Multi-sensor cooperative fault diagnosis method using data delay

Abstract

The application discloses a multi-sensor cooperative fault diagnosis method using data delay, which extracts the communication transmission delay, internal processing delay and sampling delay of each sensor as basic parameters, adopts a sliding window to obtain a first-level abstract delay feature, extracts a second-level abstract delay feature through a deep learning network, generates a comprehensive delay feature vector through attention weighting fusion, splices all sensor features to construct a cooperative delay matrix, and generates a cooperative fusion feature tensor through self-attention coding and bidirectional attention fusion of time-synchronized multi-source signals, and inputs the tensor into first and second fault diagnosis branch networks to obtain fault probability and classification results, and continuously optimizes the model through incremental learning, so that the accuracy and early fault recognition ability of multi-sensor cooperative fault diagnosis under non-uniform delay conditions are improved through the use of data delay information.

Description

Technical Field

[0001] This application relates to the field of computer networks and information security, specifically to a multi-sensor collaborative fault diagnosis method that utilizes data latency. Background Technology

[0002] In the field of industrial equipment condition monitoring, multi-sensor collaborative diagnosis has been widely used. However, due to differences in the sampling frequency, communication link characteristics, and internal signal processing flow of different sensors, multi-source data generally experience non-uniform and time-fluctuating data delays during acquisition, transmission, and processing. This makes it difficult for the data from each sensor to be precisely aligned in the time domain, resulting in a spatiotemporal mismatch problem. Consequently, this weakens the effectiveness of multi-source information fusion and reduces the diagnostic accuracy and early warning capability for physical faults in equipment (such as bearings, gears, rotors, etc.).

[0003] However, existing technologies generally regard data delay as an undesirable noise interference, and mainly use post-processing methods such as signal interpolation, clock synchronization compensation, or data alignment to eliminate or weaken it. This leads to a focus only on restoring data synchronization, without recognizing that the delay pattern itself may contain information related to the device's operating status. As a result, in practical applications, the actual change characteristics carried by the delay are easily discarded, leading to a significant underutilization of delay information, which in turn restricts the performance of multi-sensor collaborative fault diagnosis. Summary of the Invention

[0004] The purpose of this application is to provide a multi-sensor collaborative fault diagnosis method that utilizes data latency, so as to at least solve some of the problems mentioned in the background art.

[0005] According to one aspect of this application, a multi-sensor collaborative fault diagnosis method utilizing data latency is provided, comprising the following steps: S1. Acquire monitoring data from multiple sensors on the monitored system, synchronize the multiple sensors using a time synchronization module or network time synchronization protocol, add a timestamp to each acquisition record of each sensor, and generate a raw monitoring data sequence with timestamps. S2. Based on the original monitoring data sequence with timestamps, extract the communication transmission delay, internal processing delay, and sampling delay of each sensor as basic delay parameters. Use a sliding window to extract the statistical features of the basic delay parameter sequence within the window to obtain the first-level abstract delay feature vector. Input the basic delay parameter sequence into a pre-trained deep learning network model for feature extraction to obtain the second-level abstract delay feature vector. Weightedly fuse the first-level and second-level abstract delay feature vectors to generate the comprehensive delay feature vector for each sensor. S3. Based on the comprehensive delay feature vectors of all sensors, a multi-sensor collaborative delay feature matrix is ​​constructed by splicing them together in the order of the sensors. At the same time, data cleaning and time alignment preprocessing are performed on the original monitoring data sequence with timestamps to obtain a time-synchronized multi-source signal sequence. Then, the multi-sensor collaborative delay feature matrix and the multi-source signal sequence are subjected to cross-modal adaptive fusion to generate a collaborative fusion feature tensor. S4. Based on the collaborative fusion feature tensor, input it into the first fault diagnosis branch network to obtain the fault occurrence probability prediction value. When the fault occurrence probability prediction value is greater than the preset first threshold, input the collaborative fusion feature tensor into the second fault diagnosis branch network to obtain the probability distribution vector of the fault type. S5. Based on the estimated probability of failure and the probability distribution vector, output a comprehensive diagnostic report including the diagnostic timestamp, preliminary identification results, fine classification results, and corresponding probability values.

