Airport luggage system speed reducer fault prediction and diagnosis method based on digital twinning

Abstract

The invention provides an airport luggage system speed reducer fault prediction and diagnosis method based on digital twinning, which relates to the technical field of fault prediction and comprises the steps of acquiring real-time operation data of a speed reducer, establishing a digital twinning model and updating the model in real time, constructing a multi-dimensional feature vector and then performing feature extraction and fusion through deep learning, identifying a fault type and evaluating the degree. And optimizing model parameters by using a particle swarm optimization algorithm to establish a residual life prediction model, generating early warning information and determining a maintenance scheme. Accurate prediction and early warning of faults of the speed reducer are realized, the maintenance cost is reduced, and the operation reliability of an airport luggage system is improved.

Description

Technical Field

[0001] The present invention relates to fault prediction technology, and in particular to a fault prediction and diagnosis method for a speed reducer of an airport baggage system based on digital twin. Background Art

[0002] The airport baggage handling system is a key infrastructure for modern air transportation. Its efficient and stable operation is crucial for ensuring the normal takeoff and landing of flights and the safe and timely delivery of passengers' baggage. In the airport baggage system, the speed reducer, as the core component of the transmission system, undertakes the important functions of transmitting power and adjusting speed. Any fault in the speed reducer may cause the baggage system to stop, resulting in baggage handling delays, flight delays, and even significant economic losses and negative social impacts.

[0003] Traditional maintenance of speed reducers in airport baggage systems mainly relies on regular inspections and passive reactive repairs. This maintenance method usually conducts equipment inspections at fixed intervals or repairs the equipment only after obvious faults have occurred. With the development of intelligent manufacturing and Internet of Things technologies, predictive maintenance technologies have gradually received attention. By real-time monitoring and analyzing the equipment status, potential faults can be predicted, and maintenance measures can be taken in advance.

[0004] Currently, the fault prediction and diagnosis technology for speed reducers in airport baggage systems has the following defects and deficiencies: First, most existing fault diagnosis methods analyze based on single-type sensor data, making it difficult to comprehensively reflect the operating status of the speed reducer, resulting in inaccurate diagnosis results or omission of certain types of faults; Second, traditional fault prediction models have insufficient adaptability to changes in equipment operating environment and load, and cannot dynamically adjust prediction parameters according to actual operating conditions, affecting prediction accuracy; Finally, existing technologies lack effective data fusion and information processing mechanisms, and fail to fully utilize the correlation between multi-source heterogeneous data, resulting in limited accuracy of prediction models and difficulty in coping with various complex fault modes that may occur at different operating stages of the speed reducer. Summary of the Invention

[0005] The embodiments of the present invention provide a fault prediction and diagnosis method for a speed reducer of an airport baggage system based on digital twin, which can solve the problems in the prior art.

[0006] In the first aspect of the embodiments of the present invention, a fault prediction and diagnosis method for a speed reducer of an airport baggage system based on digital twin is provided, including: Obtaining real-time operation data of the speed reducer of the airport baggage system, where the real-time operation data includes speed reducer rotation speed data, vibration data, temperature data, current data, and noise data; Establishing a digital twin model of the speed reducer of the airport baggage system, and real-time updating the digital twin model through the real-time operation data; Construct a multi-dimensional feature vector based on the real-time operation data, use a deep learning algorithm to extract and fuse features of the multi-dimensional feature vector to obtain a fused feature vector, and input the fused feature vector into the digital twin model for fault type identification and fault degree assessment; Synchronize the results of the fault type identification and the fault degree assessment with the digital twin model, dynamically optimize the parameters of the digital twin model through a particle swarm optimization algorithm, establish a remaining life prediction model for the reducer, and predict the fault development trend of the reducer according to the remaining life prediction model of the reducer to generate a fault warning message; Determine a maintenance plan according to the fault warning message. The maintenance plan includes maintenance time, maintenance content and maintenance priority, and send the maintenance plan to the control terminal of the airport baggage system.

[0007] Constructing a multi-dimensional feature vector based on the real-time operation data, using a deep learning algorithm to extract and fuse features of the multi-dimensional feature vector to obtain a fused feature vector, and inputting the fused feature vector into the digital twin model for fault type identification and fault degree assessment includes: Preprocess the real-time operation data to obtain sampled data, and construct a multi-dimensional feature vector including time domain features and frequency domain features based on the sampled data; Input the multi-dimensional feature vector into a multi-channel convolutional neural network. Each channel of the multi-channel convolutional neural network processes the time domain features and the frequency domain features respectively, extracts features of the multi-dimensional feature vector through a convolutional layer and an activation function to obtain a feature map, and fuses the features of the feature map through an attention mechanism to obtain a fused feature vector. The attention mechanism calculates attention weights based on query vectors, key vectors and value vectors; Based on the digital twin model using a deep neural network structure, identify the fault type of the fused feature vector through a softmax function and a ReLU activation function to obtain the result of the fault type identification; Calculate a health index based on the deviation between the fused feature vector and the pre-stored healthy state baseline feature. The health index is calculated by dividing the difference between the fused feature vector and the healthy state baseline feature by the healthy state baseline feature. Input the health index into a fuzzy inference system, and the fuzzy inference system quantitatively evaluates the fault degree based on a Gaussian membership function to obtain the result of the fault degree assessment.

[0008] Based on the digital twin model using a deep neural network structure, identify the fault type of the fused feature vector through a softmax function and a ReLU activation function to obtain the result of the fault type identification includes: Input the fused feature vector into a deep neural network, which includes an input layer, a hidden layer, and an output layer. The hidden layer performs non-linear mapping processing using the ReLU activation function, and the output layer uses the softmax function to calculate the probability distribution of the fault types. Determine the fault type according to the probability distribution result, and synchronize the data of the fault type with the digital twin model. Verify and evaluate the fault type through the digital twin model to obtain the result of fault type identification.

[0009] Synchronize the results of the fault type identification and the fault degree evaluation with the digital twin model. Dynamically optimize the parameters of the digital twin model through the particle swarm optimization algorithm, including: Input the results of the fault type identification and the fault degree evaluation into the digital twin model. Establish a state mapping relationship between the results of the fault type identification, the results of the fault degree evaluation, and the digital twin model through a feature mapping function. Calculate the state synchronization error between the digital twin model and the actual system, and the state synchronization error is measured by the Frobenius norm. Construct an optimization objective function based on the state mapping relationship and the state synchronization error. The optimization objective function uses an adaptive weight coefficient to perform weighted fusion on the state mapping relationship and the state synchronization error, and the optimization objective function is used to optimize the parameter vector of the digital twin model. Adopt a particle swarm algorithm with dynamic inertia weight. Calculate the dynamic neighborhood radius of the optimization objective function based on the iteration number and the contraction factor. Obtain the particle fitness value through error fusion. Calculate the neighborhood influence factor according to the neighborhood center parameter, multiply it by the basic inertia weight to obtain the dynamic inertia weight, and update the particle velocity through a linear combination of the local optimal position and the global optimal position.

[0010] Adopt a particle swarm algorithm with dynamic inertia weight. Calculate the dynamic neighborhood radius of the optimization objective function based on the iteration number and the contraction factor. Obtain the particle fitness value through error fusion. Calculate the neighborhood influence factor according to the neighborhood center parameter, multiply it by the basic inertia weight to obtain the dynamic inertia weight, and update the particle velocity through a linear combination of the local optimal position and the global optimal position, including: Calculate the dynamic neighborhood radius based on the current iteration number, the maximum iteration number, and the contraction factor. Combine the synchronization error, performance error, and regularization error of the digital twin model to calculate the particle fitness value. Perform exponential weighted fusion on the dynamic neighborhood radius and the particle fitness value to obtain an adaptive neighborhood radius. Select the digital twin model parameters that meet the neighborhood constraint conditions and have a better optimization effect than the current particle according to the adaptive neighborhood radius as the neighborhood individuals, and calculate the neighborhood center parameters weighted by the reciprocal of the optimization effect of the neighborhood individuals. The neighborhood center parameters are used to guide the optimization direction of the digital twin model parameters; Calculate the neighborhood influence factor by the distance ratio between the neighborhood center parameters and the current digital twin model parameters and the global optimal parameters, and multiply the neighborhood influence factor by the basic inertia weight to obtain the dynamic inertia weight integrating neighborhood information. The basic inertia weight decreases nonlinearly with the iteration process; Update the particle velocity of the digital twin model parameters based on the dynamic inertia weight integrating neighborhood information.

