combine long short-term memory networks to analyze long-range correlations such as temperature slow drift and vibration spectrum time-varying characteristics, realizing precise decoupling of spatio-temporal coupled fault modes; in addition, the collaborative design of a three-class warning mechanism and a board-level positioning strategy can not only predict the critical failure risk at the device level, but also analyze the causal link of compound faults from the system perspective, providing a perfect solution for the state monitoring of the board-level circuit of unmanned equipment. A critical fault is set, and the critical fault is between having a fault and not having a fault. When a fault occurs or is about to occur, a fault warning signal is immediately sent, so as to realize the operation state monitoring and fault warning of the unmanned equipment.

[0033] The present invention uses a deep learning method to evaluate, judge, and predict the state of the board-level circuit of unmanned equipment, processes the collected data information in real time, and timely discovers abnormal operation states. Further, through the model prediction ability, abnormalities can be discovered before the unmanned equipment fails, and a warning signal can be sent in time to inform the management personnel to perform maintenance in advance;

[0034] The deep learning model constructed by the present invention has strong versatility, that is, the CNN-LSTM model, which can continuously optimize and adjust the model parameters according to the latest state of the unmanned equipment and the newly collected data, and timely adapt to the state changes of the unmanned equipment. Brief Description of the Drawings

[0035] Figure 1 is a schematic diagram of the principle of the method for monitoring and locating the state of the board-level circuit of unmanned equipment based on multi-sensing according to the present invention;

[0036] Figure 2 is a schematic diagram of the specific structure of the CNN-LSTM model. Detailed Embodiments

[0037] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0038] It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0039] A method for monitoring and positioning the state of a board-level circuit of an unmanned equipment based on multi-sensors, the method includes:

[0040] S1. Perform board-level fault injection simulation on the board-level circuit, collect the voltage signals, current signals, temperature signals of the nodes to be measured of the board-level circuit at each sampling moment, as well as the mechanical vibration signal and motor speed signal of the board-level circuit, and use all the signals at each sampling moment as a training sample. Among them, the signals collected by each node to be measured are different, and the voltage signal, current signal, temperature signal, mechanical vibration signal and motor speed signal are all used as monitoring targets;

[0041] The voltage signal of the node to be measured is the output voltage signal of the power supply module of the board-level circuit;

[0042] The current signal of the node to be measured is the current signal of the power loop of the board-level circuit;

[0043] The temperature signal of the node to be measured is the temperature signal of the heat dissipation surface of the power device in the board-level circuit;

[0044] Use the fault states injected into all monitoring targets at the corresponding moments of each training sample as labels, and the fault states are faulty, critical fault or no fault;

[0045] S2. Preprocess each training sample to obtain multi-dimensional feature data;

[0046] S3. Build a CNN-LSTM model, and use each training sample to train the CNN-LSTM model. The CNN-LSTM model predicts the fault state according to each received training sample;

[0047] S4. Preprocess the voltage signal, current signal, temperature signal, mechanical vibration signal and motor speed signal of the board-level circuit collected at the current moment, and use the obtained multi-dimensional feature data as the input of the trained CNN-LSTM model, so as to predict the fault states of each monitoring target and realize the fault monitoring and positioning of the board-level circuit.

[0048] In this embodiment, the proposed method for monitoring and positioning the board-level circuit state of unmanned equipment based on multi-sensing can realize the operation state evaluation and fault prediction of unmanned equipment through board-level circuit state monitoring. Specifically, through the deep fusion and intelligent modeling of multi-source heterogeneous sensing data, a board-level circuit state monitoring system with both multi-dimensional sensing accuracy and dynamic evolution analysis ability is constructed. It innovatively integrates electrical parameters such as voltage, current, and temperature with mechanical signals such as vibration and rotational speed, not only breaking through the perception blind area of traditional single-dimensional monitoring, but also generating a refined data set covering critical states through fault injection simulation, providing a training environment for the model to approximate the real fault evolution law; on this basis, a CNN-LSTM model is constructed to capture local features such as voltage transient distortion and current high-frequency ripple through convolution, and combine long short-term memory networks to analyze long-range correlations such as temperature slow drift and vibration spectrum time-varying characteristics, realizing the accurate decoupling of spatio-temporal coupled fault modes; in addition, the collaborative design of a three-class warning mechanism and a board-level positioning strategy can not only predict the critical failure risk at the device level, but also analyze the causal link of compound faults from a system perspective, providing a perfect solution for the board-level circuit state monitoring of unmanned equipment. A critical fault is set, and the critical fault is between having a fault and no fault. When a fault occurs or is about to occur, a fault warning signal is immediately sent, so as to realize the operation state monitoring and fault warning of unmanned equipment.

