A bolt axial force acquisition signal fault type identification method and device

Abstract

The present application relates to a kind of bolt axial force acquisition signal fault type identification method and device, belong to fault diagnosis technical field.The method includes the following steps: synchronous acquisition original axial force time series signal and sample level 16 state word sequence, based on state duration rate to wavelet packet node coefficient Adaptive band shaping is obtained Purified axial force time series signal;Extract the conditional time series pattern feature in abnormal trigger interval and state word interaction feature, construct augmented state feature vector;Build double-branch heterogeneous feature extraction and fault condition mapping model, generate channel gate and soft threshold for axial force time series branch by state condition branch Dynamic, realize state information whole-process guide feature extraction, complete training using reliability weighted loss;After processing, the data to be identified are input into the trained model, and the fault type prediction result is obtained.The present application can accurately distinguish multiple axial force faults and acquisition link abnormalities, improve the accuracy and robustness of fault identification under complex conditions.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology, specifically relating to a fault type identification method and device for bolt axial force acquisition signals. Background Technology

[0002] Bolts are critical connecting components in mechanical assembly and civil structures, and their preload state directly affects the overall load-bearing capacity and safety margin of the structure. To meet the needs of long-term online monitoring, axial force sensors and supporting data acquisition and wireless transmission systems are widely deployed on engineering sites. By continuously collecting bolt axial force sampling values, the mechanical states such as changes in preload, axial force decay, instantaneous slippage, and abnormal oscillations are reflected.

[0003] Existing methods have the following drawbacks: fixed filtering does not distinguish between anomaly types and durations, treating short-term intermittent anomalies and persistent anomalies equally, which easily weakens the true axial force changes that overlap with interference frequency bands; the status word only uses window-level "whether it is triggered", lacking axial force morphology analysis during anomaly triggering and time-series interaction modeling of multiple anomaly events; the status word and axial force features are only concatenated at the classification layer, which cannot adaptively adjust the feature extraction process according to the status conditions, thus limiting the ability to distinguish between multiple types of faults. Summary of the Invention

[0004] To solve the above problems, the present invention provides the following technical solution: This invention provides a method for identifying fault types in bolt axial force acquisition signals, comprising the following steps: S1. Collect bolt axial force monitoring samples, label the fault categories, and obtain labeled samples to form a training sample set; the labeled samples include the original axial force time sequence signal, the sampling level 16-bit status word sequence, the window level 16-bit status word vector, the window level status duration vector, and the real fault category label. S2. The fault-sensitive frequency band and its processing intensity are determined by using the window-level 16-bit state word vector and the window-level state duration vector. Continuous soft attenuation is performed on the wavelet packet node coefficients of the original axial force time series signal to obtain the purified axial force time series signal. The abnormal triggering interval is determined according to the sampling-level 16-bit state word sequence. Conditional time series morphological features and abnormal event interaction features are extracted from the purified axial force time series signal to obtain the augmented state feature vector. S3. Construct a dual-branch heterogeneous feature extraction and fault condition mapping model, including an axial force time-series branch and a state condition branch; input the purified axial force time-series signal, a window-level 16-bit state word vector, a window-level state duration vector, and an augmented state feature vector into the model to obtain a fault prediction probability vector; train and optimize the model using a loss function to obtain a trained dual-branch heterogeneous feature extraction and fault condition mapping model. S4. Real-time acquisition of the original axial force timing signal to be identified and the corresponding 16-bit status word sequence of the sampling level to be identified. The 16-bit status word sequence of the sampling level to be identified is summarized by bit and the duration ratio is statistically analyzed to obtain the 16-bit status word vector of the window level to be identified and the state duration rate vector of the window level to be identified. After processing in step S2, the input is into the trained dual-branch heterogeneous feature extraction and fault condition mapping model to obtain the fault type prediction result.

[0005] Step S1 specifically includes: An axial force sensor is placed on the bolt to be monitored, and axial force sampling values ​​are continuously collected according to a preset sampling rate. The acquisition window is divided into segments with a fixed time length to obtain the original axial force time sequence signal. ; While acquiring each axial force sample value, the 16-bit status word corresponding to the current sampling time is read to obtain a sampling-level 16-bit status word sequence that is time-aligned with the original axial force timing signal. ; Window-level summarization is performed on each state word bit in the sampled 16-bit state word sequence to obtain a window-level 16-bit state word vector. and window-level state duration vector .

[0006] In step S2, obtaining the purification axial force timing signal includes: For the The original axial force time-series signal of each sample implement Layer wavelet packet decomposition yields the th The end wavelet packet node coefficients of the layer; the first Layer The node coefficients of each terminal wavelet packet are defined as follows: According to the axial force sampling rate Wavelet packet decomposition level Determine the center frequency corresponding to each terminal wavelet packet node. ; From the 16 status word bits, select the status word bits that are associated with frequency band interference, and establish the status word bits and fault-sensitive center frequency. And bandwidth The correspondence between them; based on the distance between the center frequency of the terminal wavelet packet node and the center frequency of each fault-sensitive node, calculate the frequency band correlation coefficient of each terminal wavelet packet node for different abnormal events. ; By combining the anomaly duration ratio, frequency band correlation coefficient, and shaping intensity parameter in the window-level state persistence vector, the retention coefficient of each terminal wavelet packet node is calculated. ; Perform element-wise scaling on the corresponding terminal wavelet packet node coefficients using the retained coefficients to obtain the shaped wavelet packet node coefficients. ; Perform inverse wavelet packet transform on all integer wavelet packet node coefficients according to the original wavelet packet tree structure, and trim or pad the reconstruction result to the required length. The timing signal of the axial force for purification is obtained.

[0007] In step S2, the conditional time series morphological features obtained include: The set of state bits involved in the construction of conditional temporal morphological features is defined as... For the state word constructed based on the temporal morphological characteristics of the participating conditions, the intervals with consecutive values ​​of 1 are searched according to the sampling time sequence to obtain the set of abnormal trigger intervals. Extract the abnormal trigger interval from the purification axial force timing signal. The corresponding continuous sampled values ​​are used to obtain conditional time series segments. ; The state triggering duration ratio is calculated based on the length of each conditional time segment, and the local kurtosis coefficient of each conditional time segment is calculated separately. The state word is obtained by weighting and summing the data according to the length of the conditional time sequence segments. Conditional kurtosis characteristics corresponding to position ; For each conditional time series segment, first subtract the mean of the corresponding segment, then calculate the proportion of adjacent centered sampled values ​​whose signs change, thus obtaining the centered cross-zero density. The conditional zero-crossing features are obtained by weighting the centered cross-zero density according to the length of each conditional time series segment. ; Based on the fluctuation energy of the purification axial force time-series signal relative to the window mean, the dynamic energy proportion during abnormal triggering is calculated; the absolute difference between adjacent purification axial force values ​​is calculated within each conditional time-series segment, and the maximum value among all conditional time-series segments is taken to obtain the maximum absolute difference feature. ; The state trigger duration ratio, conditional kurtosis characteristics, conditional zero-crossing characteristics, dynamic energy ratio, and maximum absolute difference characteristics are arranged in a fixed order to form the state word. The 5-dimensional conditional temporal morphological features corresponding to the position ; set The conditional temporal morphological features corresponding to all state bits are concatenated in ascending order of state bit number to obtain the conditional temporal morphological feature subvector. .

[0008] In step S2, the augmented state feature vector is obtained as follows: Based on the fault mechanism and the connection relationship of the acquisition system, the state word combination that needs to be calculated for the second-order interaction feature is pre-set; the second-order interaction feature is calculated based on the window-level 16-bit state word vector. Co-occurrence flag of state word combinations Calculate the first 16-bit status word sequence based on the sampling level. Time overlap ratio of each status word combination The co-occurrence flags and temporal overlap ratios corresponding to all state word combinations are concatenated in a fixed order to obtain the state word interaction feature sub-vector. The conditional temporal morphological feature vector and the state word interaction feature vector are concatenated to obtain the augmented state feature vector. .

[0009] In step S3, the axial force temporal branch extracts axial force temporal features through multi-layer one-dimensional causal dilated convolutional residual blocks, and the state condition branch generates a fault condition vector, including: The window-level 16-bit state word vector, the window-level state duration vector, and the augmented state feature vector are concatenated in a fixed order to obtain the state conditional input vector. The state condition input vector is passed through two fully connected layers in sequence, and the fault condition vector is obtained through linear mapping and nonlinear activation. ; The purification shaft force time series signal is standardized using the shaft force mean and standard deviation from the training set. Using the standardized purification axial force timing signal as the initial input to the axial force timing branch, the timing feature map of layer 0 is obtained. ; build The process involves cascading one-dimensional causal dilated convolutional residual blocks, with the modulation temporal feature map from the previous layer inputting into the next layer's one-dimensional causal dilated convolutional residual block, ultimately yielding the first... Layer modulation timing feature map .

[0010] Furthermore, to obtain the first Layer modulation timing feature map ,include: Based on the fault condition vector Generate the first Layer channel gating vector According to the fault condition vector Generate the first Channel soft threshold ratio for each channel in the layer According to the The average absolute response of each channel in the layer time series feature map is used to calculate the channel amplitude reference value. Multiply the channel soft threshold ratio by the corresponding channel amplitude reference value to obtain the actual channel soft threshold. ; Use actual channel soft threshold for the first The layer temporal feature map undergoes channel-by-channel soft thresholding, using channel gate values ​​to adjust the pass-through ratio of retained features, thus obtaining the first... Layer modulation timing feature map ; will the first Layer modulation timing feature map As the first The input to the one-dimensional causal dilated convolution residual block is used to repeat the above operation to obtain the first... Layer modulation timing feature map .

[0011] In step S3, the fault prediction probability vector is obtained as follows: For the Global average pooling and global max pooling are performed on the layer-modulated temporal feature map respectively. The results of global average pooling and global max pooling are concatenated to obtain the temporal encoded feature vector. ; Fault condition vector Input the classification auxiliary mapping layer to obtain the state auxiliary feature vector. ; The time-encoded feature vector and the state-aided feature vector are concatenated to obtain the fused identification feature vector. ; The fused feature vectors are input into the fully connected classification layer to obtain the category score vectors. The category score vector is used to obtain the fault prediction probability vector through the Softmax function. .