[0011] Establish a remaining life prediction model for the reducer, and predict the fault development trend of the reducer according to the remaining life prediction model for the reducer, and generate fault warning information including: Obtain the time points and state data in the historical fault data of the reducer, establish a remaining life prediction model for the reducer based on the time points and the state data, and optimize the prediction model parameters by minimizing the sum of the absolute errors between the predicted life values and the true life values; Construct a state evolution equation based on the remaining life prediction model for the reducer, input the current state data of the reducer into the state evolution equation, and obtain the future multi-step state prediction values through recursive iteration. Combine the future multi-step state prediction values with the process noise to construct a prediction confidence interval. The prediction confidence interval is used to characterize the uncertainty range of the fault development trend; Combine the prediction confidence interval with the current remaining life estimate value to construct a composite warning index including a life ratio term and a state change rate term. The life ratio term is the ratio of the current remaining life to the life threshold, and the state change rate term is the ratio of the state change speed to the speed threshold; When the composite warning index is greater than the preset warning threshold, generate fault warning information based on the current remaining life estimate value and the fault development trend prediction result.

[0012] Construct a state evolution equation based on the remaining life prediction model for the reducer, input the current state data of the reducer into the state evolution equation, and obtain the future multi-step state prediction values through recursive iteration. Combining the future multi-step state prediction values with the process noise to construct a prediction confidence interval includes: Establish a state evolution equation based on the remaining life prediction model for the reducer, substitute the historical state sequence into the state evolution equation, calculate the sum of the squared errors between the predicted state values and the actual state values to obtain the sum of squared errors, and determine the parameters of the state evolution equation by minimizing the sum of squared errors; Substitute the current state data and the predicted remaining life value at the current moment into the state evolution equation to obtain the predicted state value at the next moment, and repeat the recursive iteration until the multi-step state prediction values for the preset number of steps are obtained; Take the partial derivative of the state variable with respect to the state evolution equation to obtain the Jacobian matrix, multiply the Jacobian matrix by the multi-step state prediction values and accumulate them to obtain the predicted error accumulation; Multiply the predicted error accumulation by the Jacobian matrix to obtain the predicted uncertainty, calculate the square root of the predicted uncertainty and multiply it by the confidence level coefficient to obtain the confidence interval width, and add and subtract the confidence interval width from the multi-step state prediction values respectively to obtain the prediction confidence interval.

[0013] In the second aspect of the embodiments of the present invention, an electronic device is provided, including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0014] In the third aspect of the embodiments of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0015] The beneficial effects of the present application are as follows: The method for predicting and diagnosing the faults of the reducer of the airport baggage system based on digital twin provided by the present invention realizes the accurate prediction and diagnosis of the faults of the reducer by establishing a digital twin model and combining deep learning algorithms, effectively improving the prediction accuracy and diagnosis efficiency.

[0016] The present invention uses the particle swarm optimization algorithm to dynamically optimize the parameters of the digital twin model and establish a remaining life prediction model, which can scientifically predict the fault development trend of the reducer, so as to realize early warning, avoid the shutdown of the baggage system caused by sudden equipment failures, and reduce the operating losses of the airport.

[0017] The present invention automatically determines the maintenance plan according to the fault warning information, including the maintenance time, content and priority, realizes the precise maintenance and preventive maintenance of the reducer of the baggage system, improves the reliability and service life of the equipment, reduces the maintenance cost, and ensures the stable and efficient operation of the airport baggage system. Description of the Drawings

[0018] Figure 1 It is a schematic flowchart of the method for predicting and diagnosing the faults of the reducer of the airport baggage system based on digital twin in the embodiments of the present invention; Figure 2Flowchart of the device fault type recognition method based on a deep neural network according to an embodiment of the present invention; Figure 3 Flowchart of the digital twin model parameter optimization based on particle swarm optimization according to an embodiment of the present invention; Figure 4 Comparison analysis diagram of the fault degree evaluation errors under different optimization methods according to an embodiment of the present invention; Figure 5 Comparison analysis diagram of the fault recognition accuracies under different working condition types according to an embodiment of the present invention. Detailed implementation manners

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are only a part rather than all of the embodiments of the present invention. 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.

[0020] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments may be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0021] Figure 1 Schematic flowchart of the fault prediction and diagnosis method for the reducer of the airport baggage system based on digital twin according to an embodiment of the present invention, as Figure 1 shown, the method includes: Obtain the real-time operation data of the reducer of the airport baggage system, where the real-time operation data includes reducer rotation speed data, vibration data, temperature data, current data, and noise data; Establish a digital twin model of the reducer of the airport baggage system, and update the digital twin model in real time through the real-time operation data; Construct a multi-dimensional feature vector based on the real-time operation data, use a deep learning algorithm to perform feature extraction and fusion on the multi-dimensional feature vector to obtain a fused feature vector, and input the fused feature vector into the digital twin model for fault type recognition and fault degree evaluation; Synchronize the results of the fault type recognition and the fault degree evaluation with the digital twin model, dynamically optimize the parameters of the digital twin model through a particle swarm optimization algorithm, establish a remaining life prediction model for the reducer, and predict the fault development trend of the reducer according to the remaining life prediction model for the reducer to generate a fault warning message; Determine a maintenance plan based on the fault warning information. The maintenance plan includes maintenance time, maintenance content, and maintenance priority, and send the maintenance plan to the control terminal of the airport baggage system.

[0022] In an alternative embodiment, a multi-dimensional feature vector is constructed based on the real-time operation data, and a deep learning algorithm is used to perform feature extraction and fusion on the multi-dimensional feature vector to obtain a fused feature vector. Inputting the fused feature vector into the digital twin model for fault type identification and fault degree assessment includes: Preprocess the real-time operation data to obtain sampled data, and construct a multi-dimensional feature vector including time-domain features and frequency-domain features based on the sampled data; Input the multi-dimensional feature vector into a multi-channel convolutional neural network. Each channel of the multi-channel convolutional neural network processes the time-domain features and the frequency-domain features respectively. Feature extraction is performed on the multi-dimensional feature vector through a convolutional layer and an activation function to obtain a feature map, and the feature map is subjected to feature fusion through an attention mechanism to obtain a fused feature vector. The attention mechanism calculates attention weights based on a query vector, a key vector, and a value vector; Based on the digital twin model, adopt a deep neural network structure, and perform fault type identification on the fused feature vector through a softmax function and a ReLU activation function to obtain the result of fault type identification; Calculate a health index based on the deviation between the fused feature vector and a pre-stored healthy state baseline feature. The health index is calculated by dividing the difference between the fused feature vector and the healthy state baseline feature by the healthy state baseline feature. Input the health index into a fuzzy inference system, and the fuzzy inference system quantitatively evaluates the fault degree based on a Gaussian membership function to obtain the result of fault degree assessment.

[0023] Preprocess the collected real-time operation data to obtain sampled data. The real-time operation data may include various sensor data such as the vibration signal, temperature, pressure, and sound of the device. In a specific example, a sampling frequency of 20 kHz can be adopted to collect 10 seconds of vibration signal data, obtaining 200,000 data points. The preprocessing includes denoising, normalization, and segmentation. Wavelet transform threshold method is used for denoising to filter out the noise components with an amplitude lower than 0.05 in the signal; normalization is achieved by subtracting the mean and dividing by the standard deviation to make the data distributed within the range of [-1, 1]; segmentation divides 200,000 data points into 100 segments, with 2,000 data points in each segment.