[0049] In specific applications, during the signal acquisition stage, voltage, current, rotational speed, vibration, and temperature sensors are placed at corresponding key nodes of the board-level circuit. The key nodes are the nodes to be measured. Among them, the voltage sensor is deployed at the output end of the power supply module to monitor the voltage fluctuation of the power supply; the current sensor is connected in series in the power loop to capture the current harmonics and transient changes of the power loop; the rotational speed sensor is deployed at the output shaft end of the motor to measure the rotational speed fluctuation of the motor in real time; the vibration sensor is deployed at the connection of key structures to monitor mechanical vibration; the temperature sensor is deployed on the heat dissipation surface of power devices to monitor the heat distribution. After the deployment, the original signals are synchronously collected at a sampling rate of 1 to 20 kHz.

[0050] Furthermore, the multi-dimensional feature data includes the kurtosis and FFT spectral peak density SPD of the voltage signal, the total harmonic distortion rate THD and sideband energy ratio SBER of the current signal, the heating rate and fluctuation variance of the temperature signal, the vibration kurtosis and waveform factor of the mechanical vibration signal, and the volatility and pulse loss rate of the rotational speed signal.

[0051] In this preferred embodiment, multi-dimensional feature data is selected to characterize whether there is a fault. Its advantage lies in that it constructs a cross-domain feature topology network with both physical interpretability and dynamic sensitivity; the kurtosis of the voltage signal and the spectral peak density SPD of the FFT not only accurately capture the transient spikes at the output end of the power supply module and the spread of low-frequency ripples, but also analyze the deterioration trajectory of power supply stability by quantifying the energy concentration of the main frequency; the total harmonic distortion rate THD and the sideband energy ratio SBER of the current signal not only map the distortion level of the harmonic components in the power loop, but also demodulate the sideband energy distribution, forming a dual diagnosis of the dynamic load characteristics of power devices; the heating rate and fluctuation variance of the temperature signal depict the thermal evolution process from the dual perspectives of differentiation and integration, which can not only sensitively capture the abnormal delay of the thermal inertia of the heat dissipation surface, but also quantify the random perturbation of the steady-state thermal distribution; the kurtosis and form factor of the mechanical vibration signal realize the physical traceability of mechanical-electrical coupling faults through the quantification of time-domain impact components and the normalization of energy distribution; the volatility and pulse loss rate of the rotational speed signal are analyzed through time-frequency joint analysis, which not only analyzes the non-linear distortion of the dynamic response of the motor load, but also verifies the integrity of the encoder signal link. These features form a cross-validation mechanism through cross-domain physical associations, which not only suppresses the risk of misjudgment caused by single-sensor noise interference, but also extracts the causal links of compound faults from multi-scale association rules, and finally establishes a highly robust classification boundary within the critical state fuzzy domain, providing a feature basis with both diagnostic accuracy and engineering applicability for the sub-healthy state identification of unmanned equipment under complex working conditions.

[0052] Furthermore, the implementation methods for preprocessing each training sample to obtain multi-dimensional feature data include:

[0053] After performing median filtering and Z-score normalization on the voltage signal in sequence, kurtosis calculation and FFT calculation are carried out to obtain the kurtosis of the voltage signal and the spectral peak density SPD of the FFT;

[0054] After performing median filtering and Z-score normalization on the current signal in sequence, total harmonic distortion rate calculation and sideband energy ratio calculation are carried out to obtain the total harmonic distortion rate THD and the sideband energy ratio SBER of the current signal;

[0055] After performing moving average filtering on the temperature signal, rate calculation and variance calculation are carried out to obtain the heating rate and fluctuation variance of the temperature signal;

[0056] After performing median filtering and Z-score normalization on the mechanical vibration signal in sequence, kurtosis calculation and form factor calculation are carried out to obtain the vibration kurtosis and form factor of the mechanical vibration signal;

[0057] After performing moving average filtering on the rotational speed signal, volatility and pulse loss rate calculation are carried out to obtain the volatility and pulse loss rate of the rotational speed signal.