[0012] In step S3, the loss function includes: The proportion of state duration directly related to data reliability is read from the window-level state duration vector. This reading is done directly from the window-level state duration vector. The system reads the duration of sensor disconnection and the duration of data verification anomalies; based on the duration of these two types of reliability anomalies and their corresponding weighting parameters, it calculates the... Sample reliability weights for each sample; Convert the real fault category labels of the training samples into one-hot label vectors, count the number of samples contained in each fault category in the training set, and set the class balance weights according to the number of samples in each category. The overall training loss is calculated based on the sample reliability weight, the class balance weight corresponding to the true fault category, and the fault prediction probability for each sample.

[0013] This invention also provides a fault type identification device for bolt axial force acquisition signals, which executes the above-described fault type identification method for bolt axial force acquisition signals, including: Data acquisition module: used to collect bolt axial force monitoring samples, label fault categories, obtain labeled samples, and form a training sample set; The data processing module is used to determine the fault-sensitive frequency band and its processing intensity using the window-level 16-bit state word vector and the window-level state duration vector; to perform continuous soft attenuation on the wavelet packet node coefficients of the original axial force time series signal to obtain the purified axial force time series signal; to determine the abnormal triggering interval based on the sampling-level 16-bit state word sequence; and to extract the conditional time series morphological features and abnormal event interaction features from the purified axial force time series signal to obtain the augmented state feature vector. Model building module: used to build a dual-branch heterogeneous feature extraction and fault condition mapping model, including axial force time-series branch and state condition branch; The fault prediction module is used to collect the original axial force timing signal to be identified and the corresponding 16-bit status word sequence of the sampling level to be identified in real time. It performs bit-by-bit aggregation and duration ratio statistics on the 16-bit status word sequence of the sampling level to be identified to obtain the 16-bit status word vector of the window level to be identified and the state duration rate vector of the window level to be identified. After processing in step S2, it is input into the trained dual-branch heterogeneous feature extraction and fault condition mapping model to obtain the fault type prediction result.

[0014] The advantages of this invention are: This invention determines the fault-sensitive frequency band based on the anomaly type indicated by the status word and dynamically adjusts the attenuation intensity of the wavelet packet coefficients according to the anomaly duration ratio, replacing fixed filtering and avoiding erroneous weakening of the true axial force changes. It extracts the precise trigger interval of the abnormal event from the sampling-level status word sequence, calculates the axial force morphological features and the temporal interaction features between abnormal events within the interval, and breaks through the coarse-grained limitation of using only window-level flags. The state condition branch dynamically generates channel gating and soft thresholds for each layer of the axial force temporal branch, so that the status word information continuously guides the entire feature extraction process, replacing the shallow fusion method that only splices at the classification layer. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0016] Figure 1 This is a flowchart of the steps of the method of the present invention; Figure 2 A heatmap of the sampling-level 16-bit status word sequence; Figure 3 Heatmap of the wavelet packet node coefficients at the fourth layer end; Figure 4 A schematic diagram illustrating the construction of augmented state eigenvectors; Figure 5 This is a diagram of the architecture of a dual-branch heterogeneous feature extraction and fault condition mapping model. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1 In this embodiment, as Figure 1 As shown, this invention provides a method for fault type identification of bolt axial force acquisition signals, the specific steps of which include: S1. Bolt axial force monitoring sample collection and fault category labeling This invention first synchronously acquires bolt axial force sampling values ​​and a 16-bit status word output by the acquisition device. The bolt axial force sampling values ​​reflect changes in bolt preload, axial force attenuation, instantaneous slippage, and abnormal oscillations. The 16-bit status word records sensor disconnection, abnormal aviation connector contact, wireless link malfunction, power supply failure, data verification failure, and other acquisition statuses. The specific steps are as follows: 1) Arrange axial force sensors on the bolts to be monitored, continuously collect axial force samples according to a preset sampling rate, and divide the acquisition window into fixed time periods to obtain the original axial force time-series signal. Specifically, for the first... Each acquisition window defines the original axial force timing signal as... .in, It refers to the first The original axial force timing signal corresponding to each sample; Indicates the first The sample at the th The axial force sampling value at each sampling time; Indicates the sampling time index. ; This indicates the number of sampling points contained in a single acquisition window. The axial force sampling rate is defined as... The unit is Hz. Used to determine the number of axial force sampling points collected per unit time. In one implementation, The Hz frequency can be set to 500Hz, 1000Hz, or 2000Hz, and the duration of a single acquisition window can range from 1 second to 5 seconds. For example, when... When the Hz frequency is 2s and the acquisition window duration is 2s, .

[0019] In one implementation, samples can be formed by overlapping sliding between acquisition windows, with the window sliding step size ranging from 25% to 50% of the acquisition window length. Using overlapping sliding can increase the number of training samples and reduce information truncation caused when a fault event happens to cross an adjacent acquisition window.

[0020] 2) While acquiring each axial force sample value, read the 16-bit status word corresponding to the current sampling time to obtain a 16-bit status word sequence at the sampling level that is time-aligned with the original axial force timing signal. Specifically, the 16-bit status word sequence at the sampling level is... The 16-bit status word sequence for each sample level is defined as follows: .in, It refers to the first Each sample corresponds to a 16-bit state word sequence at the sampling level, with a dimension of [dimensionality missing]. (16 corresponds to the number of bits in the status word) (Number of sampling points within the corresponding acquisition window) Indicates the first The sample at the th The state word at each sampling time Bit value; Indicates the position index of the status word. ; .when When, it indicates the first The abnormal event corresponding to the bit is in the 1st Each sampling moment is in the triggered state; when When, it indicates the first The abnormal event corresponding to the bit is in the 1st No sampling time was triggered.

[0021] It should be noted that the 16-bit status word sequence at the sampling level uses the same sampling time index as the original shaft force timing signal. When the original output frequency of the status word is lower than the shaft force sampling rate, the value of each status word can be copied to the corresponding shaft force sampling interval according to the status word hold time, so that the 16-bit status word sequence at the sampling level has the same time length as the original shaft force timing signal. .

[0022] 3) Perform window-level aggregation on each state word bit in the sampled 16-bit state word sequence to obtain the window-level 16-bit state word vector and the window-level state duration vector.

[0023] In the specific implementation, a window-level 16-bit state word vector is defined. .in, It refers to the first Each sample corresponds to a window-level 16-bit state word vector; Indicates the first The state word in each sample Whether the bit has been triggered at least once within the current acquisition window, specifically according to... Sure.

[0024] For example, suppose that within a capture window, the third bit of the status word is only... and If the value is 1 at two sampling times and 0 at all other times, then according to the formula... The calculation result is 1, indicating that the abnormal event was triggered at least once within this window.

[0025] At the same time, define a window-level state duration vector. .in, It refers to the first The window-level state duration vector corresponding to each sample; Indicates the status word The proportion of time a position is in the triggered state within the current acquisition window, specifically according to... Sure.

[0026] For example, assuming the acquisition window length If the 0th bit of the status word (sensor disconnection) takes a value of 1 for 200 sampling times, then according to the formula... , This indicates that the sensor was offline for 20% of the time in this window.

[0027] It should be noted that the window-level 16-bit state word vector is used to characterize whether the corresponding abnormal event has occurred, while the window-level state persistence vector is used to characterize the persistence of the corresponding abnormal event within the current acquisition window. Using two types of state information can distinguish between short-term, occasional abnormalities and long-term, persistent abnormalities, avoiding treating anomalies that only trigger at one sampling moment with anomalies that cover the entire acquisition window as having the same intensity.

[0028] 4) Based on the results of manual inspection, on-site test records, axial force variation patterns, and data acquisition system logs, set a real fault category label for each sample.

[0029] The total number of fault categories is defined as In one implementation, The eight fault categories are: normal, slow axial force decay, sudden axial force drop, abnormal axial force oscillation, sensor disconnection, poor connection, wireless link anomaly, and combined anomaly. Specifically, the normal category corresponds to samples where axial force changes are within the allowable range and no obvious abnormalities are observed; the slow axial force decay category corresponds to samples where axial force continuously decreases across multiple acquisition windows; the sudden axial force drop category corresponds to samples where bolt slippage, loose connections, or local structural changes cause a significant short-term decrease in axial force; the abnormal axial force oscillation category corresponds to samples where axial force sampling values ​​exhibit continuous periodic fluctuations or high-frequency jitter; the sensor disconnection category corresponds to samples where the sensor is interrupted, sampling values ​​are missing, or sampling values ​​remain unchanged for a long time; the poor connection category corresponds to samples where unstable contact of aviation plugs, cables, or terminals causes intermittent fluctuations; the wireless link anomaly category corresponds to samples where degraded wireless link quality, data retransmission, or communication interference causes discontinuous data; and the combined anomaly category corresponds to samples where two or more types of anomalies occur simultaneously within the same acquisition window.

[0030] It should be noted that the window-level 16-bit status word vector serves only as fault condition and data reliability information and is not directly equivalent to the actual fault category label. For example, an abnormal power supply status word may cause a slow drift in shaft force or it may only cause short-term data fluctuations. Therefore, it is still necessary to combine it with the original shaft force timing signal to complete the fault category determination.

[0031] 5) Combine the original axial force timing signal, the sampled 16-bit status word sequence, the windowed 16-bit status word vector, the windowed state duration vector, and the real fault category labels into labeled samples to form a training sample set. ,Right now .in, Indicates the total number of training samples; Indicates the first The actual fault category label corresponding to each sample .

[0032] Furthermore, the training sample set is divided into a training set, a validation set, and a test set for model training, hyperparameter selection, and performance evaluation.

[0033] In one embodiment, such as Figure 2 As shown, the sampling-level 16-bit status word sequence is analyzed, and the 16-bit status word value at each sampling moment within the entire acquisition window is represented in heatmap form. Each row corresponds to one status word bit, and each column corresponds to one sampling moment. The trigger distribution of different abnormal events (such as sensor disconnection, power failure, wireless link interference, etc.) on the time axis is shown, reflecting the duration and location of the fault event within the window. The horizontal axis represents time in seconds; the vertical axis represents the status word bit index, which has no unit, and each bit takes a value of 0 or 1.