[0024] Construct a multi-dimensional feature vector based on the preprocessed sampled data. The time-domain features include statistical features such as mean, standard deviation, peak value, peak-to-peak value, skewness, and kurtosis. These statistics are calculated for each segment of data. For example, for a certain segment of vibration data, the mean is 0.02, the standard deviation is 0.35, the peak value is 1.85, the peak-to-peak value is 3.42, the skewness is 0.15, and the kurtosis is 2.78. The frequency-domain features are obtained by performing a fast Fourier transform on the sampled data and include the main frequency, frequency band energy distribution, spectral peak, etc. In the embodiment, the 0-10 kHz frequency band is equally divided into 10 sub-frequency bands, and the energy ratio of each sub-frequency band is calculated to form a frequency-domain feature vector. For example, the frequency band energy distribution of a certain device is [0.05, 0.12, 0.25, 0.40, 0.08, 0.04, 0.03, 0.01, 0.01, 0.01]. The time-domain features and frequency-domain features are combined into a multi-dimensional feature vector, and the vector dimension is 16 (6 time-domain features plus 10 frequency-domain features).

[0025] Input the constructed multi-dimensional feature vector into a multi-channel convolutional neural network for feature extraction and fusion. The network contains two independent channels, which process the time-domain features and frequency-domain features respectively. The time-domain feature channel contains 3 convolutional layers, and the size of the convolutional kernel in each layer is 3, and the number of convolutional kernels is 32, 64, and 128 respectively; the frequency-domain feature channel contains 2 convolutional layers, the size of the convolutional kernel is 2, and the number of convolutional kernels is 16 and 32 respectively. After each convolutional layer, a ReLU activation function is connected to introduce non-linearity. For example, the first convolutional layer converts the 6-dimensional time-domain features into 32 feature maps, and each feature map captures different time-domain feature patterns.

[0026] An attention mechanism is used for feature fusion, and the attention mechanism calculates the attention weights based on the query vector, key vector, and value vector. In the example, the output features of the time-domain channel are used as the query vector, and the output features of the frequency-domain channel are used as the key vector and value vector. Calculate the similarity between the query vector and the key vector to obtain the attention scores, such as [0.12, 0.28, 0.35, 0.15, 0.10]. The attention scores are normalized by the softmax function to weights, such as [0.10, 0.30, 0.40, 0.12, 0.08]. The weights are weighted and summed with the value vector to obtain a fused feature vector containing the information of both channels, and the dimension is 64.

[0027] Based on the digital twin model, a deep neural network structure is used to identify the fault type of the fused feature vector. The network contains 3 fully connected layers, with the number of nodes being 128, 64, and the number of fault types (e.g., 10 types) respectively. The ReLU activation function is used in the first two fully connected layers, and the softmax function is used in the last layer to output the probabilities of each fault type. For example, for a certain test sample, the probability distribution output by the network is [0.01, 0.02, 0.85, 0.05, 0.02, 0.01, 0.01, 0.01, 0.01, 0.01], indicating that the probability that this sample belongs to the 3rd type of fault is 85%.

[0028] The health index is calculated based on the deviation between the fused feature vector and the pre-stored baseline features of the healthy state. The baseline features of the healthy state are obtained by extracting the features of the long-term operation data of healthy devices. For example, the baseline feature vector of a certain device's healthy state is [0.02, 0.30, 1.50, 3.00, 0.10, 2.50, 0.05, 0.10, 0.20, 0.35, 0.10, 0.05, 0.05, 0.03, 0.03, 0.04]. The health index is calculated by dividing the difference between the fused feature vector and the baseline features of the healthy state by the baseline features of the healthy state. The obtained health index vector describes the degree of deviation of all aspects of the device's performance, such as [0.00, 0.17, 0.23, 0.14, 0.50, 0.11, 0.00, 0.20, 0.25, 0.14, 0.20, 0.20, 0.40, 0.67, 0.67, 0.75].

[0029] The health index is input into a fuzzy inference system for fault degree assessment. The fuzzy inference system maps the health index to three fault degree levels: minor, medium, and severe based on the Gaussian membership function. Based on the above health index vector, the fault degree membership degrees output by the fuzzy inference system are [0.65, 0.30, 0.05], indicating that the probability that the device is in a minor fault state is 65%, the probability of being in a medium fault state is 30%, and the probability of being in a severe fault state is 5%.

[0030] In an alternative embodiment, based on the digital twin model, a deep neural network structure is adopted, and the softmax function and the ReLU activation function are used to identify the fault type of the fused feature vector. The results of the fault type identification include: The fused feature vector is input into a deep neural network. The deep neural network includes an input layer, a hidden layer, and an output layer. The hidden layer uses the ReLU activation function for non-linear mapping processing, and the output layer uses the softmax function to calculate the probability distribution of the fault types; Determine the fault type according to the probability distribution result, synchronize the data of the fault type with the digital twin model, and verify and evaluate the fault type through the digital twin model to obtain the result of fault type recognition.

[0031] The fused feature vector is obtained by fusing multi-dimensional feature information of the device, including vibration features, temperature features, acoustic features, and historical operation data, etc. In a specific example, assume that the dimension of the fused feature vector of an industrial pump is 128 dimensions, including 64-dimensional vibration features, 32-dimensional temperature features, and 32-dimensional historical operation data features.

[0032] The structure design of the deep neural network includes an input layer, hidden layers, and an output layer. The number of neurons in the input layer is the same as the dimension of the fused feature vector, which is 128 neurons in this example. The hidden layers adopt a multi-layer design, specifically including three hidden layers, with the number of neurons being 256, 128, and 64 respectively. The number of neurons in the output layer corresponds to the types of fault types, which is set to 8 neurons in this example, corresponding to 8 states: bearing fault, blade fracture, cavitation, seal leakage, motor overload, valve jamming, shaft misalignment, and normal state.

[0033] The ReLU activation function is used for non-linear mapping processing. The characteristic of the ReLU activation function is that when the input is positive, the output is equal to the input value; when the input is negative, the output is zero. This design can effectively alleviate the problem of gradient disappearance and improve the training efficiency of the deep network. In practical applications, the ReLU function processes each hidden layer node as follows: First, calculate the sum of the weighted inputs received by the node, then compare this value with zero, and take the larger value of the two as the output of the node.

[0034] For the first hidden layer, each node receives the sum of the products of 128 input values and the corresponding weights, and then passes it to the next layer after being processed by the ReLU function. For example, for the first node in the first hidden layer, its input is the sum of the products of the 128-dimensional fused feature vector and the 128 weight parameters corresponding to this node plus the bias term. Assume the calculation result is 2.35, then the output after being processed by the ReLU function is still 2.35; if the calculation result is -0.87, then the output is 0. This processing method enables the network to learn complex non-linear relationships and improve the ability to extract fault features.

[0035] The softmax function is used to calculate the probability distribution of the fault types. The softmax function converts the raw values of each node in the output layer into probability form, making the sum of the values of all output nodes equal to 1. The value of each node represents the probability of the corresponding fault type. In specific implementation, first calculate the raw output values of each node in the output layer, then perform exponential operation on the raw value of each node, and then divide the exponential value by the sum of the exponential values of all nodes to obtain the probability value of the corresponding fault type for this node.

[0036] Suppose after the neural network calculation, the raw values of 8 nodes in the output layer are [2.1, 0.8, -0.5, 1.2, 0.3, -0.2, 0.9, 0.4] respectively. After being processed by the softmax function, it is converted into probability form as [0.42, 0.11, 0.03, 0.17, 0.07, 0.04, 0.12, 0.08]. This means the probability of bearing fault is 0.42, the probability of blade fracture is 0.11, and so on. Based on these probability values, the system identifies the current fault type as bearing fault because its corresponding probability value is the highest.