[0058] In this preferred embodiment, an implementation method for obtaining multi-dimensional data is presented. Through a preprocessing chain architecture adapted to heterogeneous signal characteristics and a feature engineering strategy driven by physical mechanisms, a data representation paradigm with deep coupling of fault sensitivity and anti-interference is constructed: for high-frequency transient signals such as voltage, current, and vibration, a rigid denoising combination of median filtering - Z-score normalization is adopted, which not only suppresses the interference of pulse noise on kurtosis calculation but also retains the nanosecond-level spike features of the precursor to switch device breakdown. At the same time, the frequency-domain diagnosis basis for power supply stability is formed by quantifying the degree of main frequency energy leakage through the FFT spectral peak density SPD; for slow-varying signals such as temperature and rotational speed, a flexible smoothing strategy of moving average filtering is adopted. While filtering out the random temperature fluctuations caused by air flow disturbances, the mutation inflection point of the heating rate is extracted through differential operation, and the boundary oscillation of the thermal steady state is captured by variance calculation; the total harmonic distortion rate THD and sideband energy ratio SBER of the current signal are explicitly modeled mathematically. The former is directly related to the harmonic energy accumulation of the power circuit switch loss, and the latter demodulates the abnormal change in the sideband energy ratio to achieve cross-verification of electrical-control compound faults; the vibration waveform factor and rotational speed pulse loss rate are further decoupled in the time-frequency-amplitude three dimensions. The former identifies the periodic energy accumulation of mechanical resonance with the effective value - peak value ratio, and the latter diagnoses the signal integrity of the encoder link with the pulse sequence dispersion degree. Finally, a cross-domain feature topology network is formed in the CNN-LSTM model, providing a data engine with both physical interpretability and engineering robustness for the condition monitoring of unmanned equipment.

[0059] See Figure 2 , and further presents the specific structure of the constructed CNN-LSTM model. The CNN-LSTM model is implemented by cascading a CNN and a bidirectional LSTM in sequence;

[0060] The CNN is used to sequentially perform convolution, activation, and pooling processing on the received multi-dimensional feature data, extract local features, and fuse them to obtain a sequence of fused time-series feature vectors; specifically,

[0061] The CNN includes 3 branches and processes in parallel in 3 ways;

[0062] The first branch is used to sequentially perform convolution, activation, and pooling processing on both the kurtosis and FFT spectral peak density SPD of the voltage signal in the multi-dimensional feature data;

[0063] The second branch is used to sequentially perform convolution, activation, and pooling processing on both the total harmonic distortion rate THD and sideband energy ratio SBER of the current signal in the multi-dimensional feature data;

[0064] The third branch is used to sequentially perform convolution, activation, and pooling processing on the heating rate and fluctuation variance of the temperature signal, the vibration kurtosis and waveform factor of the mechanical vibration signal, and the volatility and pulse loss rate of the rotational speed signal in the multi-dimensional feature data;

[0065] The time series feature vector sequences output by the three branches are concatenated in the channel dimension to obtain the fused time series feature vector sequence.

[0066] The bidirectional LSTM is used to perform bidirectional time series modeling and feature fusion processing on the received fused time series feature vector sequence to obtain the fault state classification result.

[0067] In this preferred embodiment, the specific structure of the constructed CNN-LSTM model is given. Through the collaborative architecture of heterogeneous feature partitioning - spatio-temporal joint modeling, it realizes the efficient analysis of multi-physical field signals and the deep coupling of fault evolution laws: the model implements feature domain decoupling with a three-branch parallel CNN architecture - the voltage branch accurately captures the high-frequency ripple transient output by the power supply module and the frequency domain energy dispersion characteristics of the FFT spectral peak density SPD through a narrowband convolution kernel, and the current branch synchronously analyzes the accumulated switching losses of the power loop through the cross-channel convolution fusion of the total harmonic distortion THD and the sideband energy ratio SBER; the temperature-vibration-rotation branch uses a wide time series convolution kernel to perform cross-modal feature extraction on the sudden change in the heating rate of the temperature signal, the periodic aggregation of the vibration waveform factor, and the discretization event of the rotational speed pulse loss rate, revealing the non-linear correlation of the mechanical-thermal-electric coupling effect; the time series features output by the three branches are concatenated in the channel dimension to form a high-dimensional vector of cross-domain fusion, which is input into the bidirectional LSTM for forward and backward time series modeling - the forward network captures the progressive accumulation law of fault precursors, and the backward network traces the propagation link of fault consequences. Through the gating mechanism, the contribution degree of cross-branch features is dynamically weighted to achieve the decoupling of the primary and secondary factors of compound faults; this architecture realizes a closed-loop design of partitioning extraction and spatio-temporal fusion, which not only ensures the inference speed but also greatly improves the fault location accuracy, providing an intelligent diagnosis solution with both theoretical completeness and engineering robustness for the early fault detection, multi-fault concurrent analysis, and critical state warning of unmanned equipment under electromagnetic-mechanical-thermal multi-field coupling disturbances.