[0034] S2. Adaptive signal cleanup and augmented state feature construction guided by prior knowledge of state word fault bands. The original axial force timing signal is susceptible to sensor disconnection, poor connection, wireless signal interference, power supply abnormalities, and data verification anomalies. Different abnormal events have different frequency ranges. For example, power supply abnormalities can easily cause low-frequency drift, wireless link abnormalities can easily form periodic interference in a specific frequency range, and poor connection and momentary disconnection can easily cause high-frequency jumps. This invention uses a window-level 16-bit state word vector and a window-level state duration vector to determine the fault-sensitive frequency band and its processing intensity. Continuous soft attenuation is performed on the wavelet packet node coefficients of the original axial force timing signal to obtain a purified axial force timing signal. Furthermore, the anomaly triggering interval is determined based on the sampling-level 16-bit state word sequence. Conditional timing morphological features and abnormal event interaction features are extracted from the purified axial force timing signal to obtain an augmented state feature vector. The specific steps are as follows: S201, Wavelet packet band selective shaping under state duration constraint Fixed-band filtering applies the same frequency suppression range to all original axial force timing signals, which can easily weaken the true axial force attenuation and axial force drop characteristics.

[0035] This invention attenuates the wavelet packet node coefficients close to the fault-sensitive frequency band of the abnormal event only when the corresponding status word is triggered, and adjusts the attenuation intensity according to the duration of the abnormal event. The specific steps are as follows: 1) Regarding the first The original axial force time-series signal of each sample implement Layer wavelet packet decomposition yields the th The end wavelet packet node coefficients of the layer, and the first Layer contains One terminal wavelet packet node.

[0036] In practical implementation, a wavelet basis suitable for the axial force variation characteristics is first selected, such as db4, sym4, or coif3. Then, the original axial force time-series signal is processed layer by layer according to low-pass decomposition and high-pass decomposition. Selecting a wavelet basis suitable for the axial force variation characteristics mainly refers to the high similarity between the waveform of the wavelet basis and the characteristics of the axial force signal to be analyzed (such as sudden drops and oscillations). This is usually done through comparative experiments. For example, the db4 (Daubechies 4th order) wavelet is good at capturing abrupt changes and spikes in the signal; the sym4 (Symlet 4th order) wavelet has good symmetry, which can reduce phase distortion during signal reconstruction to some extent; the coif3 (Coiflet 3rd order) wavelet has high vanishing moments in both its scaling function and wavelet function, making it suitable for analyzing relatively smooth variation trends. When selecting a basis, the reconstruction error or classification performance under different wavelet bases can be tested on simulation or historical data of various fault types, and the optimal one can be chosen.

[0037] Furthermore, the first Layer The node coefficients of each terminal wavelet packet are defined as follows: ,in, . This refers to the original axial force timing signal at the 1st... Layer Wavelet packet coefficient sequence within each frequency band.

[0038] In practice, the wavelet packet node order obtained by different software may be arranged in a tree structure rather than in ascending frequency order. Therefore, before calculating the node frequency, the theoretical frequency band range corresponding to each node should be used to determine the order of the wavelet packet nodes. The wavelet packet nodes at the end of the layer are reordered according to their center frequencies from low to high; subsequent node indexes... Both refer to the node indices sorted in ascending order of frequency. For example, when... At that time, The tree structure might initially consist of nodes 0, 1, 3, 2, 6, 7, 5, and 4, before being reordered according to theoretical frequency bands from low to high. After determining the theoretical frequency band range for each node based on its corresponding low-pass / high-pass filter combination, it is reordered as nodes 0, 1, 2, 3, 4, 5, 6, and 7, with the corresponding center frequencies increasing sequentially. Subsequent node indices... Both refer to the index after sorting by this frequency in ascending order.

[0039] Wavelet packet decomposition level Used to control the accuracy of frequency band division. In one implementation, The original axial force timing signal was divided into 16 end frequency bands; when it is necessary to further distinguish adjacent interference frequency bands, You can choose 5.

[0040] 2) Based on axial force sampling rate Wavelet packet decomposition level The center frequency corresponding to each terminal wavelet packet node is determined. Specifically, the... The bandwidth of each terminal wavelet packet node in the layer is defined as ,Right now .in, It refers to the first The bandwidth covered by each terminal wavelet packet node of the layer; This indicates the Nyquist frequency.

[0041] Furthermore, after arranging them in ascending order of frequency, the first... Layer The center frequency of the terminal wavelet packet nodes .in, It refers to the first Layer The center frequencies of the terminal wavelet packet nodes are used to calculate the proximity between the current terminal wavelet packet node and the fault-sensitive frequency band. For example, when Hz and hour, The center frequency of the 0th terminal wavelet packet node is 15.625 Hz, and the center frequency of the 1st terminal wavelet packet node is 46.875 Hz.

[0042] 3) Select status words that are clearly associated with frequency band interference from the 16 status word bits, and establish the correspondence between status word bits, fault-sensitive center frequency and frequency band influence width.

[0043] Specifically, the number of abnormal events participating in frequency band selective shaping is defined as... , will the The actual state bit number corresponding to the exception event is defined as follows: ,in, For the first Setting the frequency of fault sensitivity centers for exception events And bandwidth Among them, the frequency of the fault-sensitive center. Used to determine the frequency location primarily affected by the abnormal event; bandwidth of influence. This is used to control the extent to which the anomalous event attenuates the terminal wavelet packet nodes near the fault-sensitive center frequency.

[0044] In one implementation, the first, second, and fifth bits of the status word are selected to participate in frequency band selective shaping, i.e. , , The first bit of the status word corresponds to a power supply abnormality and the fault-sensitive center frequency. It can be set to 0Hz; the frequency band affects the bandwidth. The frequency can be from 20Hz to 40Hz; the second bit of the status word corresponds to wireless link interference and the fault-sensitive center frequency. The interference frequency can be set based on the pre-experiment results, for example, 50Hz, and the bandwidth can be adjusted accordingly. The frequency can range from 15Hz to 30Hz; the 5th bit of the status word corresponds to a high-frequency oscillation abnormality, and the fault-sensitive center frequency. 200Hz can be used; the frequency band affects the bandwidth. You can choose frequencies from 20Hz to 50Hz.

[0045] In one implementation, the fault-sensitive center frequency and bandwidth can be determined through no-load acquisition tests or fault injection tests. Specifically, axial force signals are collected when a certain abnormal event has not occurred and when it has occurred. The average spectra of the two types of signals are compared, and the frequency position with the larger increase in spectral energy is taken as the fault-sensitive center frequency, and the frequency range with an increase continuously exceeding a preset proportion is taken as the bandwidth. For example, under no-load (bolts are not stressed) conditions, the output signals of the axial force sensor are collected under normal power module conditions and under noise conditions. Fourier transforms are performed on the two signals to obtain the spectrum. If, under the noise condition, the average increase in spectral energy in the range of 0Hz to 30Hz is 5 times that of the normal state, while the increase above 30Hz is not significant, then the fault-sensitive center frequency of the power supply abnormality can be determined. Setting it to 0Hz affects bandwidth. Set to 30Hz.

[0046] It should be noted that the correspondence between status word position, fault-sensitive center frequency, and bandwidth is a pre-defined rule based on physical mechanism analysis and preliminary experiments. Physically, power supply anomalies typically cause a slow drift in the signal baseline, thus corresponding to low frequencies (around 0Hz); wireless link interference is often related to power frequency or its harmonics, thus corresponding to a specific interference frequency; poor connection can cause instantaneous signal jumps, manifesting as high-frequency components. These correspondences and their parameters can be further precisely determined through no-load or fault injection tests.

[0047] 4) Based on the distance between the center frequency of the terminal wavelet packet node and the center frequencies of each fault-sensitive node, calculate the frequency band correlation coefficient of each terminal wavelet packet node for different abnormal events. Specifically, the... Class of exceptions and the first Layer The frequency band correlation coefficient between terminal wavelet packet nodes is defined as follows: It shall be determined in the following manner: , in, The value range is from 0 to 1; when near hour, A value close to 1 indicates that the current terminal wavelet packet node is close to the first... The fault-sensitive frequency bands of similar abnormal events are highly correlated; when keep away hour, The value gradually approaches 0, indicating that the current anomalous event does not require significant suppression of the terminal wavelet packet node.

[0048] It should be noted that the frequency band correlation coefficient adopts a continuous variation method, rather than simply dividing the terminal wavelet packet nodes into suppressed nodes and non-suppressed nodes, which can reduce the reconstruction ringing caused by abrupt changes in the processing intensity of adjacent frequency bands.

[0049] 5) Combining the anomaly persistence ratio, frequency band correlation coefficient, and shaping intensity parameter in the window-level state persistence vector, calculate the retention coefficient for each terminal wavelet packet node. Specifically, the... In the nth sample Layer The retention coefficient of each terminal wavelet packet node is defined as follows: The calculation method is expressed as follows: , in, Used to control the Layer The retention ratio of the node coefficients of each terminal wavelet packet; Indicates the minimum frequency band retention ratio. This is used to prevent coefficients in fault-sensitive frequency bands from being completely deleted; Indicates the first The integer strength parameter corresponding to the abnormal event. ; Indicates the first The status word corresponding to the exception event is in the 1st position. The proportion of persistence in a sample; Indicates the first Class of exceptions and the first Layer The frequency band correlation coefficients between the terminal wavelet packet nodes. In one implementation, The value can be between 0.1 and 0.3, representing the shaping strength parameter corresponding to power supply anomalies, wireless link interference, and high-frequency oscillation anomalies. The values ​​can be 0.8, 1.2, and 1.0 respectively.

[0050] In practical implementation, when none of the status word bits involved in frequency band selective shaping are triggered, the corresponding All are 0, at this time All terminal wavelet packet node coefficients remain unchanged.

[0051] It should be noted that when an abnormal event is triggered for only a short period of time, the corresponding... The frequency band is relatively small, and the fault-sensitive frequency band is attenuated only slightly; when a certain abnormal event continuously covers most of the acquisition window, the corresponding frequency band is attenuated only slightly. The frequency band is significantly attenuated due to the larger frequency, thus preventing over-filtering caused by occasional status word settings.

[0052] In one embodiment, as an example, when Hz, , Hz, At Hz, the center frequency of the 0th terminal wavelet packet node is 15.625Hz, corresponding to a frequency band correlation coefficient of approximately 0.737. If the power supply anomaly persists for a certain period... Shaping strength parameters and Ignoring other abnormal events, the retention coefficient of the current terminal wavelet packet node is approximately 0.714. If the power failure is not triggered, the retention coefficient of the current terminal wavelet packet node is 1.