[0037] After completing the identification of the fault type, this embodiment synchronizes the identification result with the digital twin model. The digital twin model is an accurate mapping of the physical device in the virtual environment, containing information such as the geometric model, physical characteristics, and operating status of the device. The data synchronization process updates the fault type information to the digital twin model in real time, enabling the model to reflect the current state of the device.

[0038] When a bearing fault is identified, the system transmits this information to the digital twin model to update the status identifier of the bearing component in the model. On the graphical interface, the affected bearing component may be marked in red or other warning colors, and at the same time, information such as the fault type, fault probability, and occurrence time is displayed. In addition, the system will also adjust relevant parameters in the model according to the fault type, such as simulating the impact of the bearing fault on the vibration, temperature, and efficiency of the device.

[0039] The digital twin model is used to verify and evaluate the fault type. The digital twin model constructs the normal and abnormal behavior patterns of the device based on physical laws and historical data, and can be used to verify the rationality of the fault identification result. The verification and evaluation mainly include the following aspects: Compare the identified fault type with the possible faults predicted by the digital twin model. Suppose the identification result is a bearing fault, and the digital twin model also predicts the highest possibility of abnormality in the bearing area based on the current operating parameters, which enhances the credibility of the identification result.

[0040] Analyze the consistency between the fault type and the current operating state of the device. For example, if a bearing fault is identified and the vibration spectrum of the device indeed shows abnormal bearing characteristic frequencies, and the temperature sensor also detects an increase in temperature at the bearing, then the accuracy of the fault identification is further confirmed.

[0041] Verify by combining historical fault data. For example, this device or similar devices have experienced bearing faults under similar operating conditions and shown similar characteristics, which also supports the current fault identification result.

[0042] Through the above verification and evaluation, the system finally confirms the result of the fault type identification and triggers corresponding alarms, maintenance suggestions or automatic protection measures as needed. For example, for the confirmed bearing fault, the system may recommend replacing the bearing at the next planned downtime, and at the same time adjust the operating parameters to reduce the bearing load and extend its remaining service life.

[0043] Figure 2 The following is a flowchart of the method for identifying the fault type of a device based on a deep neural network according to an embodiment of the present invention: This flowchart shows a complete fault type identification process. First, the fused feature vector is input into the deep neural network structure for processing. The deep neural network includes three main layers: the input layer is responsible for receiving the fused feature vector data; the hidden layer uses the ReLU activation function for non-linear mapping processing to enhance the feature extraction and expression ability of the network; the output layer uses the Softmax activation function to calculate the probability distribution of the fault type, mapping the feature space to the fault type probability space. After obtaining the probability distribution result of the fault type, the system determines the specific fault type according to the maximum probability. Then, the identified fault type is synchronized with the digital twin model to achieve the state unity of the physical device and the digital model. Finally, the system will verify and evaluate the fault type predicted by the digital twin model, and verify the accuracy of the model by comparing the actual fault situation with the predicted result, so as to output the final fault type identification result. This deep learning-based fault diagnosis method can effectively identify various fault modes of complex industrial devices through multi-level feature extraction and non-linear mapping.

[0044] In an alternative embodiment, synchronizing the results of the fault type identification and the fault degree evaluation with the digital twin model, and dynamically optimizing the parameters of the digital twin model through the particle swarm optimization algorithm includes: Input the results of the fault type identification and the results of the fault degree assessment into the digital twin model. Establish a state mapping relationship between the results of the fault type identification, the results of the fault degree assessment, and the digital twin model through a feature mapping function. Calculate the state synchronization error between the digital twin model and the actual system, and the state synchronization error is measured by the Frobenius norm. Construct an optimization objective function based on the state mapping relationship and the state synchronization error. The optimization objective function uses an adaptive weight coefficient to perform weighted fusion on the state mapping relationship and the state synchronization error, and the optimization objective function is used to optimize the parameter vector of the digital twin model. Adopt a particle swarm algorithm with dynamic inertia weight. Calculate the dynamic neighborhood radius of the optimization objective function based on the number of iterations and the contraction factor. Obtain the particle fitness value through error fusion. Calculate the neighborhood influence factor according to the neighborhood center parameter, multiply it by the basic inertia weight to obtain the dynamic inertia weight, and update the particle velocity through a linear combination of the local optimal position and the global optimal position.

[0045] Input the results of the fault type identification and the results of the fault degree assessment into the digital twin model. The system establishes a state mapping relationship between the results of the fault type identification, the results of the fault degree assessment, and the digital twin model through a feature mapping function. For example, assume that the fault identification result of an engine system shows a bearing overheating fault, and the fault degree assessment result is a 72% damage level. The system will map this information to the corresponding state variables of the digital twin model through the feature mapping function. Specifically, the feature mapping function can adopt a multi-layer perceptron structure. The input layer receives the fault type and degree information, the hidden layer performs non-linear transformation through an activation function, and the output layer generates the mapping value of the corresponding digital twin model state variable.

[0046] Calculate the state synchronization error between the digital twin model and the actual system, and this error is measured by the Frobenius norm. For example, if the temperature sensor of the actual system shows that the bearing temperature is 95°C, while the predicted value of the digital twin model is 92°C, the actual value of the pressure sensor is 2.8 MPa, and the model predicted value is 3.0 MPa, then the state synchronization error is calculated by the Frobenius norm of the error matrix of all state variables. In practical applications, for an industrial system with 15 state variables, the state synchronization error usually needs to be controlled within 5% to ensure the reliability of the model.

[0047] Based on the above state mapping relationship and state synchronization error, the system constructs an optimization objective function. This function uses an adaptive weight coefficient to perform weighted fusion on the state mapping relationship and state synchronization error. The adaptive weight coefficient is dynamically adjusted according to the convergence situation in the current iteration process. For example, when the state synchronization error is large, the system will automatically increase the weight coefficient of the synchronization error term, making the optimization process pay more attention to reducing the difference between the model and the actual system. In an application case of a factory equipment monitoring system, the initial weight of the state mapping relationship is set to 0.4, and the weight of the state synchronization error is set to 0.6. As the optimization progresses, the system gradually adjusts to 0.35 and 0.65 to accelerate model convergence.

[0048] After constructing the optimization objective function, the particle swarm algorithm with dynamic inertia weight is used to optimize the parameter vector of the digital twin model. First, the dynamic neighborhood radius of the optimization objective function is calculated based on the number of iterations and the contraction factor. For example, in the initial stage of the optimization process, the system sets a relatively large neighborhood radius such as 0.8. As the number of iterations increases, when the iteration reaches the 50th time, the neighborhood radius may shrink to 0.5, and further shrink to 0.3 at the 100th iteration. This helps the algorithm to perform global search first and then local fine search.

[0049] The particle fitness value is obtained through error fusion. In practical applications, such as in a power system fault diagnosis scenario, the initial population is set to 50 particles, and each particle represents a group of possible model parameter combinations. The system calculates the objective function value corresponding to each particle as the fitness. For example, the fitness value of parameter combination A is 0.25, and the fitness value of parameter combination B is 0.18, indicating that combination B is better.

[0050] The neighborhood influence factor is calculated according to the neighborhood center parameter. For each particle, the system determines other particles within its neighborhood and calculates the neighborhood influence factor based on the parameter values of the central particle. For example, when the parameter value of a certain particle is close to the local optimal solution, its neighborhood influence factor may be 0.85, while when it is far away, it may be 0.65.

[0051] The dynamic inertia weight is obtained by multiplying the neighborhood influence factor by the basic inertia weight. Assuming that the basic inertia weight is set to 0.7, then according to the above example, the dynamic inertia weights are 0.595 and 0.455 respectively. During the optimization process, the basic inertia weight usually linearly decays from 0.9 to 0.4 to balance the global search and local convergence capabilities.

[0052] Update the particle velocity through the linear combination of the local optimal position and the global optimal position. For example, in the current iteration, the position of a certain particle is [0.5, 0.3, 0.8], the velocity is [0.1, -0.05, 0.02], the local optimal position is [0.6, 0.25, 0.75], and the global optimal position is [0.55, 0.2, 0.85]. Then the system calculates a new velocity vector by multiplying the dynamic inertia weight by the current velocity and adding the weighted terms for moving towards the local optimal and global optimal positions, thereby updating the particle position.