[0006] Preferably, step S2 involves weighted fusion of the first-level and second-level abstract delayed feature vectors, specifically including the following steps: S21. Map the first-level abstract delayed feature vector to the query vector through the first linear transformation; S22. Map the second-level abstract delayed feature vector to a key vector through a second linear transformation, and then to a value vector through a third linear transformation. S23. Determine the scaling dot product similarity between the query vector and the key vector, and obtain the attention weights after normalization by the normalized exponential function; S24. Use attention weights to perform a weighted summation of the value vectors to obtain attention-enhanced features; S25. Perform a residual connection between the attention-enhanced features and the second-level abstract delayed feature vector to obtain the enhanced second-level features; S26. The enhanced second-level feature is concatenated with the first-level abstract delay feature vector and transformed through a fully connected layer to output the comprehensive delay feature vector.

[0007] Preferably, in step S2, a sliding window is used to extract the statistical features of the basic delay parameter sequence within the window to obtain the first-level abstract delay feature vector, specifically including the following steps: S31. Set the sliding window width W and the sliding step size S, where W is a preset multiple of the number of data points corresponding to the nominal sampling period of the sensor; S32. Using the sliding window, traverse the basic delay parameter sequence. For the subsequence {τ1,τ2,...,τ_W} within the k-th window, extract the statistical features: S321. Statistical characteristics include: mean It is used to characterize the central tendency of data.

[0008] Standard deviation It is used to characterize the degree of dispersion of data. sqrt

[0009] Skewness It is used to characterize the asymmetry of data distribution. i W

[0010] Kudo It is used to characterize the steepness of the data distribution. i=1 W

[0011] S322. Determine the time-series correlation characteristics, i.e., the autocorrelation coefficient with a delay of 1. It is used to characterize the degree of linear correlation between adjacent values ​​in a sequence.

[0012]

[0013] S323. Calculate the approximate entropy of the subsequence using the template matching method. It is used to quantify the randomness or unpredictability of sequences, where the embedding dimension m=2, and the similarity tolerance r is a preset scaling factor. The product;

[0014] S33. Calculate the feature group obtained from the current sliding window.

[0015] This serves as the output of the first-level abstract delayed feature vector at time k.

[0016] Preferably, the deep learning network model in step S2 specifically includes the following steps: S41. Receive the basic delay parameter sequence, perform one-dimensional convolution operations in parallel using at least two different width convolution kernels, and generate multi-scale feature maps. S42. Receive the feature map output from the one-dimensional convolutional layer, perform recursive operations along the time dimension through the gated recurrent unit, and output the hidden state sequence. S43. Receive the hidden state sequence, determine the weights of the features at each time step in the hidden state sequence through learnable parameters, and perform a weighted summation of the hidden state sequence based on the weights to generate a context feature vector. S44. Receive the context feature vector, and output the second-level abstract delay feature vector through one or more linear transformations and nonlinear activations.

[0017] Preferably, the cross-modal adaptive fusion in step S3 specifically includes the following steps: A one-dimensional convolutional neural network is used to extract local spatiotemporal features of multi-source signal sequences, and the signal feature vector is obtained after global average pooling. The multi-sensor collaborative delay feature matrix is ​​input into the self-attention-based sensor association encoder, and the updated delay feature vector is output. The signal feature vector and the updated delay feature vector are input into the cross-modal bidirectional attention fusion module, which outputs a collaboratively fused feature tensor.

[0018] Preferably, the self-attention-based sensor-associative encoder performs the following steps: The multi-sensor collaborative delay feature matrix is ​​mapped to a query matrix, a key matrix, and a value matrix through a linear transformation to obtain a self-attention output. The self-attention output is residually connected to the delay feature matrix, and then passed through layer normalization and feedforward neural network in sequence to output the updated delay feature vector.