[0053] 6) Perform element-wise scaling on the corresponding terminal wavelet packet node coefficients using the retained coefficients to obtain the shaped wavelet packet node coefficients, i.e. .in, Indicates the first In the nth sample Layer Each integer wavelet packet node coefficient; Represents the terminal wavelet packet node coefficients before shaping; Broadcast along all coefficient positions of the current terminal wavelet packet node.

[0054] It should be noted that by multiplying all coefficients within each wavelet packet node by the same retention coefficient, the phase and temporal relationships of the wavelet packet node coefficients are preserved, changing only the amplitude of the corresponding frequency band. This is achieved by setting a minimum frequency band retention ratio. The true axial force variation components in the fault-sensitive frequency band can still be partially preserved, especially reducing the risk that the slow decay characteristics of the true axial force will be completely deleted during the low-frequency power supply drift suppression process.

[0055] 7) Perform inverse wavelet packet transform on all integer wavelet packet node coefficients according to the original wavelet packet tree structure, and trim or pad the reconstruction result to the required length. The purification axial force timing signal is obtained, i.e. .in, Indicates the first The purification axial force time series signal corresponding to each sample refers to the axial force signal obtained after wavelet packet frequency band selective shaping and inverse wavelet packet reconstruction under state duration constraint, and its sampling time is in one-to-one correspondence with the original axial force time series signal. Indicates the first The sample at the th The axial force value of purification at each sampling time.

[0056] It should be noted that this step uses the abnormal event type, the duration of the abnormal event, and the fault band distance to determine the band retention coefficient. The abnormal event type determines the frequency location that needs to be processed, the duration of the abnormal event determines the processing intensity, and the fault band distance ensures a smooth transition between adjacent bands. Based on this, while suppressing status word indication interference, it is possible to reduce the false attenuation of the real axial force change by the fixed filter.

[0057] In one embodiment, such as Figure 3 As shown, the coefficients of the fourth-layer wavelet packet nodes are analyzed. A heatmap is used to represent the coefficient sequence of the sixteen terminal nodes obtained after the original axial force time-series signal undergoes four-layer wavelet packet decomposition. The nodes are arranged from low to high center frequency. The distribution of signal energy within different frequency bands, as well as the differences in coefficient amplitude between low-frequency trend components and high-frequency noise components, can be observed. The horizontal axis represents the coefficient sequence number (unitless); the vertical axis represents the node index (in ascending frequency order) (unitless).

[0058] S202. Construction of the timing morphological features corresponding to the status word trigger interval. A window-level 16-bit status word vector can only indicate whether an abnormal event has occurred, but cannot reflect the axial force fluctuation pattern during the occurrence of the abnormal event. This invention determines the actual trigger interval of each abnormal event based on the sampling-level 16-bit status word sequence, and extracts the duration ratio, kurtosis, centered cross-zero density, dynamic energy proportion, and maximum absolute difference value within the corresponding purification axial force timing signal segment, such as... Figure 4 As shown, the specific steps are as follows: 1) For the state bits constructed based on the temporal morphological features of the participating conditions, search for intervals with consecutive values ​​of 1 according to the sampling time sequence to obtain the set of abnormal trigger intervals. Specifically, the set of state bits constructed based on the temporal morphological features of the participating conditions is defined as follows: The number of state bits contained in the set is defined as .

[0059] For the state word Bit, ,Will Each time interval in which a continuous value of 1 is taken is defined as an exception trigger interval, resulting in... .in, Indicates the first The state word in each sample The set of abnormal trigger intervals corresponding to each bit; Indicates the first One abnormal trigger interval; Indicates the number of intervals where an anomaly is triggered.

[0060] Furthermore, the abnormal triggering interval is extracted from the purification axial force timing signal. The corresponding continuous sampled values ​​are used to obtain conditional time series segments. Meanwhile, the number of sampling points contained in a conditional time series segment is defined as .

[0061] In practical implementation, when there are multiple abnormal triggering intervals for the same state word, each conditional time sequence segment calculates the adjacent difference and cross-zero features separately, instead of directly connecting the beginning and end of multiple conditional time sequence segments. This avoids the formation of non-existent artificial jumps between the end sampled value of the previous abnormal triggering interval and the beginning sampled value of the next abnormal triggering interval.

[0062] 2) Calculate the state trigger duration ratio based on the length of each conditional time segment, specifically by directly using the values ​​obtained in S1. As the first The sustained proportional characteristics corresponding to the position. For example, for If the window's status word is triggered at 200 sampling points, then... This indicates that the duration of the abnormal event within the current window is 0.2%.

[0063] Among them, the persistence ratio feature is used to characterize the temporal coverage of the corresponding abnormal event within the current acquisition window. When When the value is close to 0, it indicates that the abnormal event only occurs briefly; when... A value close to 1 indicates that the abnormal event persists for most of the acquisition window.

[0064] 3) Calculate the local kurtosis coefficients for each conditional time segment, and then weight and sum them according to the length of the conditional time segment to obtain the state word. The conditional kurtosis characteristics corresponding to the position.

[0065] Specifically, conditional time series segments The mean is defined as Standard deviation is defined as .when and At that time, the first Local kurtosis coefficients of a conditional time segment Determined in the following manner: , in, Representing conditional timing segments The first in One purification axial force value; Used to characterize the degree of spikes and heavy-tailed fluctuations in the current conditional time series segment.

[0066] In specific implementation, when or At that time, Set to 0.

[0067] Furthermore, the status word Conditional kurtosis characteristics corresponding to position It is obtained by weighting the lengths of each conditional time segment and is expressed as: , The larger the conditional kurtosis characteristic, the more obvious the spikes, pulses, or instantaneous jumps during the triggering of the abnormal event.

[0068] 4) For each conditional time series segment, first subtract the mean of the corresponding segment, and then calculate the proportion of adjacent centered sampled values ​​whose signs change, to obtain the centered cross-zero density.

[0069] Specifically, the first The centered zero-crossing density of a conditional time series segment is defined as follows: The calculation method is expressed as follows: , in, This represents a conditional indicator function, which takes the value 1 when the condition is true and 0 when the condition is false.

[0070] It should be noted that since bolt axial force sampling values ​​are usually positive, directly statistically analyzing whether the original sampling values ​​cross 0 cannot effectively reflect the degree of oscillation. This invention first subtracts the mean of the conditional time series segment, and then statistically analyzes the sign change of the centered sampling values, so that the centered zero-point density can reflect the frequency of rapid back-and-forth changes of axial force on both sides of the local average level.

[0071] In specific implementation, when At that time, Set to 0. Status word number The conditional zero-crossing feature corresponding to the position The density across the center is obtained by weighting the time segment lengths of each condition. For example, suppose the 5th bit of the status word triggers two intervals within a window, and the segment length is... density across zero point ; Length of fragment 2 density across zero point The global weighted average is then... .

[0072] 5) Calculate the dynamic energy ratio during abnormal triggering based on the fluctuation energy of the purification shaft force timing signal relative to the window mean.

[0073] Specifically, the window mean of the purification axial force timing signal is defined as... ,Right now Then the status word The corresponding dynamic energy percentage Determined in the following manner: , in, Used to characterize the state word The proportion of axial force fluctuation energy during the bit triggering period to the total axial force fluctuation energy in the entire acquisition window.

[0074] It should be noted that when the purification axial force timing signal remains constant throughout the entire acquisition window, making the denominator 0, then... Set to 0. Using dynamic energy relative to the window mean, instead of directly using the sum of squares of axial force samples, reduces the dominance of large DC axial force values ​​on the energy percentage.

[0075] 6) Calculate the absolute difference between adjacent purification axial force values ​​within each conditional time segment, and take the maximum value among all conditional time segments to obtain the maximum absolute difference feature. Specifically, regarding the first... Each conditional time segment is calculated sequentially. Then, the maximum value is taken from the adjacent absolute differences of all conditional time series segments as... .in, It refers to the first In each sample, the status word is... Within all the abnormal trigger intervals corresponding to the bit, the maximum absolute difference between adjacent sampling points of the purification shaft force value. For example, if the second bit of the status word triggers two intervals, where the purification shaft force value in interval 1 is [10.1, 10.5, 10.2], its absolute difference is [0.4, 0.3], with a maximum of 0.4, and the purification shaft force value in interval 2 is [12.0, 14.5, 14.4], its absolute difference is [2.5, 0.1], with a maximum of 2.5, then the maximum absolute difference feature of this status bit is... .

[0076] It should be noted that the maximum absolute difference feature is used to characterize the maximum axial force change between adjacent sampling times during the occurrence of the corresponding abnormal event, and it is highly sensitive to sensor momentary disconnection, poor connection contact and data jump.

[0077] 7) Arrange the state trigger duration ratio, conditional kurtosis characteristics, conditional zero-crossing characteristics, dynamic energy ratio, and maximum absolute difference characteristics in a fixed order to form the state word. The 5-dimensional conditional temporal morphological features corresponding to each position, namely .in, Indicates the first The state word in each sample The conditional temporal morphological features corresponding to the position.

[0078] In the specific implementation, when the state word... If the position is not triggered within the current acquisition window, All 5 elements of the set are set to 0. The set is then... The conditional temporal morphological features corresponding to all state bits are concatenated in ascending order of state bit number to obtain the conditional temporal morphological feature subvector. (Right now It is a set All state bits corresponding to (Constructed by concatenating status digit numbers from smallest to largest), with dimensions of... .

[0079] S203, Construction of second-order co-occurrence and temporal overlap features of state words The occurrence of two abnormal events within the same acquisition window does not necessarily mean that the two abnormal events occurred simultaneously. Using only window-level logic and results cannot distinguish between events that occur separately or simultaneously.

[0080] This invention simultaneously calculates the co-occurrence flag of abnormal events and the temporal overlap ratio of abnormal events, thereby enabling a fine characterization of the temporal concurrency and independence of acquisition link anomalies indicated by multiple status bits. The specific steps are as follows: 1) Based on the fault mechanism and the connection relationship of the acquisition system, pre-set the state word bit combination that needs to be calculated for the second-order interactive features.

[0081] Specifically, the number of state word combinations is defined as , No. The combination of state bits is defined as ,in, ,and .