[0053] Through the above method, the system can effectively synchronize the fault information with the digital twin model, continuously optimize the model parameters, and improve the model's representation ability of the actual fault state. In an application case of a chemical plant, after adopting this method, the fault prediction accuracy rate has increased from 83% to 91%, and the fault warning lead time has been extended from an average of 2 hours to 3.5 hours, providing more sufficient preparation time for equipment maintenance.

[0054] Figure 3 The flowchart for optimizing the parameters of the digital twin model based on particle swarm optimization according to the embodiments of the present invention is as follows: This flowchart describes a complete process for optimizing the parameters of the digital twin model. First, the system inputs the fault type recognition result and the fault degree evaluation result into the digital twin model, and establishes the corresponding relationship between the fault recognition result and the model through the feature mapping function. At the same time, the system calculates the state mapping relationship between the fault degree evaluation result and the digital twin model, and uses the Frobenius norm to calculate the state synchronization error between the model and the actual system. On this basis, an optimization objective function is constructed. This function adopts an adaptive weight system to perform weighted fusion of the state mapping relationship and the state synchronization error for optimizing the parameter vector of the digital twin model. Finally, the system uses the particle swarm algorithm with dynamic inertia weight for optimization and solution. It calculates the dynamic neighborhood radius of the optimization objective function through the number of iterations and the contraction factor, obtains the particle fitness value through error fusion, calculates the neighborhood influence factor according to the neighborhood center parameter, multiplies it by the basic inertia weight to obtain the dynamic inertia weight, updates the particle velocity through the linear combination of the local optimal position and the global optimal position, and finally realizes the optimization of the model parameters. This method based on particle swarm optimization can effectively improve the parameter accuracy and state synchronization performance of the digital twin model.

[0055] In an optional implementation manner, the particle swarm algorithm with dynamic inertia weight calculates the dynamic neighborhood radius of the optimization objective function based on the number of iterations and the contraction factor, obtains the particle fitness value through error fusion, calculates the neighborhood influence factor according to the neighborhood center parameter, multiplies it by the basic inertia weight to obtain the dynamic inertia weight, and updates the particle velocity through the linear combination of the local optimal position and the global optimal position, including: Calculate the dynamic neighborhood radius based on the current iteration number, the maximum iteration number, and the contraction factor. Combine the synchronization error, performance error, and regularization error of the digital twin model to calculate the particle fitness value. Perform exponential weighted fusion on the dynamic neighborhood radius and the particle fitness value to obtain the adaptive neighborhood radius; Select the digital twin model parameters that meet the neighborhood constraint conditions and have a better optimization effect than the current particle as the neighborhood individuals according to the adaptive neighborhood radius. Calculate the neighborhood center parameters weighted by the reciprocal of the optimization effect of the neighborhood individuals. The neighborhood center parameters are used to guide the optimization direction of the digital twin model parameters; Calculate the neighborhood influence factor by the ratio of the distances between the neighborhood center parameters and the current digital twin model parameters and the global optimal parameters. Multiply the neighborhood influence factor by the basic inertia weight to obtain the dynamic inertia weight that fuses neighborhood information. The basic inertia weight decreases non-linearly with the iteration process; Update the particle velocity of the digital twin model parameters based on the dynamic inertia weight that fuses neighborhood information.

[0056] Calculate the dynamic neighborhood radius based on the current iteration number. In the t-th iteration, the system calculates the dynamic neighborhood radius R_t according to the current iteration number t, the maximum iteration number T_max, and the preset contraction factor α. Specifically, the dynamic neighborhood radius R_t decreases non-linearly with the increase of the iteration number. For example, when the current iteration number is 50, the maximum iteration number is 200, and the contraction factor is 0.9, the calculated dynamic neighborhood radius is about 0.7 times the initial neighborhood radius, so that the search space gradually shrinks with the progress of the optimization process, enhancing the local search ability.

[0057] Calculate the particle fitness value. In the optimization of the digital twin model, the system combines three types of errors: synchronization error E_sync, performance error E_perf, and regularization error E_reg to calculate the particle fitness value. The synchronization error represents the difference between the output of the digital model and the state of the physical entity. The performance error measures the deviation between the model prediction performance and the target performance. The regularization error is used to prevent overfitting. For example, for a digital twin model of a wind turbine, the synchronization error can be the root mean square error between the predicted power and the actual power (such as 5.2 kW), the performance error can be the difference between the model efficiency and the target efficiency (such as 0.03), and the regularization error can be the L2 norm of the model parameters (such as 0.15). The system performs weighted summation of these three errors to obtain the comprehensive fitness value F. The weights can be set to 0.5, 0.3, and 0.2 respectively. Then the comprehensive fitness value of this example is 2.94.

[0058] The dynamic neighborhood radius and the particle fitness value are exponentially weighted and fused to obtain the adaptive neighborhood radius \(R_{adaptive}\). For example, when the dynamic neighborhood radius is 0.7, the particle fitness value is 2.94, and the exponential weighting coefficient is 0.8, the adaptive neighborhood radius is approximately 0.82. Particles with higher fitness values obtain larger neighborhood radii, enhancing their exploration ability.

[0059] According to the adaptive neighborhood radius, the system selects the digital twin model parameters that satisfy the neighborhood constraint conditions and have a better optimization effect than the current particle as neighborhood individuals. The neighborhood constraint condition refers to the Euclidean distance between parameters being less than the adaptive neighborhood radius. For example, when the current particle parameters are [0.25, 0.4, 0.6] and the adaptive neighborhood radius is 0.82, the distance of the particle with parameters [0.3, 0.5, 0.7] is approximately 0.17, which is less than the neighborhood radius, and its fitness value is 2.1, which is less than 2.94 of the current particle. Therefore, it is selected as a neighborhood individual.

[0060] For the selected neighborhood individuals, the system weights them based on the reciprocal of the optimization effect and calculates the neighborhood center parameters. The better the optimization effect (the smaller the fitness value), the greater the corresponding weight. For example, three neighborhood individuals are selected, with fitness values of 2.1, 2.3, and 2.5 respectively, and parameters [0.3, 0.5, 0.7], [0.28, 0.45, 0.65], and [0.27, 0.42, 0.63] respectively. The weighted neighborhood center parameters are [0.29, 0.47, 0.67]. These neighborhood center parameters represent the distribution center of local high-quality solutions and are used to guide the optimization direction of the digital twin model parameters.

[0061] Calculate the distance \(d_{center}\) between the neighborhood center parameters and the current digital twin model parameters, and the distance \(d_{global}\) between the current parameters and the global optimal parameters. For example, when the current parameters are [0.25, 0.4, 0.6], the neighborhood center parameters are [0.29, 0.47, 0.67], and the global optimal parameters are [0.35, 0.55, 0.75], then \(d_{center}\) is approximately 0.11 and \(d_{global}\) is approximately 0.22. The neighborhood influence factor is calculated as the ratio of \(d_{center}\) to \(d_{global}\), which is approximately 0.5. This indicates that the distance between the current particle and the neighborhood center is relatively smaller than the distance to the global optimum, and local information should have a higher influence.

[0062] Multiply the neighborhood influence factor by the basic inertia weight to obtain the dynamic inertia weight \(w_{dynamic}\) that fuses neighborhood information. The basic inertia weight \(w_{base}\) decreases non-linearly with the iteration process. For example, the initial value of \(w_{base}\) is 0.9, which drops to 0.7 when the iteration reaches 1 / 4, to 0.5 when the iteration reaches half, and to 0.3 when the iteration reaches 3 / 4. When the neighborhood influence factor is 0.5 and the basic inertia weight is 0.7, the fused dynamic inertia weight is 0.35, indicating that the current particle is more inclined to adjust towards the neighborhood center and the global optimal direction, reducing the influence of the inertia component.