[0068] Furthermore, in the first branch, the number of convolution kernels is 16, the size of the convolution kernel is 3, the ReLU function is used for activation, the pooling operation is max pooling, and the size of the max pooling is 2 and the stride is 1;

[0069] In the second branch, the number of convolution kernels is 16, the size of the convolution kernel is 3, the ReLU function is used for activation, the pooling operation is average pooling, and the size of the average pooling is 2 and the stride is 2;

[0070] In the third branch, the number of convolution kernels is 32, the size of the convolution kernel is 5, the ReLU function is used for activation, the pooling operation is max pooling, and the size of the max pooling is 2 and the stride is 2.

[0071] In this preferred embodiment, by further defining the number of convolutional kernels, the size of convolutional kernels, and the specific method of pooling operation in each branch, the advantage is that through the physical adaptability of heterogeneous signal feature extraction and the Pareto optimal design of computing resources and information entropy, an optimal solution space for multi-modal feature extraction is constructed: the voltage branch uses 16 3×1 narrowband time-domain convolutional kernels in conjunction with max pooling with a stride of 1 to capture the high-frequency transient distortion output by the power module at sub-sampling level time resolution. At the same time, the significance of kurtosis mutation is enhanced through the local extreme value retention strategy to avoid the smoothing effect of the traditional pooling stride on weak fault features; although the current branch still uses 16 3×1 convolutional kernels, anti-aliasing dimensionality reduction of harmonic components is achieved through average pooling with a stride of 2 - while suppressing the interference of EMI noise on THD calculation, the energy distribution stability of the fundamental wave and low-order harmonics is maintained through sliding mean operation, greatly reducing the analytical error of the modulation feature of the sideband energy ratio SBER; the multi-modal branch covers the long-period correlation of temperature, vibration, and rotational speed signals with 32 5×1 wide-time convolutional kernels, and its max pooling with a stride of 2 highlights the mutation features of mechanical shock events and thermal runaway critical points through an extreme value screening mechanism. At the same time, the combination mode diversity of 6D input signals is adapted through the high-dimensional feature channel number; with the cooperation of the three, the use of the ReLU activation function ensures the stable propagation of cross-branch gradients. The differential pooling strategy enables the fused vector after channel splicing to have both transient anomaly significance and steady-state trend smoothness, providing a high-dynamic-range input for the time-series modeling of bidirectional LSTM and providing a solution with both sensitivity and robustness for real-time multi-physical field diagnosis in the edge computing scenario of unmanned equipment.

[0072] Although the present invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the present invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed, as long as they do not depart from the spirit and scope of the present invention as defined by the appended claims. It should be understood that different dependent claims and the features described herein can be combined in a manner different from that described in the original claims. It should also be understood that the features described in connection with a single embodiment can be used in other described embodiments.

Claims

1. A multi-sensor based unmanned equipment board-level circuit status monitoring and positioning method, characterized in that: The method includes: S1, perform board-level fault injection simulation on the board-level circuit, collect voltage signals, current signals and temperature signals of the nodes to be tested of the board-level circuit at each sampling time, as well as mechanical vibration signals and motor speed signals of the board-level circuit, and use all the signals at each sampling time as a training sample, wherein the signals collected by each node to be tested are different, and the voltage signal, current signal, temperature signal, mechanical vibration signal and motor speed signal are all used as monitoring targets; The voltage signal of the node to be tested is the output voltage signal of the board-level circuit power module; The current signal of the node to be measured is the current signal of the power loop of the board-level circuit; The temperature signal of the node to be measured is the temperature signal of the heat dissipation surface of the power device in the board-level circuit; The fault state injected into all monitoring targets at the time corresponding to each training sample is used as a label, and the fault state is faulty, critically faulty, or not faulty; S2, preprocessing each training sample to obtain multi-dimensional feature data; S3, construct a CNN-LSTM model, use each training sample to train the CNN-LSTM model, and the CNN-LSTM model predicts the fault status according to each received training sample; S4. Preprocess the voltage signal, current signal, temperature signal, mechanical vibration signal and motor speed signal of the board-level circuit collected at the current moment, and use the obtained multi-dimensional feature data as the input of the trained CNN-LSTM model to predict the fault status of each monitoring target and realize fault monitoring and positioning of the board-level circuit.