[0082] In one implementation, a combination of status word bits can be selected, including power failure and wireless link failure, power failure and data verification failure, sensor disconnection and poor connection, wireless link failure and data verification failure, and poor connection and high-frequency oscillation failure. For example, five status word bit combinations can be selected: "Power failure (bit 1) and wireless link failure (bit 2)," "Power failure (bit 1) and data verification failure (bit 4)," "Sensor disconnection (bit 0) and poor connection (bit 3)," "Wireless link failure (bit 2) and data verification failure (bit 4)," and "Poor connection (bit 3) and high-frequency oscillation failure (bit 5)." These bit numbers are for illustrative purposes only.

[0083] 2) Calculate the first... based on the window-level 16-bit state word vector. Co-occurrence flag of state word combinations ,Right now .in, Indicates the status word Bit and status word Each bit has been triggered at least once within the current acquisition window; This indicates that at least one status word has not been triggered.

[0084] 3) Calculate the number of samples based on the 16-bit status word sequence of the sampling level. Time overlap ratio of each status word combination The calculation method is expressed as follows: , in, Indicates the status word Bit and status word The proportion of time a location is simultaneously in a triggered state within the current acquisition window.

[0085] In one embodiment, for example, if both status bits are triggered within the acquisition window, but appear in the first and second halves of the acquisition window respectively, then ,and If two status bits together trigger 30% of the acquisition window length, then and .

[0086] 4) Concatenate the co-occurrence flags and temporal overlap ratios corresponding to all state word combinations in a fixed order to obtain the state word interaction feature sub-vector: , in, Indicates the first The state word interaction feature sub-vectors corresponding to each sample have a dimension of . .

[0087] 5) Concatenate the conditional temporal morphological feature vector and the state word interaction feature vector to obtain the augmented state feature vector, i.e. .in, Indicates the first The augmented state feature vector corresponding to each sample, and the dimension of the augmented state feature vector. .

[0088] In one implementation, four state words are selected to extract conditional temporal morphological features, and five state words are selected in combination to extract state word interaction features. The dimension of the conditional temporal morphological feature subvector is 20, and the dimension of the state word interaction feature subvector is 10, resulting in the final dimension... The augmented state feature vector. For example, in this implementation, four state word bits are selected: bit 0 (sensor disconnection), bit 1 (power supply failure), bit 3 (poor connection), and bit 5 (high-frequency oscillation failure) to extract conditional timing morphological features. Five state word bits are selected in combination: (bit 1, bit 2), (bit 1, bit 4), (bit 0, bit 3), (bit 2, bit 4), and (bit 3, bit 5) to extract state word interaction features.

[0089] It should be noted that this step not only encodes whether abnormal events occur together, but also encodes the actual degree of overlap of abnormal events on the time axis. Among them, the co-occurrence flag is used to preserve the window-level combination relationship, and the time overlap ratio is used to distinguish between independent abnormalities before and after and synchronous compound abnormalities, which is helpful to identify compound faults caused by common abnormalities of multiple acquisition components.

[0090] S3. Dual-branch heterogeneous feature extraction and fault condition mapping The axial force time series signal is continuous time series data, which includes changes at different time scales such as sudden drop in axial force, slow decay, abnormal oscillation and data jump; the window-level 16-bit state word vector, the window-level state duration vector and the augmented state feature vector are window-level conditional data, which include abnormal event type, abnormal duration, axial force morphology during abnormality and abnormal event combination relationship.

[0091] This invention constructs an axial force temporal branch and a state condition branch. The axial force temporal branch extracts axial force temporal features through multi-layer one-dimensional causal dilated convolutional residual blocks. The state condition branch generates a fault condition vector and, based on this vector, generates channel gating vectors and channel soft threshold vectors for each layer of the axial force temporal branch. This ensures that different abnormal conditions correspond to different feature channel retention ratios and response thresholds. Figure 5 As shown, the specific steps are as follows: S301, Generation of fault condition vectors for state-condition branches This step takes the state condition input vector, which contains discrete anomaly indicators, anomaly duration, and anomaly-related axial force morphological features, and performs nonlinear mapping through a multi-layer fully connected network. This compressed vector is then integrated into a compact fault condition vector to guide feature extraction of the axial force temporal branch. The specific steps are as follows: 1) Before entering the state conditional branch, perform feature channel standardization on the window-level state duration vector and the augmented state feature vector using training set statistics. Specifically, the window-level 16-bit state word vector remains 0 or 1 and is not standardized. The window-level state duration vector remains within the physical range of 0 to 1; it can be directly input into the state conditional branch or standardized using training set statistics (Z-score standardization can be used). The kurtosis, energy, and difference features in the augmented state feature vector are standardized using training set statistics (Z-score standardization can be used).

[0092] It should be noted that the same statistics saved in the training phase are used in the verification, testing, and online identification phases, and the standardized parameters are not recalculated based on a single identification acquisition window.

[0093] 2) Concatenate the window-level 16-bit state word vector, the window-level state duration vector, and the augmented state feature vector in a fixed order to obtain the state condition input vector, i.e. .in, Indicates the first The state condition input vector corresponding to each sample has a dimension of . .

[0094] 3) The state condition input vector is passed through two fully connected layers in sequence. Through linear mapping and nonlinear activation, the fault condition vector is obtained, which is expressed as: , in, Indicates the first The fault condition vector corresponding to each sample is used to compactly encode window-level anomaly information and generate channel-level modulation parameters for different layers of the axial force timing branch, with a dimension of [missing information]. ; and This represents the trainable weight matrix of the two fully connected layers in the state-conditional branch; and This represents the corresponding trainable bias vector; This represents the linear rectification activation function.

[0095] In one implementation, the output dimension of the first fully connected layer can be 128, and the output dimension of the second fully connected layer can be... You can choose 64.

[0096] S302, One-dimensional causal dilated convolution feature extraction of axial force temporal branches This step uses cascaded one-dimensional causal dilated convolutional residual blocks to extract multi-scale features from the purified axial force time-series signal. While ensuring temporal causality, it captures axial force variation patterns at different time scales, from instantaneous jumps to long-term trends, layer by layer. The specific steps are as follows: 1) Standardize the purification axial force time series signal using the mean and standard deviation of axial force from the training set (Z-score standardization can be used) to obtain the standardized purification axial force time series signal. Standardized purification axial force timing signal This refers to the purified axial force time series signal after uniform scaling of the training set, with dimensions of [missing information]. Standardization is used to reduce the overall amplitude differences caused by different sensor ranges, different initial bolt preloads, and different installation positions.

[0097] 2) Using the standardized purification axial force timing signal as the initial input to the axial force timing branch, the 0th layer timing feature map is obtained, i.e. .in, Indicates the first The temporal feature map of layer 0 corresponding to each sample has a dimension of . (1 indicates the number of input channels, (Indicates the length of time).

[0098] 3) Construction The process involves cascading one-dimensional causal dilated convolutional residual blocks and inputting the modulation temporal feature map from the previous layer into the next layer's one-dimensional causal dilated convolutional residual block. Specifically, for the first... One-dimensional causal dilated convolution residual block, , will the Layer modulation timing feature map As input; when season .

[0099] In specific implementation, the first The one-dimensional causal dilation convolution residual block consists of two one-dimensional causal dilation convolutional layers with a kernel length of 3. The first one-dimensional causal dilation convolutional layer is followed by a batch normalization layer and a ReLU activation function, and the second one-dimensional causal dilation convolutional layer is followed by a batch normalization layer. The output of the main convolutional branch is added to the output of the residual branch, and then passed through a ReLU activation function to obtain the first... Timing characteristics of the layer without fault condition modulation .

[0100] Specifically, in the l-th layer one-dimensional causal dilated convolution residual block: The main convolution branch output refers to the feature map obtained after the input feature map passes through "the first one-dimensional causal dilation convolutional layer → batch normalization layer → ReLU activation function → the second one-dimensional causal dilation convolutional layer → batch normalization layer" in sequence. This output has not undergone ReLU activation.

[0101] The residual branch output refers to the feature map that is directly passed from the residual block input or after channel alignment. When the number of channels in the modulation timing feature map of layer l-1 is the same as the number of channels in the output of layer l, the residual branch output is the input feature map itself (identity mapping); when the number of channels is different, the residual branch output is the channel-aligned feature map obtained by passing the input feature map through a one-dimensional convolution with a kernel length of 1 (stride of 1).

[0102] It should be noted that the one-dimensional causal dilation convolution only uses the temporal information of the current position and the position before the current position; by supplementing the left side of the convolution input with sampled values ​​that match the dilation rate and the length of the convolution kernel, the time length of the convolution output remains unchanged, so that the model can be directly used for online sliding window recognition and avoids the leakage of temporal information caused by using future sampled values.

[0103] In practical implementation, when the first Number of input channels per layer and the first layer When the number of output channels is the same, the residual branch directly passes the input features; when the number of channels is different, the residual branch uses a one-dimensional convolution with a kernel length of 1 to align the number of channels.

[0104] In one implementation, setting The number of output channels in the four layers are 32, 64, 128, and 256, respectively, corresponding to dilation rates of 1, 2, 4, and 8. The convolutional layer with a dilation rate of 1 focuses on extracting instantaneous jumps between adjacent sampling points, the convolutional layers with dilation rates of 2 and 4 focus on extracting oscillation and decay patterns over a medium time range, and the convolutional layer with a dilation rate of 8 focuses on extracting axial force variation trends over a longer time range.

[0105] S303, Fault Condition-Driven Channel Soft Threshold Selection The absolute threshold is generated directly from the state condition branch, which is easily affected by the difference in feature amplitude of different network layers.

[0106] This invention first generates a channel soft threshold ratio between 0 and a preset upper limit based on the fault condition vector, and then calculates the actual channel soft threshold by combining it with the average absolute response of each channel in the current layer, so that the threshold matches the feature scale of the current channel. The specific steps are as follows: 1) Based on the fault condition vector Generate the first The channel gating vector of the layer. Specifically, the first... Layer The channel gate value corresponding to each channel Determined in the following manner: , in, Represents the time-series feature channel index. ; Indicates the first Number of output channels of the layer; Indicates the first The sample at the th Layer The channel gating value corresponding to each channel; Indicates the minimum channel retention ratio; Indicates the first The layer is used to generate a trainable weight matrix for the channel gating vector; Indicates the first The layer is used to generate trainable bias vectors for channel gating vectors; Indicates the first in the read vector One element; This represents the Sigmoid activation function.