[0063] Update the particle velocity \(V\) based on the dynamic inertia weight that fuses neighborhood information. For example, the current particle velocity is \([0.05, 0.03, 0.04]\), the dynamic inertia weight is 0.35, the learning factors \(c1 = c2 = 2.0\), the random numbers \(r1 = 0.6\), \(r2 = 0.8\), the current position is \([0.25, 0.4, 0.6]\), the individual optimal position is \([0.28, 0.45, 0.65]\), and the global optimal position is \([0.35, 0.55, 0.75]\). Then the updated velocity is approximately \([0.056, 0.079, 0.084]\). The particle position is correspondingly updated to \([0.306, 0.479, 0.684]\), moving towards the neighborhood center and the global optimal direction.

[0064] The experimental results show that the particle swarm optimization algorithm with this dynamic inertia weight, in the parameter optimization of the digital twin model of wind turbines, compared with the traditional PSO, the convergence speed is increased by about 35%, the final optimization accuracy is increased by about 25%, and it has better stability under different initial conditions.

[0065] Figure 4 This is the comparative analysis chart of the fault degree evaluation error under different optimization methods in the embodiment of the present invention: This figure shows the comparison of the evaluation errors of various optimization methods under different degrees of faults. The horizontal axis in the figure represents the degree of fault (slight fault, moderate fault, severe fault), and the vertical axis represents the relative error percentage of fault evaluation (0 - 15%). From the data, this technical solution (diamond mark) shows the lowest evaluation error under various degrees of faults: 2.9% for slight faults, dropping to 2.1% for moderate faults, and further decreasing to 1.3% for severe faults, with the average error remaining at 2.1%. In contrast, the errors of other methods are significantly higher: the error range of the linearly decreasing weight PSO (triangle mark) is between 3.8% - 5.8%, the error of the standard PSO (square mark) is between 5.2% - 7.9%, and the error of the parameter - free optimization method (circle mark) is the highest, between 7.2% - 11.5%. This technical solution reduces the error by 77.7% compared to the parameter - free optimization solution, 67.7% compared to the standard PSO, and 56.3% compared to the linearly decreasing weight PSO. These data fully demonstrate that this technical solution has significant advantages in the accuracy of fault degree evaluation, especially showing higher precision when dealing with severe faults.

[0066] In an alternative implementation, a remaining - life prediction model of the speed reducer is established, and based on the remaining - life prediction model of the speed reducer, the fault development trend of the speed reducer is predicted, and the generated fault warning information includes: Obtain the time points and status data in the historical fault data of the speed reducer, establish a remaining - life prediction model of the speed reducer based on the time points and the status data, and optimize the prediction model parameters by minimizing the sum of the absolute errors between the predicted life values and the true life values. Based on the remaining - life prediction model of the speed reducer, construct a state evolution equation, input the current state data of the speed reducer into the state evolution equation, obtain the future multi - step state prediction values through recursive iteration calculation, and construct a prediction confidence interval by combining the future multi - step state prediction values with process noise. The prediction confidence interval is used to characterize the uncertainty range of the fault development trend. Combine the prediction confidence interval with the current remaining - life estimate value to construct a composite warning index including a life - ratio term and a state - change - rate term. The life - ratio term is the ratio of the current remaining life to the life threshold, and the state - change - rate term is the ratio of the state - change speed to the speed threshold. When the composite warning index is greater than the preset warning threshold, generate fault warning information based on the current remaining - life estimate value and the fault development trend prediction result.

[0067] Obtain the historical fault data of the reducer, including time points and status data. The historical fault data can be obtained from the reducer monitoring system. For example, obtain the vibration signal data, temperature data, noise data, and oil analysis data, etc. of a reducer in a manufacturing enterprise for three consecutive months. Specifically, the root mean square value of the vibration of the reducer gearbox can be selected as the status feature, and 10 complete data sequences from normal operation to failure are extracted from the historical data, with each group of data containing the status data of 200 time points.

[0068] Based on the obtained time points and status data, establish a remaining life prediction model for the reducer. In this embodiment, a double-exponential trend model is used as the prediction model, and this model can represent the relationship between the status value and time. To optimize the model parameters, the sum of the absolute errors between the predicted life value and the true life value is minimized through an iterative method. Specifically, set the initial parameter values. For example, initialize the decay coefficient in the model to 0.05 and the growth coefficient to 0.02, and then use the gradient descent method for parameter optimization. In practical applications, after training with 10 groups of historical data, the average prediction error of the model can be controlled within 7.5%, where the decay coefficient is optimized to 0.038 and the growth coefficient is optimized to 0.025.

[0069] Based on the established prediction model, construct a state evolution equation. The state evolution equation describes the relationship of how the equipment state evolves over time. Input the current status data of the reducer into the state evolution equation, and obtain the future multi-step state prediction values through recursive iteration. For example, for a running reducer, its current root mean square value of vibration is 2.8 mm / s, and the state change trend within the next 30 days can be predicted through the state evolution equation. In the recursive calculation process, each prediction value is used as the input for the next calculation. For example, the prediction value on the first day is 2.85 mm / s, and it is used as the input to predict the state value on the second day as 2.91 mm / s, and so on.

[0070] To characterize the uncertainty in the prediction process, combine the future multi-step state prediction values with process noise to construct a prediction confidence interval. The process noise can be determined by analyzing the volatility of historical data. In this embodiment, the standard deviation of the process noise is obtained by calculating the standard deviation of historical data as 0.15 mm / s. The upper limit of the prediction confidence interval is the prediction value plus twice the standard deviation, and the lower limit is the prediction value minus twice the standard deviation. For example, the prediction value on the 10th day is 3.25 mm / s, then the prediction confidence interval is [2.95 mm / s, 3.55 mm / s]. This prediction confidence interval indicates that at a 95% confidence level, the actual state value of the reducer on the 10th day will fall within this interval.

[0071] Combine the predicted confidence interval with the current remaining life estimate to construct a composite warning index. The composite warning index includes a life ratio term and a state change rate term. The life ratio term is the ratio of the current remaining life to the life threshold, and the state change rate term is the ratio of the state change speed to the speed threshold. In this embodiment, based on the prediction model, the remaining life of the current reducer is estimated to be 25 days, and the life threshold is set to 30 days; the current state change speed is 0.06 mm / s / day, and the speed threshold is set to 0.05 mm / s / day. Therefore, the life ratio term is calculated as 25 / 30 = 0.833, and the state change rate term is calculated as 0.06 / 0.05 = 1.2. The composite warning index can be obtained by weighted average. In this example, the weights are set to 0.6 and 0.4 respectively, and the composite warning index value is 0.833×0.6 + 1.2×0.4 = 0.98.

[0072] Set the warning threshold to 0.95. When the composite warning index value of 0.98 is greater than the warning threshold of 0.95, the system generates a fault warning message. The warning message includes the following content: "Equipment ID: GBX-1025, Current status: Vibration value 2.8 mm / s exceeds the warning line, Estimated remaining life: 25 days, Suggested operation: Scheduled maintenance, Priority: Medium". At the same time, the warning message can also include a state trend chart, intuitively showing the state evolution trend and predicted confidence interval in the next 30 days.

[0073] Through the above method, the remaining life of the reducer can be accurately predicted, and based on the prediction results, a fault warning message can be generated in a timely manner, providing decision support for equipment maintenance, effectively avoiding unscheduled downtime of the equipment, and improving equipment utilization and production efficiency.