2. The multi-sensor based unmanned equipment board-level circuit status monitoring and positioning method according to claim 1 is characterized in that: The multidimensional feature data include the kurtosis and FFT spectrum peak density SPD of the voltage signal, the total harmonic distortion THD and sideband energy ratio SBER of the current signal, the heating rate and fluctuation variance of the temperature signal, the vibration kurtosis and waveform factor of the mechanical vibration signal, and the fluctuation rate and pulse loss rate of the speed signal.

3. The multi-sensor based unmanned equipment board-level circuit status monitoring and positioning method according to claim 1, characterized in that: The implementation methods of preprocessing each training sample to obtain multi-dimensional feature data include: After the voltage signal is subjected to median filtering and Z-score standardization, kurtosis calculation and FFT calculation are performed to obtain the kurtosis of the voltage signal and the FFT spectrum peak density SPD; After the current signal is subjected to median filtering and Z-score standardization, the total harmonic distortion rate and the sideband energy ratio are calculated to obtain the total harmonic distortion rate THD and the sideband energy ratio SBER of the current signal; After the temperature signal is processed by sliding average filtering, the rate calculation and variance calculation are performed to obtain the heating rate and fluctuation variance of the temperature signal; After the mechanical vibration signal is subjected to median filtering and Z-score standardization in sequence, the kurtosis and waveform factor are calculated to obtain the vibration kurtosis and waveform factor of the mechanical vibration signal; After the speed signal is processed by sliding average filtering, the fluctuation rate and pulse loss rate are calculated to obtain the fluctuation rate and pulse loss rate of the speed signal.

4. The multi-sensor based unmanned equipment board-level circuit status monitoring and positioning method according to claim 2 is characterized in that: The CNN-LSTM model is implemented by cascading CNN and bidirectional LSTM in sequence; CNN is used to perform convolution, activation, and pooling on the received multi-dimensional feature data in sequence, extract local features, and fuse them to obtain a fused time series feature vector sequence; The bidirectional LSTM is used to perform bidirectional time series modeling and feature fusion processing on the received fused time series feature vector sequence to obtain the fault state classification result.

5. The multi-sensor based unmanned equipment board-level circuit status monitoring and positioning method according to claim 4 is characterized in that: CNN consists of 3 branches, and 3 branches are processed in parallel; The first branch is used to sequentially perform convolution, activation and pooling processing on the kurtosis of the voltage signal and the FFT spectrum peak density SPD in the multi-dimensional feature data; The second branch is used to sequentially perform convolution, activation and pooling processing on the total harmonic distortion rate THD and the sideband energy ratio SBER of the current signal in the multi-dimensional feature data; The third branch is used to sequentially perform convolution, activation and pooling processing on the heating rate and fluctuation variance of the temperature signal, the vibration kurtosis and waveform factor of the mechanical vibration signal, and the fluctuation rate and pulse loss rate of the speed signal in the multidimensional feature data; The time series feature vector sequences output by the three branches are concatenated in the channel dimension to obtain a fused time series feature vector sequence.

6. The multi-sensor based unmanned equipment board-level circuit status monitoring and positioning method according to claim 5, characterized in that: In the first branch, the number of convolution kernels is 16, the convolution kernel size is 3, the ReLU function is used for activation, the pooling operation is the maximum pooling, and the maximum pooling size is 2 and the step size is 1; In the second branch, the number of convolution kernels is 16, the convolution kernel size is 3, the ReLU function is used for activation, the pooling operation is average pooling, and the average pooling size is 2 and the step size is 2; In the third branch, the number of convolution kernels is 32, the convolution kernel size is 5, the ReLU function is used for activation, the pooling operation is maximum pooling, and the maximum pooling size is 2 and the step size is 2.

Images