[0019] Preferably, the cross-modal bidirectional attention fusion module linearly projects the signal feature vector and the updated delay feature vector onto the same dimension before calculating the attention, specifically including the following steps: S71. Using the projected signal feature vector as the query and the projected updated delay feature vector as the key and value, calculate the first attention-weighted feature by scaling dot product attention. S72. Using the projected and updated delayed feature vector as the query and the projected signal feature vector as the key and value, calculate the second attention-weighted feature by scaling dot product attention. S73. The first attention-weighted feature and the second attention-weighted feature are concatenated and then passed through a joint fully connected layer to output a collaborative fusion feature tensor.

[0020] Preferably, the second fault diagnosis branch network includes at least three fully connected hidden layers, each followed by a batch normalization layer, a random drop-out layer, and a linear rectified activation layer, and the output layer uses a normalized exponential function to output a probability distribution vector.

[0021] Preferably, the following steps are also included: S6. Store the original multi-sensor monitoring data of the current diagnostic cycle, the basic delay parameter sequence of each sensor and the corresponding evaluation result label as training samples into the historical sample set. S7. When the number of samples in the historical sample set exceeds the preset update threshold, the incremental learning process is started. Based on the original monitoring data and basic delay parameter sequence stored in the historical sample set, steps S2 to S3 are re-executed to generate a new collaborative fusion feature tensor. The parameters of the deep learning network model, cross-modal adaptive fusion module, first health assessment branch network and second health assessment branch network are updated using the mini-batch gradient descent method. The loss function is the weighted cross-entropy loss function. S8. After the parameter update is complete, replace the original model parameters with the updated model parameters.

[0022] This application addresses the technical bottleneck in existing multi-sensor collaborative fault diagnosis, which generally treats data delay as interference noise and fails to extract the fault information it contains, resulting in the masking of early fault features and severe spatiotemporal mismatch of multiple sensors, which severely restricts the accuracy of diagnosis. Through a multi-collaborative design of active extraction of multi-level features of data delay, cross-modal adaptive fusion, multi-level diagnosis and incremental self-optimization, the accuracy and robustness of collaborative fault diagnosis of multi-source asynchronous signals in non-uniform delay scenarios are improved.

[0023] Specifically, this application extracts the communication transmission delay, internal processing delay, and sampling delay of each sensor as basic parameters, uses sliding window statistics to obtain the first-level abstract delay features, and simultaneously extracts the second-level abstract delay features through a deep learning network. A comprehensive delay feature vector is generated through attention-weighted fusion, and the features of all sensors are concatenated to construct a collaborative delay matrix. Multi-source signals synchronized with time are fused through self-attention encoding and bidirectional attention to generate a collaborative fusion feature tensor. This tensor is then input into the first and second fault diagnosis branch networks to obtain the fault probability and classification results. The model is continuously optimized through incremental learning, thereby improving the accuracy of multi-sensor collaborative fault diagnosis and the ability to identify early faults under non-uniform delay conditions by utilizing data delay information. Attached Figure Description

[0024] Figure 1 This is a block diagram of a multi-sensor collaborative fault diagnosis method utilizing data delay according to an embodiment of this application. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] Please see Figure 1According to one embodiment of this application, a multi-sensor collaborative fault diagnosis method utilizing data delay is provided. This application addresses the technical bottleneck in existing multi-sensor collaborative fault diagnosis methods, which generally treat data delay as interference noise, failing to extract the fault information it contains. This leads to the masking of early fault characteristics and severe spatiotemporal mismatch among multiple sensors, severely limiting diagnostic accuracy. The method extracts the communication transmission delay, internal processing delay, and sampling delay of each sensor as basic parameters. A sliding window statistical method is used to obtain the first-level abstract delay features. Simultaneously, a deep learning network extracts the second-level abstract delay features, and attention-weighted fusion is used to generate a comprehensive delay feature vector. This vector is then used to construct a collaborative delay matrix and perform cross-modal adaptive fusion with multi-source signals. Finally, a multi-level diagnostic network is used to achieve high-precision fault diagnosis under non-uniform delay conditions. Specifically, the method includes the following steps S1-S5.