[0074] Figure 5 This is a comparative analysis chart of the fault identification accuracy rates under different working conditions in the embodiment of the present invention: This figure shows the comparison of the fault recognition accuracy rates of three different methods under various working condition types. The horizontal axis in the figure represents eight different working condition types (low-speed light load, low-speed heavy load, medium-speed light load, medium-speed medium load, medium-speed heavy load, high-speed light load, high-speed medium load, high-speed heavy load), and the vertical axis represents the fault recognition accuracy rate (70% - 100%). The technical solution of the present invention (circular mark) maintains a high recognition accuracy rate under all working conditions, ranging from 94% at low-speed light load to 87% at high-speed heavy load, and the overall performance is the most stable. The deep learning method (square mark) ranks second, with the accuracy rate fluctuating between 82% and 91%, but the performance significantly drops to 82% under the high-speed heavy load working condition. The traditional regression method (triangle mark) performs the worst, with the overall low accuracy rate and large fluctuations, gradually decreasing from 85% at low-speed light load to 71% at high-speed heavy load, especially with a significant performance decay under high-speed working conditions. The data shows that the technical solution of the present invention maintains a high recognition accuracy rate under various working condition conditions, especially still maintaining good performance under the complex high-speed heavy load working condition, demonstrating strong adaptability and reliability. This stable performance advantage has important practical significance for the fault diagnosis of industrial equipment.

[0075] In an alternative embodiment, a state evolution equation is constructed based on the remaining life prediction model of the speed reducer, and the current state data of the speed reducer is input into the state evolution equation. The future multi-step state prediction values are obtained through recursive iterative calculation. Constructing a prediction confidence interval by combining the future multi-step state prediction values with process noise includes: Based on the remaining life prediction model of the speed reducer, a state evolution equation is established. The historical state sequence is substituted into the state evolution equation, and the sum of the squared errors between the predicted state value and the actual state value is calculated and summed to obtain the sum of squared errors. The parameters of the state evolution equation are determined by minimizing the sum of squared errors; Substitute the current state data and the remaining life prediction value at the current moment into the state evolution equation to obtain the predicted state value at the next moment, and repeat the recursive iteration until the multi-step state prediction values of the preset number of steps are obtained; Take the partial derivative of the state variable of the state evolution equation to obtain the Jacobian matrix, multiply the Jacobian matrix by the multi-step state prediction values and accumulate to obtain the prediction error accumulation; Multiply the prediction error accumulation by the Jacobian matrix to obtain the prediction uncertainty, calculate the square root of the prediction uncertainty and multiply it by the confidence level coefficient to obtain the confidence interval width, and add and subtract the confidence interval width from the multi-step state prediction values respectively to obtain the prediction confidence interval.

[0076] Obtain the historical operation data of the reducer, including vibration signals, temperature data, lubricating oil parameters and other condition monitoring data. Taking the planetary reducer in a certain mining equipment as an example, collect its operation data for three consecutive months, with a sampling frequency of once per hour, including parameters such as vibration root mean square value, temperature, input shaft speed, etc. Through preprocessing the original data, such as denoising and normalization, a standardized historical state sequence is obtained, denoted as X_1, X_2, ..., X_n, where X_i represents the state value of the reducer at the i-th moment.

[0077] Based on the preprocessed historical state sequence, establish a remaining useful life prediction model for the reducer. This model can be constructed using machine learning methods such as BP neural network, support vector regression or random forest. Taking the BP neural network as an example, the number of input layer nodes is set to 5, corresponding to 5 historical state values; the hidden layer is set to two layers, with the number of nodes being 10 and 6 respectively; the output layer has 1 node, corresponding to the predicted remaining useful life value. By training this model, the mapping relationship from the current state to the remaining useful life can be obtained.

[0078] Based on the above remaining useful life prediction model of the reducer, establish a state evolution equation. The state evolution equation describes the transition relationship from the current state to the next moment state, and this equation can be expressed as a functional relationship between the state value, its historical values and the remaining useful life. In specific implementation, take the historical state values X_t, X_(t - 1), ..., X_(t - p) and the current remaining useful life prediction value RUL_t as input variables, and X_(t + 1) as the output variable to construct the general form of the state evolution equation.

[0079] To determine the specific parameters of the state evolution equation, substitute the historical state sequence into this equation, calculate the sum of the squared errors between the predicted state value and the actual state value. For example, for each set of adjacent state values in the historical data, substitute the state at the previous moment into the equation to get the predicted value, calculate the square of the difference from the actual observed value, and then sum to obtain the total sum of squared errors. By minimizing this sum of squared errors, the parameter values in the state evolution equation can be determined. In actual calculation, optimization algorithms such as gradient descent method or least squares method can be used to solve for the optimal parameters. Taking the data of a certain reducer as an example, the parameters of the state evolution equation obtained through training are: weight coefficients w_1 = 0.72, w_2 = 0.18, w_3 = 0.05, w_4 = 0.03, w_5 = 0.02, and the coefficient related to the remaining useful life is -0.015.

[0080] After determining the parameters of the state evolution equation, substitute the state data \(X_t\) at the current time \(t\) and the predicted remaining useful life value \(RUL_t\) at the current time into the state evolution equation to obtain the predicted state value \(X_{(t + 1)}\) at the next time \(t+1\). Then, substitute \(X_{(t + 1)}\) and the updated predicted remaining useful life value \(RUL_{(t + 1)}\) into the state evolution equation again to obtain the predicted state value \(X_{(t + 2)}\) at time \(t + 2\). And so on. Through recursive iteration calculation, finally obtain the multi-step state prediction values for the preset number of steps, such as the state prediction sequence for the next 24 hours.

[0081] To evaluate the uncertainty of the prediction results, it is necessary to construct a prediction confidence interval. First, take the partial derivative of the state variable with respect to the state evolution equation to obtain the Jacobian matrix. Specifically, calculate the partial derivative values of the equation with respect to each input state variable to form the elements of the Jacobian matrix. For the above parameter example, the main elements of the Jacobian matrix are 0.72, 0.18, 0.05, 0.03, 0.02, and -0.015.

[0082] Multiply the Jacobian matrix by the multi-step state prediction values and accumulate them to obtain the cumulative prediction error. This step reflects the propagation and cumulative effect of errors in the multi-step prediction process. In specific implementation, for each step of prediction, calculate the product of the Jacobian matrix and the current prediction error, and then accumulate the results of all steps to obtain the total cumulative prediction error.

[0083] Multiply the cumulative prediction error by the Jacobian matrix to obtain the prediction uncertainty. This calculation takes into account the influence of model parameter uncertainty on the prediction results. Calculate the square root of the obtained prediction uncertainty, and then multiply it by the confidence level coefficient (for example, the coefficient corresponding to a 95% confidence level is 1.96) to obtain the confidence interval width. Finally, add and subtract the confidence interval width from the multi-step state prediction values respectively to form the upper and lower boundaries of the prediction results, that is, the prediction confidence interval.

[0084] Taking the 24-hour prediction of a certain reducer as an example, the current state value is 0.65 (after normalization). Through recursive calculation, the predicted state sequence for the next 24 hours is: 0.67, 0.69, 0.72, 0.74...1.05. The corresponding 95% confidence interval widths are: 0.03, 0.05, 0.08, 0.11...0.32. Therefore, the 95% prediction confidence interval for the state value after 24 hours is [0.73, 1.37].

[0085] Through the above method, multi-step prediction of the future state of the reducer can be realized, and the confidence interval of the prediction results can be given, providing a reliable decision-making basis for the preventive maintenance and fault warning of the equipment.

[0086] In the second aspect of the embodiments of the present invention, an electronic device is provided, including: A processor; A memory for storing processor-executable instructions; Wherein, the processor is configured to call the instructions stored in the memory to execute the foregoing method.

[0087] In a third aspect of the embodiments of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the foregoing method is implemented.

[0088] The present invention may be a method, an apparatus, a system, and / or a computer program product. The computer program product may include a computer-readable storage medium having thereon computer-readable program instructions for performing various aspects of the present invention.