[0107] In one implementation, It can be set to 0.05 to 0.2. Setting a minimum channel retention ratio can prevent a channel from being completely shut down, allowing weak fault characteristics to still retain a certain information transmission path.

[0108] 2) Based on the fault condition vector Generate the first The soft threshold ratio of each channel in each layer. Specifically, the... Layer Channel soft threshold ratio for each channel Determined in the following manner: , in, Indicates the first The sample at the th Layer The channel soft threshold ratio corresponding to each channel; This indicates the upper limit of the channel soft threshold ratio. ; Indicates the first The layer is used to generate a trainable weight matrix for the channel soft threshold ratio; This represents the corresponding trainable bias vector.

[0109] In one implementation, A value of 0.8 can be used. Limiting the upper limit of the channel soft threshold ratio can prevent the actual channel soft threshold from being significantly higher than the typical response amplitude of the current channel, reducing the risk of the entire channel being over-zeroed.

[0110] 3) According to the first The average absolute response of each channel in the layer time series characteristic map is used to calculate the channel amplitude reference value. Specifically, the first... Layer The channel amplitude reference value corresponding to each channel Determined in the following manner: , in, Indicates the first The sample at the th Layer The average absolute response of each channel; Indicates the first The time length of the layer temporal feature map; Indicates the first Layer The first channel in the Feature values ​​at each time location.

[0111] 4) Multiply the channel soft threshold ratio by the corresponding channel amplitude reference value to obtain the actual channel soft threshold, i.e. .in, Indicates the first The sample at the th Layer The actual soft threshold for each channel.

[0112] Therefore, the actual channel soft thresholds of different network layers and different feature channels adapt to the current feature amplitude. Even if shallow high-frequency features and deep trend features have different numerical ranges, feature selection can be completed using a relatively consistent threshold ratio.

[0113] 5) Use the actual channel soft threshold for the first... The layer temporal feature map undergoes channel-by-channel soft thresholding, and then the channel gate value is used to adjust the passing ratio of retained features to obtain the first layer. The layer modulation timing feature map is represented as follows: , in, Indicates the first Layer modulation timing feature map in the first The first channel, the first Feature values ​​at each time location; Represents a symbolic function.

[0114] In practice, the same channel gating value and the same actual channel soft threshold are broadcast along the time dimension of the corresponding channel. The actual channel soft threshold is used to suppress responses with small amplitudes that do not match the current fault conditions; the channel gating value is used to adjust the pass-through ratio of the retained responses.

[0115] In one embodiment, for example, when the window-level 16-bit state word vector indicates that a wireless link anomaly is continuously occurring, the state condition branch can increase the channel soft threshold of the channel related to minor fluctuations in communication interference and decrease the channel gate value of the corresponding channel; when the window-level state word does not record an acquisition anomaly, but there is a significant sudden drop in axial force in the purification axial force timing signal, the state condition branch can maintain the channel related to the sudden change in axial force with a lower channel soft threshold and a higher channel gate value, so that the true structural change characteristics are preserved.

[0116] 6) The first Layer modulation timing feature map As the first The input of the one-dimensional causal dilated convolutional residual block is used to repeatedly perform temporal feature extraction (corresponding to steps 2) and 3) in S302), channel gating value generation (corresponding to step 1) in S303), channel soft threshold calculation (corresponding to steps 2), 3) and 4) in S303), and channel modulation (corresponding to step 5) in S303), until the first... Layer modulation timing feature map .

[0117] It should be noted that both the channel gating vector and the channel soft threshold vector are dynamically generated for each sample, and are not fixed constants after training. The same purification axial force time-series signal can yield different channel selection results under different state conditions, so that the state word information continuously participates in the extraction of time-series features at each layer, rather than being concatenated with time-series features only once in the final classification layer.

[0118] In one implementation, to make the model training process resemble a state where strong feature suppression is not performed during the initial stages, the initial bias of the channel-gated generation layer can be set to a positive value, making the initial channel gating value close to 1; the initial bias of the channel soft-threshold ratio generation layer can be set to a negative value, making the initial channel soft-threshold ratio close to 0. This initialization method can reduce the risk of prematurely deleting temporal features before the network has learned effective fault condition mappings.

[0119] S304, Dual-branch feature fusion and fault probability output This step involves pooling and encoding the multi-scale temporal features extracted from the axial force temporal branch, and then concatenating and fusing them with the state-aided features generated by the state condition branch to form a comprehensive identification feature that combines dynamic changes in axial force with collected state information. Finally, a fully connected classification layer is used to output the predicted probabilities of various faults. The specific steps are as follows: 1) Regarding the first Global average pooling and global max pooling are performed on the layer modulation temporal feature maps respectively.

[0120] Among them, global average pooling calculates the average value of each feature channel at all time positions to preserve stable fault evidence such as slow axial force decay and continuous abnormal oscillation; global max pooling calculates the maximum response of each feature channel at all time positions to preserve short-term sudden fault evidence such as sudden drop in axial force, instantaneous disconnection and connection jump.

[0121] Furthermore, the global average pooling result and the global max pooling result are concatenated to obtain the temporal encoded feature vector. When the first The number of output channels of the layer is At that time, the dimension of the time-encoded feature vector is .

[0122] 2) Transfer the fault condition vector Inputting the classification auxiliary mapping layer yields the state auxiliary feature vector, represented as: , in, Indicates the first The state auxiliary feature vector corresponding to each sample; This represents the trainable weight matrix of the classification auxiliary mapping layer. This represents the trainable bias vector of the classification auxiliary mapping layer.

[0123] It should be noted that the state-assisted feature vector is used to directly retain the type of abnormal event, the degree of abnormality, and the interaction relationship of abnormal events in the final classification stage, so as to avoid over-compression when state condition information is only indirectly transmitted through channel modulation.

[0124] 3) Concatenate the time-encoded feature vector and the state-aided feature vector to obtain the fused identification feature vector, i.e. .in, Indicates the first The fused identification feature vector corresponding to each sample includes the axial force time-series feature modulated by the fault condition and the window-level state condition feature. for transpose; for transpose; This indicates the transpose operation.

[0125] 4) Input the fused feature vector into the classification fully connected layer to obtain the category score vector. Then, the fault prediction probability vector is obtained through the Softmax function, i.e. .in, Indicates the first The fault prediction probability vector corresponding to each sample; Indicates the first The sample belongs to the first Predicted probabilities for each fault category; Indicates the fault category index, ; This represents the Softmax function; This represents the trainable weight matrix of the fully connected layer for classification. This represents the trainable bias vector of the fully connected classification layer.

[0126] In practice, the sum of all elements in the fault prediction probability vector is 1, and the element with the highest probability corresponds to the initial fault identification result of the current sample.

[0127] The specific hierarchical structure corresponding to the dual-branch heterogeneous feature extraction and fault condition mapping operations is shown in Table 1: Table 1. Hierarchical structure table corresponding to dual-branch heterogeneous feature extraction and fault condition mapping operations. S305, Sample reliability weighted training based on the persistence of state word anomalies Sensor disconnections and data verification anomalies reduce the integrity and reliability of the original axial force timing signal. If fixed weighting is applied only based on whether the window-level status word is set, short-term and long-term anomalies will receive the same sample reliability weight, which cannot accurately reflect the degree of data contamination.

[0128] This invention calculates sample reliability weights based on the duration of sensor disconnections and data verification anomalies within the acquisition window. This ensures that samples with longer anomaly durations have less impact on network parameter updates. Simultaneously, a minimum sample reliability weight is set to prevent anomaly samples containing valid fault information from being completely ignored. The specific steps are as follows: S3051, Sample Reliability Weight Calculation This step dynamically calculates the reliability weight of each sample based on the persistence ratio of two states that directly affect data reliability: sensor disconnection and data verification anomalies. This ensures that samples with more severe data contamination contribute less to the training process. The specific steps are as follows: 1) Read the state duration ratio, which is directly related to data reliability, from the window-level state duration vector. When reading, directly extract the percentage from the window-level state duration vector. Read the value at a specific index position.

[0129] In one implementation, bit 0 of the status word corresponds to a sensor disconnection, and bit 3 corresponds to a data verification error. (Read) and ,in, Indicates the first The percentage of samples where the sensor remained offline for an extended period Indicates the first The proportion of data verification anomalies persisting in each sample.

[0130] 2) Based on the persistence ratio of the two types of reliability anomalies and the corresponding deweighting strength parameters, calculate the first... The sample reliability weight of each sample is expressed as: , in, Indicates the first The sample reliability weights corresponding to each sample; Represents the minimum sample reliability weight. ; This indicates the weighting intensity parameter corresponding to the sensor disconnection. This parameter represents the weighting intensity corresponding to data verification anomalies.

[0131] In specific implementation, when and hour, This indicates that the current sample participates in training with full weights. When the proportion of sensor disconnections or data verification anomalies continues to increase, Gradually decrease, but will not fall below... .

[0132] In one implementation, , , .like and The sample reliability weight is approximately 0.441. The deweighting intensity parameter corresponding to sensor disconnection is higher than that corresponding to data verification anomaly, reflecting that sensor disconnection usually has a more direct impact on the integrity of continuous axial force waveform.

[0133] 3) Associating sample reliability weights with corresponding training samples is only used during the training loss calculation stage.

[0134] It should be noted that the sample reliability weight does not change the purification shaft force time series signal, does not modify the real fault category label, and is not directly multiplied into the fault prediction probability during the online identification stage. It only controls the degree of contribution of different training samples to the network parameter update.

[0135] S3052, Category Equalization and Reliability Weighted Loss Calculation This step introduces both class balance weights and sample reliability weights when calculating cross-entropy loss. The former is used to alleviate class imbalance, while the latter is used to reduce the negative impact of unreliable samples. Normalization is then used to ensure the numerical stability of the training process. The specific steps are as follows: 1) Convert the true fault category labels of the training samples into one-hot label vectors, i.e. .in, Indicates the first The true one-hot label vector corresponding to the nth sample; when the nth sample... The sample belongs to the first When there are multiple fault categories The remaining positions are set to 0.

[0136] 2) Count the number of samples in each fault category in the training set and set the category balance weights according to the number of samples in each category.