[0027] In step S1, monitoring data from multiple sensors on the monitored device are acquired. The monitored device is a CNC machine tool spindle system. Multiple sensors are deployed near the bearing housing, housing, and drive motor of the spindle, including an accelerometer (model PCB352C33, sampling frequency 25.6kHz), a temperature sensor (PT100, sampling frequency 1Hz), and a current sensor (Hall effect, sampling frequency 10kHz). A GPS timing module (model TrimbleResolutionT) is used as a unified time reference. Hardware synchronization of each sensor is performed through a second pulse signal. The timestamp error of each sensor's acquisition time is less than 1 microsecond. A timestamp is added to each acquisition record of each sensor to generate a raw monitoring data sequence with timestamps. The timing synchronization adopts the IEEE1588 precise time protocol or simplified network time protocol, or uses the second pulse calibration module inside the bus module to generate a reference trigger signal. A microsecond or millisecond-level timestamp is added to each data record.

[0028] In one embodiment, before performing step S2 to extract the delay parameters, it is necessary to understand their potential correlation with the device state. This invention does not assert a simple linear causal relationship between delay and mechanical failure, but rather is based on the observation that changes in device operating state (such as increased vibration, load fluctuations, or localized overheating) can trigger coordinated, nonlinear patterns of change in the communication, processing, and sampling delay sequences of multiple sensors by affecting the overall system's temporal behavior, bus load, or sensor operating points. These patterns, individually weak and submerged in noise, can be effectively characterized and amplified through the multi-level feature extraction (capturing statistical and deep temporal features) and multi-sensor collaborative analysis (constructing a coordinated delay matrix) proposed in this invention. Furthermore, through a cross-modal attention mechanism and complementary physical signal features, they can be used together for state assessment.

[0029] In step S2, based on the timestamped original monitoring data sequence, the communication transmission delay, internal processing delay, and sampling delay of each sensor are extracted as basic delay parameters. Taking the accelerometer as an example, the communication transmission delay is obtained by sending a data request from the sensor to the central processing unit and recording half of the round-trip time, which is 0.2ms. The communication transmission delay is set to 0.1ms. The internal processing delay is given by the sensor hardware manual and is set to 0.05ms. For periodically acquired sensors, the difference between the actual acquisition interval and the nominal sampling period is used as the sampling delay. The nominal sampling period is 39.0625μs, and the measured sampling delay is between -2μs and +3μs. The above three delay parameters are associated with the timestamp of the acquisition time and stored to form a basic delay parameter sequence. The basic delay parameters of the temperature sensor and the current sensor are obtained in the same way.

[0030] Furthermore, in engineering practice, the basic delay parameter can be estimated based on the timestamp as follows:

[0031] 1. Communication transmission delay: In a request-response system, it can be calculated by the host as (receive timestamp, request sending timestamp) / 2. For streaming data, it can be estimated by the jitter of the time interval between data packets.

[0032] 2. Internal processing delay: This is typically the inherent fixed or small-range fluctuating delay of the sensor, the nominal value of which can be obtained from the datasheet. If no specific value is available, it can be set to zero. The ability of the method of this invention to extract its variation characteristics still mainly depends on other delay components.

[0033] 3. Sampling delay: This is a key parameter characterizing timing jitter. For a nominal sampling period of... The sensor, its first The theoretical sampling time for each data point should be: ,in The theoretical start time of the sequence, sampling delay That is, the actual timestamp carried by the data point. Deviation from theoretical time: This calculation only requires a unified time synchronization and known... This is a feasible solution from an engineering perspective.

[0034] A sliding window is used to extract the statistical features of the basic delay parameter sequence within the window, obtaining the first-level abstract delay feature vector. The sliding window width W and sliding step size S are set, where W is a preset multiple of the number of data points corresponding to the nominal sampling period of the sensor; W = 256 data points (corresponding to a 10ms duration) and S = 64. The sliding window is used to traverse the basic delay parameter sequence. For the subsequence {τ1, τ2, ..., τ_W} within the k-th window, the mean is determined. It is used to characterize the central tendency of data. Determine the standard deviation It is used to characterize the degree of dispersion of data. sqrt Determine skewness It is used to characterize the asymmetry of data distribution. i W

[0035] Determine kurtosis It is used to characterize the steepness of the data distribution. i=1 W

[0036] Determine the time-series correlation characteristics, i.e., the autocorrelation coefficient with a delay of 1. It is used to characterize the degree of linear correlation between adjacent values ​​in a sequence. ] / [ ]

[0037] Parameter settings: Set the embedding dimension Similarity tolerance ,in This represents the standard deviation of the current window subsequence.