[0089] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting and diagnosing reducer faults in airport baggage systems based on digital twins, characterized in that: include: Acquire real-time operating data of the reducer of the airport baggage system, wherein the real-time operating data includes reducer speed data, vibration data, temperature data, current data and noise data; Establishing a digital twin model of the airport baggage system reducer, and updating the digital twin model in real time through the real-time operation data; Constructing a multidimensional feature vector based on the real-time operation data, extracting and fusing the multidimensional feature vector using a deep learning algorithm to obtain a fused feature vector, and inputting the fused feature vector into the digital twin model to perform fault type identification and fault degree assessment; The results of the fault type identification and the fault degree assessment are synchronized with the digital twin model, the parameters of the digital twin model are dynamically optimized by a particle swarm optimization algorithm, a reducer remaining life prediction model is established, the fault development trend of the reducer is predicted according to the reducer remaining life prediction model, and fault warning information is generated; A maintenance plan is determined according to the fault warning information, wherein the maintenance plan includes maintenance time, maintenance content and maintenance priority, and the maintenance plan is sent to the airport baggage system control terminal.

2. The method according to claim 1, characterized in that Constructing a multidimensional feature vector based on the real-time operation data, extracting and fusing the multidimensional feature vector using a deep learning algorithm to obtain a fused feature vector, and inputting the fused feature vector into the digital twin model for fault type identification and fault degree assessment includes: Preprocessing the real-time operation data to obtain sampling data, and constructing a multidimensional feature vector including time domain features and frequency domain features based on the sampling data; Inputting the multidimensional feature vector into a multi-channel convolutional neural network, wherein each channel of the multi-channel convolutional neural network processes the time domain feature and the frequency domain feature respectively, extracting features from the multidimensional feature vector through a convolutional layer and an activation function to obtain a feature map, and performing feature fusion on the feature map through an attention mechanism to obtain a fused feature vector, wherein the attention mechanism calculates an attention weight based on a query vector, a key vector, and a value vector; Based on the digital twin model, a deep neural network structure is adopted to identify the fault type of the fused feature vector through a softmax function and a ReLU activation function to obtain a result of fault type identification; A health index is calculated based on the deviation between the fused feature vector and the pre-stored health status baseline feature. The health index is calculated by dividing the difference between the fused feature vector and the health status baseline feature by the health status baseline feature. The health index is input into a fuzzy inference system. The fuzzy inference system quantitatively evaluates the degree of fault based on a Gaussian membership function to obtain a fault degree evaluation result.

3. The method according to claim 2, characterized in that Based on the deep neural network structure adopted by the digital twin model, the fault type is identified on the fused feature vector through the softmax function and the ReLU activation function, and the results of the fault type identification include: Input the fused feature vector into a deep neural network, the deep neural network includes an input layer, a hidden layer and an output layer, the hidden layer uses a ReLU activation function to perform nonlinear mapping processing, and the output layer uses a softmax function to calculate the probability distribution of the fault type; The fault type is determined according to the probability distribution result, and the fault type is synchronized with the digital twin model. The fault type is verified and evaluated through the digital twin model to obtain the result of fault type identification.

4. The method according to claim 1, characterized in that: The method of synchronizing the results of the fault type identification and the fault degree assessment with the digital twin model and dynamically optimizing the parameters of the digital twin model by using a particle swarm optimization algorithm includes: Inputting the result of the fault type identification and the result of the fault degree assessment into the digital twin model, establishing a state mapping relationship between the result of the fault type identification, the result of the fault degree assessment and the digital twin model through a feature mapping function, calculating a state synchronization error between the digital twin model and the actual system, and measuring the state synchronization error through a Frobenius norm; An optimization objective function is constructed based on the state mapping relationship and the state synchronization error, wherein the optimization objective function uses an adaptive weight coefficient to perform weighted fusion on the state mapping relationship and the state synchronization error, and the optimization objective function is used to optimize the parameter vector of the digital twin model; A particle swarm algorithm with dynamic inertia weight is adopted. The dynamic neighborhood radius of the optimization objective function is calculated based on the number of iterations and the shrinkage factor. The particle fitness value is obtained by error fusion. The field influence factor is calculated according to the neighborhood center parameter, and the dynamic inertia weight is obtained by multiplying it with the basic inertia weight. The particle velocity is updated by a linear combination of the local optimal position and the global optimal position.

5. The method according to claim 4, characterized in that The particle swarm algorithm with dynamic inertia weight is used to calculate the dynamic neighborhood radius of the optimization objective function based on the number of iterations and the shrinkage factor, and the particle fitness value is obtained by error fusion. The field influence factor is calculated according to the neighborhood center parameter, and the dynamic inertia weight is obtained by multiplying it with the basic inertia weight. The particle speed is updated through a linear combination of the local optimal position and the global optimal position, including: The dynamic neighborhood radius is calculated based on the current number of iterations, the maximum number of iterations and the shrinkage factor, the particle fitness value is calculated in combination with the synchronization error, performance error and regularization error of the digital twin model, and the dynamic neighborhood radius is exponentially weighted fused with the particle fitness value to obtain an adaptive neighborhood radius; According to the adaptive neighborhood radius, a digital twin model parameter that satisfies the neighborhood constraint condition and has an optimization effect better than the current particle is selected as a neighborhood individual, and a neighborhood center parameter of the neighborhood individual weighted by the inverse of the optimization effect is calculated. The neighborhood center parameter is used to guide the optimization direction of the digital twin model parameter; The neighborhood influence factor is calculated by the distance ratio between the neighborhood center parameter and the current digital twin model parameter and the global optimal parameter, and the neighborhood influence factor is multiplied by the basic inertia weight to obtain the dynamic inertia weight of the fused neighborhood information, and the basic inertia weight decreases nonlinearly with the iteration process; The particle velocity of the digital twin model parameters is updated based on the dynamic inertia weight of the fused neighborhood information.

6. The method according to claim 1, characterized in that Establishing a reducer remaining life prediction model, predicting the failure development trend of the reducer according to the reducer remaining life prediction model, and generating fault warning information including: Obtaining time points and status data from historical fault data of the reducer, establishing a reduction gear remaining life prediction model based on the time points and the status data, and optimizing the prediction model parameters by minimizing the sum of absolute errors between the predicted life value and the actual life value; A state evolution equation is constructed based on the reducer remaining life prediction model, current state data of the reducer is input into the state evolution equation, future multi-step state prediction values ​​are obtained through recursive iterative calculation, and a prediction confidence interval is constructed by combining the future multi-step state prediction values ​​with process noise, wherein the prediction confidence interval is used to characterize the uncertainty range of the fault development trend; Combining the prediction confidence interval with the current estimated value of remaining life, constructing a composite early warning indicator including a life ratio term and a state change rate term, wherein the life ratio term is the ratio of the current remaining life to the life threshold, and the state change rate term is the ratio of the state change speed to the speed threshold; When the composite warning index is greater than a preset warning threshold, fault warning information is generated based on the current remaining life estimate and the fault development trend prediction result.

7. The method according to claim 6, characterized in that Based on the reducer remaining life prediction model, a state evolution equation is constructed, the current state data of the reducer is input into the state evolution equation, and the future multi-step state prediction value is obtained by recursive iterative calculation. The future multi-step state prediction value is combined with process noise to construct a prediction confidence interval, including: Establishing a state evolution equation based on the reducer remaining life prediction model, substituting the historical state sequence into the state evolution equation, calculating the square error between the predicted state value and the actual state value and summing them to obtain the square error sum, and determining the parameters of the state evolution equation by minimizing the square error sum; Substituting the current state data and the current remaining life prediction value into the state evolution equation to obtain the predicted state value at the next moment, and repeating the recursive iteration until a multi-step state prediction value of a preset number of steps is obtained; The partial derivatives of the state variables of the state evolution equation are obtained to obtain a Jacobian matrix, and the Jacobian matrix is ​​multiplied by the multi-step state prediction value and accumulated to obtain a prediction error accumulation; The prediction error accumulation is multiplied by the Jacobian matrix to obtain the prediction uncertainty, the square root of the prediction uncertainty is calculated and multiplied by the confidence level coefficient to obtain the confidence interval width, and the multi-step state prediction value is added or subtracted from the confidence interval width to obtain the prediction confidence interval.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images