[0137] Specifically, the first The number of training samples for each fault category is defined as The maximum number of samples across all fault categories is defined as Then the first Category balance weights corresponding to each fault category It can be determined in the following way: , in, Indicates the first Category balance weights corresponding to each fault category; This represents a function that takes the minimum value. This represents the upper limit of the category balance weight, which can be between 2 and 5.

[0138] It should be noted that when the training sample class distribution is extremely imbalanced, using the square root method instead of directly using the inverse ratio of sample size can reduce the risk of training instability caused by a very small number of classes having excessively large weights. Furthermore, the class balance weights are obtained solely from statistics on the training set.

[0139] 3) Define the number of samples contained in a training batch as... Based on the sample reliability weights, the class balance weights corresponding to the true fault categories, and the fault prediction probabilities for each sample, the overall training loss is calculated, expressed as: , in, This represents the reliability-weighted cross-entropy loss corresponding to the current training batch; Indicates the number of samples in the training batch; Indicates the first The sample reliability weights corresponding to each sample; Indicates the first Category balance weights corresponding to each fault category; This represents the actual unique hot tag value; This represents the probability of fault prediction.

[0140] It should be noted that in each training batch, the denominator is normalized using the sum of the effective weights in the current training batch, so that the numerical scale of the overall loss remains relatively stable when the number of unreliable samples changes in different training batches.

[0141] It should also be noted that sample reliability weights and class balance weights serve different purposes. Sample reliability weights are used to reduce the impact of samples contaminated by sensor disconnections or data verification anomalies, while class balance weights are used to increase the attention given to a few fault classes during training.

[0142] S3053, Model Parameter Update and Optimal Model Saving This step, based on the calculated loss, iteratively updates the trainable parameters of the model using the gradient backpropagation algorithm and the Adam optimizer. After each training epoch, the model performance is evaluated by the macro-averaged F1 score on the validation set. Finally, the model with the best generalization ability is saved. The specific steps are as follows: 1) Perform S2 and S3 sequentially for each training sample to obtain the corresponding fault prediction probability vector, and calculate the reliability-weighted cross-entropy loss of the current training batch according to S402.

[0143] 2) Perform gradient backpropagation based on the reliability-weighted cross-entropy loss to update the trainable network parameters other than the prior parameters of the state word fault band. The trainable network parameters involved in the update include the fully connected layer parameters of the state condition branch, the convolution parameters and batch normalization parameters of each one-dimensional causal dilated convolutional residual block in the axial force temporal branch, the channel gating generation parameters of each layer, the soft thresholding ratio generation parameters of each layer, the classification auxiliary mapping layer parameters, and the classification fully connected layer parameters.

[0144] It should be noted that the fault-sensitive center frequency, bandwidth, minimum bandwidth retention ratio, and basic shaping strength parameters can be fixed according to the pre-experiment, or determined by finite parameter search on the validation set, and do not participate in gradient backpropagation.

[0145] 3) Update all trainable network parameters using the Adam optimizer.

[0146] In one implementation, the initial learning rate can be 0.001, the training batch size can be 32 or 64, and the number of training epochs can be 100 to 200. If the validation set loss does not decrease for 5 consecutive training epochs, the current learning rate is multiplied by 0.5; if the macro-average F1 score of the validation set does not improve for 10 consecutive training epochs, training can be stopped early.

[0147] 4) After each training cycle, calculate the precision, recall and F1 score for each fault category, and take the arithmetic mean of the F1 scores for all fault categories to obtain the macro average F1 score.

[0148] Furthermore, the network parameters corresponding to the highest macro-average F1 score of the validation set are saved to obtain the bolt axial force monitoring fault type identification model.

[0149] It should be noted that the macro-average F1 score assigns equal importance to each fault category, which can prevent the model from achieving a high overall accuracy by only improving the recognition rate of normal categories when there are a large number of normal samples, while ignoring a few categories such as sudden drop in axial force, poor connection contact, or compound abnormalities.

[0150] S4. Online fault identification for bolt axial force monitoring After model training is completed, the trained bolt axial force monitoring fault type identification model is deployed on bolt axial force online monitoring terminals, edge computing devices, or servers. During the online operation phase, real-time monitoring data is processed using the same sampling rate, acquisition window length, status word definition, and feature order as during the training phase. The specific steps are as follows: 1) Extract the current acquisition window to be identified from the real-time axial force data stream to obtain the original axial force timing signal to be identified. and the corresponding 16-bit state word sequence of the sampling level to be identified The duration and number of sampling points of the original axial force timing signal to be identified remain consistent with those in the training phase. The online acquisition window can slide continuously in fixed steps; for example, the acquisition window length is 2 seconds, and the window sliding step is 0.5 seconds.

[0151] 2) Perform bitwise summarization and duration proportional statistics on the 16-bit status word sequence to be identified at the sampling level to obtain the 16-bit status word vector at the window level to be identified. and the duration vector of the window-level state to be identified .

[0152] 3) Based on the wavelet basis, wavelet packet decomposition level, correspondence between state word position and fault-sensitive frequency band, frequency band influence width, shaping intensity parameters, and minimum frequency band retention ratio stored in S201, perform frequency band selective shaping and inverse wavelet packet reconstruction on the original axial force time series signal to be identified, to obtain the axial force time series signal to be purified. .

[0153] When all state words involved in frequency band selective shaping are not triggered, the retention coefficients of all end wavelet packet nodes are 1, and the original axial force timing signal to be identified is not subject to additional frequency band attenuation.

[0154] 4) Based on the state word selection range, state word combination, and feature arrangement order stored in S202 and S203, extract conditional timing morphological features and state word interaction features from the axial force timing signal to be identified and the 16-bit state word sequence of the sampling level to be identified, to obtain the augmented state feature vector to be identified. .

[0155] 5) Using the standardized statistics of axial force and state features saved during the training phase, perform the same standardization process as during the training phase on the axial force time series signal to be identified, the window-level state duration vector to be identified, and the augmented state feature vector to be identified.

[0156] During the online identification phase, the standardized range must not be recalculated based on a single acquisition window to be identified; otherwise, the relative amplitude of the same axial force change in different acquisition windows will be altered, reducing the consistency of continuous identification results.

[0157] 6) Input the axial force timing signal to be identified into the axial force timing branch, and input the 16-bit state word vector at the window level, the state duration vector at the window level, and the augmented state feature vector to be identified into the state condition branch. Then, sequentially complete the fault condition vector generation, one-dimensional causal dilation convolution feature extraction, channel soft threshold selection, dual-branch feature fusion, and fault probability calculation to obtain the fault prediction probability vector to be identified. .in, This represents the fault prediction probability vector corresponding to the current acquisition window to be identified; This indicates that the current window to be identified belongs to the first... The predicted probability of each fault category.

[0158] 7) Read the maximum predicted probability from the fault prediction probability vector to be identified, and determine whether the current fault category should be output in the current acquisition window based on the confidence threshold. Define the confidence threshold as follows: , This represents the minimum prediction probability required to accept the current fault identification result, and can be between 0.6 and 0.8. When the maximum prediction probability is less than... When the probability of failure is not less than 1, a low confidence level alert is output, but a formal fault alarm is not immediately output; when the maximum predicted probability is not less than 1, the alert is displayed. When the maximum predicted probability is used, the fault category corresponding to the fault category in the current acquisition window is selected as the candidate fault category.

[0159] In one embodiment, for example, if: , and The maximum predicted probability is 0.62, corresponding to the third fault category. The candidate fault category in the current acquisition window to be identified is sudden drop in axial force.

[0160] For example, if the maximum predicted probability is 0.38, which is less than the confidence threshold of 0.6, a low confidence warning will be output, and subsequent collection windows to be identified will continue to be collected.

[0161] 8) Perform time consistency verification on candidate fault categories for multiple consecutive data acquisition windows. Specifically, the number of consecutive verification windows is defined as... , This indicates the number of consecutive acquisition windows with the same candidate fault category that need to appear before a formal fault alarm is output; it can be 3 to 5.

[0162] When the same candidate fault category is in succession Each of the data collection windows to be identified corresponds to the category with the highest predicted probability, and the corresponding highest predicted probability is not less than the confidence threshold. When the candidate fault category changes frequently in adjacent data collection windows, a formal fault alarm will not be output, and subsequent data collection will continue.

[0163] For the sensor disconnection or data verification anomaly rate that is consistently higher than the preset limit for the acquisition window to be identified, a data reliability prompt can be output simultaneously, enabling maintenance personnel to distinguish between actual axial force anomalies and acquisition link anomalies.

[0164] Example 2 This embodiment provides a fault type identification device for bolt axial force acquisition signals, which executes the fault type identification method for bolt axial force acquisition signals described in Embodiment 1, including: Data acquisition module: used to collect bolt axial force monitoring samples, label fault categories, obtain labeled samples, and form a training sample set; The data processing module is used to determine the fault-sensitive frequency band and its processing intensity using the window-level 16-bit state word vector and the window-level state duration vector; to perform continuous soft attenuation on the wavelet packet node coefficients of the original axial force time series signal to obtain the purified axial force time series signal; to determine the abnormal triggering interval based on the sampling-level 16-bit state word sequence; and to extract the conditional time series morphological features and abnormal event interaction features from the purified axial force time series signal to obtain the augmented state feature vector. Model building module: used to build a dual-branch heterogeneous feature extraction and fault condition mapping model, including axial force time-series branch and state condition branch; The fault prediction module is used to collect the original axial force timing signal to be identified and the corresponding 16-bit status word sequence of the sampling level to be identified in real time. It performs bit-by-bit aggregation and duration ratio statistics on the 16-bit status word sequence of the sampling level to be identified to obtain the 16-bit status word vector of the window level to be identified and the state duration rate vector of the window level to be identified. After processing in step S2, it is input into the trained dual-branch heterogeneous feature extraction and fault condition mapping model to obtain the fault type prediction result.