[0038] Vector reconstruction: Reconstructing subsequences within a window Recorded as Construct in sequence Dimensional vector:

[0040] Distance calculation: for each vector Calculate its relationship with all vectors distance The distance is defined as the maximum absolute value of the difference between corresponding components of two vectors.

[0041] Similarity statistics: For each Statistics meet the conditions of The quantity, denoted as Then calculate .

[0042] Calculate the mean: Calculate the mean for all corresponding Average value: Add dimension: Increase the embedding dimension to Repeat steps b) to e) to obtain .

[0043] Calculate the approximate entropy: Finally, the approximate entropy of the window subsequence is:

[0044]

[0045] The approximate entropy of the subsequence is calculated using template matching. It is used to quantify the randomness or unpredictability of sequences, where the embedding dimension m=2, and the similarity tolerance r is a preset scaling factor. The product of these factors, the similarity tolerance r = α·σ_k, where α is a preset proportionality coefficient, taken as 0.2;

[0046] The feature set calculated by the current sliding window This serves as the output of the first-level abstract delayed feature vector at time k.

[0047] The basic delay parameter sequence is input into a pre-trained deep learning network model for feature extraction to obtain a second-level abstract delay feature vector. This deep learning network model has the following hierarchical structure and data flow: a one-dimensional convolutional layer receives the basic delay parameter sequence and performs one-dimensional convolution operations in parallel using at least two different width convolutional kernels to generate multi-scale feature maps; a recurrent neural network layer receives the feature maps output by the one-dimensional convolutional layer and performs recursive calculations along the time dimension using gated recurrent units or long short-term memory network units to output a hidden state sequence; a temporal attention layer receives the hidden state sequence, calculates the weights of features at each time step in the hidden state sequence using learnable parameters, and performs a weighted summation of the hidden state sequence based on the weights to generate a context feature vector; a fully connected layer receives the context feature vector and outputs the second-level abstract delay feature vector through one or more linear transformations and nonlinear activations.

[0048] In one embodiment, the one-dimensional convolutional layer uses 64 convolutional kernels with widths of 3, 5, and 7 respectively, the gated recurrent unit layer has 128 hidden units, the temporal attention layer outputs a 128-dimensional context, and the fully connected layer maps the 128 dimensions to a 32-dimensional second-level abstract delayed feature vector.

[0049] The first-level and second-level abstract delay feature vectors are weighted and fused to generate a comprehensive delay feature vector for each sensor. The first-level abstract delay feature vector is mapped to a query vector through a first linear transformation, and the second-level abstract delay feature vector is mapped to a key vector through a second linear transformation and then to a value vector through a third linear transformation. The scaling dot product similarity between the query vector and the key vector is calculated, and attention weights are obtained after normalization using a normalized exponential function. The attention weights are used to perform a weighted summation on the value vectors to obtain attention-enhanced features. The attention-enhanced features are residually connected with the second-level abstract delay feature vector to obtain enhanced second-level features. The enhanced second-level features are concatenated with the first-level abstract delay feature vector and transformed through a fully connected layer to output the comprehensive delay feature vector. The first-level features are 6-dimensional, the second-level features are 32-dimensional, and the comprehensive delay feature vector is obtained after attention fusion. The above process is performed on the acceleration, temperature, and current sensors respectively to obtain their respective comprehensive delay feature vectors.