[0165] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for fault type identification of bolt axial force acquisition signals, characterized in that, Includes the following steps: S1. Collect bolt axial force monitoring samples, label the fault categories, and obtain labeled samples to form a training sample set; the labeled samples include the original axial force time sequence signal, the sampling level 16-bit status word sequence, the window level 16-bit status word vector, the window level status duration vector, and the real fault category label. S2. Determine the fault-sensitive frequency band and its processing intensity using the window-level 16-bit state word vector and the window-level state duration vector. Perform continuous soft attenuation on the wavelet packet node coefficients of the original axial force time sequence signal to obtain the purified axial force time sequence signal. The abnormal triggering interval is determined based on the 16-bit status word sequence of the sampling level. Conditional timing morphological features and abnormal event interaction features are extracted from the purification axial force timing signal to obtain the augmented state feature vector. S3. Construct a dual-branch heterogeneous feature extraction and fault condition mapping model, including an axial force time-series branch and a state condition branch; input the purified axial force time-series signal, a window-level 16-bit state word vector, a window-level state duration vector, and an augmented state feature vector into the model to obtain a fault prediction probability vector; train and optimize the model using a loss function to obtain a trained dual-branch heterogeneous feature extraction and fault condition mapping model. S4. Real-time acquisition of the original axial force timing signal to be identified and the corresponding 16-bit status word sequence of the sampling level to be identified. Perform bit-by-bit summarization and duration ratio statistics on the 16-bit status word sequence of the sampling level to be identified to obtain the 16-bit status word vector of the window level to be identified and the state duration rate vector of the window level to be identified. After processing in step S2, input the trained dual-branch heterogeneous feature extraction and fault condition mapping model to obtain the fault type prediction result.

2. The fault type identification method for bolt axial force acquisition signals according to claim 1, characterized in that, Step S1 specifically includes: An axial force sensor is placed on the bolt to be monitored, and axial force sampling values ​​are continuously collected according to a preset sampling rate. The acquisition window is divided into segments with a fixed time length to obtain the original axial force time sequence signal. ; While acquiring each axial force sample value, the 16-bit status word corresponding to the current sampling time is read to obtain a sampling-level 16-bit status word sequence that is time-aligned with the original axial force timing signal. ; Window-level summarization is performed on each state word bit in the sampled 16-bit state word sequence to obtain a window-level 16-bit state word vector. and window-level state duration vector .

3. The fault type identification method for bolt axial force acquisition signals according to claim 1, characterized in that, In step S2, obtaining the purification axial force timing signal includes: For the first The original axial force time-series signal of each sample implement Layer wavelet packet decomposition yields the th The end wavelet packet node coefficients of the layer; the first Layer The node coefficients of each terminal wavelet packet are defined as follows: According to the axial force sampling rate Wavelet packet decomposition level Determine the center frequency corresponding to each terminal wavelet packet node. ; From the 16 status word bits, select the status word bits that are associated with frequency band interference, and establish the status word bits and fault-sensitive center frequency. And bandwidth The correspondence between them; based on the distance between the center frequency of the terminal wavelet packet node and the center frequency of each fault-sensitive node, calculate the frequency band correlation coefficient of each terminal wavelet packet node for different abnormal events. ; By combining the anomaly duration ratio, frequency band correlation coefficient, and shaping intensity parameter in the window-level state persistence vector, the retention coefficient of each terminal wavelet packet node is calculated. ; Perform element-wise scaling on the corresponding terminal wavelet packet node coefficients using the retained coefficients to obtain the shaped wavelet packet node coefficients. ; Perform inverse wavelet packet transform on all integer wavelet packet node coefficients according to the original wavelet packet tree structure, and trim or pad the reconstruction result to the required length. The timing signal of the axial force for purification is obtained.

4. The fault type identification method for bolt axial force acquisition signals according to claim 1, characterized in that, In step S2, the conditional time series morphological features obtained include: The set of state bits involved in the construction of conditional temporal morphological features is defined as... For the state word constructed based on the temporal morphological characteristics of the participating conditions, the intervals with consecutive values ​​of 1 are searched according to the sampling time sequence to obtain the set of abnormal trigger intervals. Extract the abnormal trigger interval from the purification axial force timing signal. The corresponding continuous sampled values ​​are used to obtain conditional time series segments. ; The state triggering duration ratio is calculated based on the length of each conditional time segment, and the local kurtosis coefficient of each conditional time segment is calculated separately. The state word is obtained by weighting and summing the data according to the length of the conditional time sequence segments. Conditional kurtosis characteristics corresponding to position ; For each conditional time series segment, first subtract the mean of the corresponding segment, then calculate the proportion of adjacent centered sampled values ​​whose signs change, thus obtaining the centered cross-zero density. The conditional zero-crossing features are obtained by weighting the centered cross-zero density according to the length of each conditional time series segment. ; Based on the fluctuation energy of the purification axial force time-series signal relative to the window mean, the dynamic energy proportion during abnormal triggering is calculated; the absolute difference between adjacent purification axial force values ​​is calculated within each conditional time-series segment, and the maximum value among all conditional time-series segments is taken to obtain the maximum absolute difference feature. ; The state trigger duration ratio, conditional kurtosis characteristics, conditional zero-crossing characteristics, dynamic energy ratio, and maximum absolute difference characteristics are arranged in a fixed order to form the state word. The 5-dimensional conditional temporal morphological features corresponding to the position ; set The conditional temporal morphological features corresponding to all state bits are concatenated in ascending order of state bit number to obtain the conditional temporal morphological feature subvector. .

5. The fault type identification method for bolt axial force acquisition signals according to claim 1, characterized in that, In step S2, the augmented state feature vector is obtained as follows: Based on the fault mechanism and the connection relationship of the acquisition system, the state word combination that needs to be calculated for the second-order interaction feature is pre-set; the second-order interaction feature is calculated based on the window-level 16-bit state word vector. Co-occurrence flag of state word combinations Calculate the first 16-bit status word sequence based on the sampling level. Time overlap ratio of each status word combination The co-occurrence flags and temporal overlap ratios corresponding to all state word combinations are concatenated in a fixed order to obtain the state word interaction feature sub-vector. The conditional temporal morphological feature vector and the state word interaction feature vector are concatenated to obtain the augmented state feature vector. .

6. The fault type identification method for bolt axial force acquisition signals according to claim 1, characterized in that, In step S3, the axial force temporal branch extracts axial force temporal features through multi-layer one-dimensional causal dilated convolutional residual blocks, and the state condition branch generates a fault condition vector, including: The window-level 16-bit state word vector, the window-level state duration vector, and the augmented state feature vector are concatenated in a fixed order to obtain the state conditional input vector. The state condition input vector is passed through two fully connected layers in sequence, and the fault condition vector is obtained through linear mapping and nonlinear activation. ; The purification shaft force time series signal is standardized using the shaft force mean and standard deviation from the training set. Using the standardized purification axial force timing signal as the initial input to the axial force timing branch, the timing feature map of layer 0 is obtained. ; build The process involves cascading one-dimensional causal dilated convolutional residual blocks, with the modulation temporal feature map from the previous layer inputting into the next layer's one-dimensional causal dilated convolutional residual block, ultimately yielding the first... Layer modulation timing feature map .

7. The fault type identification method for bolt axial force acquisition signals according to claim 6, characterized in that, Get the first Layer modulation timing feature map ,include: Based on the fault condition vector Generate the first Layer channel gating vector According to the fault condition vector Generate the first Channel soft threshold ratio for each channel in the layer According to the The average absolute response of each channel in the layer time series feature map is used to calculate the channel amplitude reference value. Multiply the channel soft threshold ratio by the corresponding channel amplitude reference value to obtain the actual channel soft threshold. ; Use actual channel soft threshold for the first The layer temporal feature map undergoes channel-by-channel soft thresholding, using channel gate values ​​to adjust the pass-through ratio of retained features, thus obtaining the first... Layer modulation timing feature map ; will the first Layer modulation timing feature map As the first The input to the one-dimensional causal dilated convolution residual block is used to repeat the above operation to obtain the first... Layer modulation timing feature map .

8. The fault type identification method for bolt axial force acquisition signals according to claim 1, characterized in that, In step S3, the fault prediction probability vector is obtained as follows: For the first Global average pooling and global max pooling are performed on the layer-modulated temporal feature map respectively. The results of global average pooling and global max pooling are concatenated to obtain the temporal encoded feature vector. ; Fault condition vector Input the classification auxiliary mapping layer to obtain the state auxiliary feature vector. ; The time-encoded feature vector and the state-aided feature vector are concatenated to obtain the fused identification feature vector. ; The fused feature vectors are input into the fully connected classification layer to obtain the category score vectors. The category score vector is used to obtain the fault prediction probability vector through the Softmax function. .

9. The fault type identification method for bolt axial force acquisition signals according to claim 1, characterized in that, In step S3, the loss function includes: The proportion of state duration directly related to data reliability is read from the window-level state duration vector. This reading is done directly from the window-level state duration vector. The system reads the duration of sensor disconnection and the duration of data verification anomalies; based on the duration of these two types of reliability anomalies and their corresponding weighting parameters, it calculates the... Sample reliability weights for each sample; Convert the real fault category labels of the training samples into one-hot label vectors, count the number of samples contained in each fault category in the training set, and set the class balance weights according to the number of samples in each category. The overall training loss is calculated based on the sample reliability weight, the class balance weight corresponding to the true fault category, and the fault prediction probability for each sample.

10. A fault type identification device for bolt axial force acquisition signals, comprising the fault type identification method for bolt axial force acquisition signals as described in any one of claims 1-9, characterized in that, include: Data acquisition module: used to collect bolt axial force monitoring samples, label fault categories, obtain labeled samples, and form a training sample set; Data processing module: used to determine the fault-sensitive frequency band and its processing intensity using window-level 16-bit state word vector and window-level state duration vector, and to perform continuous soft attenuation on the wavelet packet node coefficients of the original axial force time sequence signal to obtain the purified axial force time sequence signal; The abnormal triggering interval is determined based on the 16-bit status word sequence of the sampling level. Conditional timing morphological features and abnormal event interaction features are extracted from the purification axial force timing signal to obtain the augmented state feature vector. Model building module: used to build a dual-branch heterogeneous feature extraction and fault condition mapping model, including axial force time-series branch and state condition branch; The fault prediction module is used to collect the original axial force timing signal to be identified and the corresponding 16-bit status word sequence of the sampling level to be identified in real time. It performs bit-by-bit aggregation and duration ratio statistics on the 16-bit status word sequence of the sampling level to be identified to obtain the 16-bit status word vector of the window level to be identified and the state duration rate vector of the window level to be identified. After processing in step S2, it is input into the trained dual-branch heterogeneous feature extraction and fault condition mapping model to obtain the fault type prediction result.

Images