[0050] In one embodiment of this application, cross-modal adaptive fusion includes: extracting local spatiotemporal features of the multi-source signal sequence using a one-dimensional convolutional neural network, obtaining a signal feature vector after global average pooling, inputting the multi-sensor collaborative delay feature matrix into a self-attention-based sensor association encoder, outputting an updated delay feature vector, and inputting the signal feature vector and the updated delay feature vector into a cross-modal bidirectional attention fusion module to output the collaboratively fused feature tensor. The self-attention-based sensor association encoder performs a linear transformation to map the multi-sensor collaborative delay feature matrix into a query matrix, a key matrix, and a value matrix to obtain a self-attention output. The self-attention output is then residually concatenated with the delay feature matrix, and subsequently processed through layer normalization and a feedforward neural network to output the updated delay feature vector.

[0051] In one embodiment, the input to the self-attention encoder is a 3×32 matrix, the output is a 3×32 matrix, and the 32-dimensional global delay feature vector is obtained by average pooling along the sensor dimension.

[0052] In one embodiment of this application, the cross-modal bidirectional attention fusion module performs the following: using the signal feature vector as a query and the updated delay feature vector as a key and value, it calculates a first attention-weighted feature by scaling dot product attention; using the updated delay feature vector as a query and the signal feature vector as a key and value, it calculates a second attention-weighted feature by scaling dot product attention; it concatenates the first attention-weighted feature and the second attention-weighted feature, and then passes them through a joint fully connected layer to output the collaboratively fused feature tensor.

[0053] In one embodiment, the signal feature vector is 64-dimensional, the updated delay feature vector is 32-dimensional, the first weighted feature is 64-dimensional, the second weighted feature is 32-dimensional, and the concatenation results in 96-dimensional features. A 64-dimensional collaborative fusion feature tensor is then output through a joint fully connected layer.

[0054] In step S4, the collaborative fusion feature tensor is input into the first fault diagnosis branch network to obtain a fault occurrence probability prediction. The first fault diagnosis branch network is a binary classification network, including two fully connected layers (128 neurons and 64 neurons) and an output layer Sigmoid. A first threshold of 0.7 is set. When the fault occurrence probability prediction is greater than the preset first threshold, the second fault diagnosis branch network is activated. The collaborative fusion feature tensor is input into the second fault diagnosis branch network to obtain a probability distribution vector of the fault type. The second fault diagnosis branch network includes at least three fully connected hidden layers. Each fully connected hidden layer is followed by a batch normalization layer, a random dropout layer, and a linear rectified activation layer. The output layer uses a normalized exponential function to output the probability distribution vector.

[0055] In one embodiment, the second branch network has three fully connected hidden layers with 256, 128, and 64 neurons respectively. The dropout rate of the random dropout layer is 0.3. The number of neurons in the output layer is the number of fault categories (preset to 6 categories, including bearing inner ring fault, bearing outer ring fault, gear tooth breakage, rotor imbalance, sensor zero drift, and data packet loss). The fault type corresponding to the maximum value in the probability distribution vector is taken as the fine diagnosis result.

[0056] In step S5, based on the estimated probability of the fault occurrence and the probability distribution vector, a comprehensive diagnostic report is output, which includes a diagnostic timestamp, preliminary identification results, fine classification results, and corresponding probability values.

[0057] In one embodiment, a comparative experiment was conducted on a measured dataset of a CNC machine tool spindle system. The dataset included data from three sensors—accelerometer, temperature sensor, and current sensor—under normal conditions and six types of fault conditions (bearing inner ring fault, outer ring fault, rolling element fault, gear tooth breakage, rotor imbalance, and slight sensor zero drift), and non-uniform communication and sampling delays were artificially introduced to reflect the characteristics of an industrial environment.

[0058] Specifically as follows: 1. Baseline Method A: Only use the time-aligned preprocessed multi-source vibration, temperature, and current signal sequences, and input a diagnostic network with the same structure as the second fault diagnosis branch network of this invention (i.e., ignore all delay information).

[0059] 2. Baseline Method B: Extract the mean and standard deviation of the delay of each sensor as simple features, directly concatenate them with the aligned signal feature vector, and then input them into the same diagnostic network as baseline A.

[0060] 3. Method of the present invention: The complete process described in this application is adopted.

[0061] Experimental results: On the same independent test set, the overall fault diagnosis accuracy of the method of the present invention reached 98.7%, which is significantly higher than the 92.1% of the baseline method A and the 95.3% of the baseline method B.

[0062] In the task of identifying "early faults" (defined as samples whose fault feature amplitude is 20% lower than the normal operating condition signal amplitude), the detection rate of the method of this invention is 89.5%, which is a significant improvement over baseline A (64.4%) and baseline B (76.8%). The experimental results demonstrate that the scheme proposed in this invention, which extracts deep features from delay and performs cross-modal fusion, can effectively utilize delay information and significantly improve diagnostic accuracy and early fault identification capabilities.

[0063] All parts not covered in this application are the same as or can be implemented using existing technology. Although embodiments of this application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-sensor collaborative fault diagnosis method utilizing data latency, characterized in that, Includes the following steps: S1. Acquire monitoring data from multiple sensors on the monitored system, synchronize the multiple sensors using a time synchronization module or network time synchronization protocol, add a timestamp to each acquisition record of each sensor, and generate a raw monitoring data sequence with timestamps. S2. Based on the original monitoring data sequence with timestamps, extract the communication transmission delay, internal processing delay, and sampling delay of each sensor as basic delay parameters. Use a sliding window to extract the statistical features of the basic delay parameter sequence within the window to obtain the first-level abstract delay feature vector. Input the basic delay parameter sequence into a pre-trained deep learning network model for feature extraction to obtain the second-level abstract delay feature vector. Weightedly fuse the first-level and second-level abstract delay feature vectors to generate the comprehensive delay feature vector for each sensor. S3. Based on the comprehensive delay feature vectors of all sensors, a multi-sensor collaborative delay feature matrix is ​​constructed by splicing them together in the order of the sensors. At the same time, data cleaning and time alignment preprocessing are performed on the original monitoring data sequence with timestamps to obtain a time-synchronized multi-source signal sequence. Then, the multi-sensor collaborative delay feature matrix and the multi-source signal sequence are subjected to cross-modal adaptive fusion to generate a collaborative fusion feature tensor. S4. Based on the collaborative fusion feature tensor, input it into the first fault diagnosis branch network to obtain the fault occurrence probability prediction value. When the fault occurrence probability prediction value is greater than the preset first threshold, input the collaborative fusion feature tensor into the second fault diagnosis branch network to obtain the probability distribution vector of the fault type. S5. Based on the estimated probability of failure and the probability distribution vector, output a comprehensive diagnostic report including the diagnostic timestamp, preliminary identification results, fine classification results, and corresponding probability values.

2. The method according to claim 1, characterized in that, Step S2 involves weighted fusion of the first-level and second-level abstract delayed feature vectors, specifically including the following steps: S21. Map the first-level abstract delayed feature vector to the query vector through the first linear transformation; S22. Map the second-level abstract delayed feature vector to a key vector through a second linear transformation, and then to a value vector through a third linear transformation. S23. Determine the scaling dot product similarity between the query vector and the key vector, and obtain the attention weights after normalization by the normalized exponential function; S24. Use attention weights to perform a weighted summation of the value vectors to obtain attention-enhanced features; S25. Perform a residual connection between the attention-enhanced features and the second-level abstract delayed feature vector to obtain the enhanced second-level features; S26. The enhanced second-level feature is concatenated with the first-level abstract delay feature vector and transformed through a fully connected layer to output the comprehensive delay feature vector.

3. The method according to claim 1, characterized in that, In step S2, a sliding window is used to extract the statistical features of the basic delay parameter sequence within the window to obtain the first-level abstract delay feature vector, which specifically includes the following steps: S31. Set the sliding window width W and the sliding step size S, where W is a preset multiple of the number of data points corresponding to the nominal sampling period of the sensor; S32. Using the sliding window, traverse the basic delay parameter sequence. For the subsequence {τ1,τ2,...,τ_W} within the k-th window, extract the statistical features: S321. Statistical characteristics include: mean It is used to characterize the central tendency of data. Standard deviation It is used to characterize the degree of dispersion of data. sqrt Skewness It is used to characterize the asymmetry of data distribution. and IN Kudo It is used to characterize the steepness of the data distribution. i=1 IN S322. Determine the time-series correlation characteristics, i.e., the autocorrelation coefficient with a delay of 1. It is used to characterize the degree of linear correlation between adjacent values ​​in a sequence. S323. Calculate the approximate entropy of the subsequence using the template matching method. It is used to quantify the randomness or unpredictability of sequences, where the embedding dimension m=2, and the similarity tolerance r is a preset scaling factor. The product; S33. Calculate the feature group obtained from the current sliding window. This serves as the output of the first-level abstract delayed feature vector at time k.

4. The method according to claim 1, characterized in that, in, The deep learning network model in step S2 specifically includes the following steps: S41. Receive the basic delay parameter sequence, perform one-dimensional convolution operations in parallel using at least two different width convolution kernels, and generate multi-scale feature maps. S42. Receive the feature map output from the one-dimensional convolutional layer, perform recursive operations along the time dimension through the gated recurrent unit, and output the hidden state sequence. S43. Receive the hidden state sequence, determine the weights of the features at each time step in the hidden state sequence through learnable parameters, and perform a weighted summation of the hidden state sequence based on the weights to generate a context feature vector. S44. Receive the context feature vector, and output the second-level abstract delay feature vector through one or more linear transformations and nonlinear activations.

5. The method according to claim 1, characterized in that, Step S3, cross-modal adaptive fusion, specifically includes the following steps: A one-dimensional convolutional neural network is used to extract local spatiotemporal features of multi-source signal sequences, and the signal feature vector is obtained after global average pooling. The multi-sensor collaborative delay feature matrix is ​​input into the self-attention-based sensor association encoder, and the updated delay feature vector is output. The signal feature vector and the updated delay feature vector are input into the cross-modal bidirectional attention fusion module, which outputs a collaboratively fused feature tensor.

6. The method according to claim 5, characterized in that, The self-attention-based sensor-associative encoder performs the following steps: The multi-sensor collaborative delay feature matrix is ​​mapped to a query matrix, a key matrix, and a value matrix through a linear transformation to obtain a self-attention output. The self-attention output is residually connected to the delay feature matrix, and then passed through layer normalization and feedforward neural network in sequence to output the updated delay feature vector.

7. The method according to claim 5, characterized in that, The cross-modal bidirectional attention fusion module linearly projects the signal feature vector and the updated delay feature vector onto the same dimension before calculating attention. Specifically, this includes the following steps: S71. Using the projected signal feature vector as the query and the projected updated delay feature vector as the key and value, calculate the first attention-weighted feature by scaling dot product attention. S72. Using the projected and updated delayed feature vector as the query and the projected signal feature vector as the key and value, calculate the second attention-weighted feature by scaling dot product attention. S73. The first attention-weighted feature and the second attention-weighted feature are concatenated and then passed through a joint fully connected layer to output a collaborative fusion feature tensor.

8. The method according to claim 1, characterized in that, The second fault diagnosis branch network includes at least three fully connected hidden layers. Each fully connected hidden layer is followed by a batch normalization layer, a random drop-out layer, and a linear rectified activation layer. The output layer uses a normalized exponential function to output a probability distribution vector.

9. The method according to claim 1, characterized in that, It also includes the following steps: S6. Store the original multi-sensor monitoring data of the current diagnostic cycle, the basic delay parameter sequence of each sensor and the corresponding evaluation result label as training samples into the historical sample set. S7. When the number of samples in the historical sample set exceeds the preset update threshold, the incremental learning process is started. Based on the original monitoring data and basic delay parameter sequence stored in the historical sample set, steps S2 to S3 are re-executed to generate a new collaborative fusion feature tensor. The parameters of the deep learning network model, cross-modal adaptive fusion module, first health assessment branch network and second health assessment branch network are updated using the mini-batch gradient descent method. The loss function is the weighted cross-entropy loss function. S8. After the parameter update is complete, replace the original model parameters with the updated model parameters.

Images