A crane steel wire rope damage identification and prediction method

Abstract

A method for identifying and predicting damage to crane wire ropes includes: deploying sensors along the axial direction of the wire rope to collect the original magnetic field signal of the wire rope to be tested; preprocessing the original magnetic field signal to obtain a preprocessed magnetic field signal; performing continuous wavelet transform on the preprocessed magnetic field signal to extract time-frequency features; inputting the time-frequency features into a pre-trained deep magnetic field perception learning model, which processes the features and outputs damage identification data and damage prediction data; and obtaining the damage identification result and damage prediction result of the wire rope based on the damage identification data and damage prediction data. This method can perform high-sensitivity non-destructive testing of wire ropes in real time, identifying minute physical damage such as corrosion, wear, fracture, and localized plastic deformation, and can predict the damage state of the wire rope at a future time based on the historical evolution data of the identified minute damage.

Description

Technical Field

[0001] This invention relates to the field of crane safety operation technology, specifically to a method for identifying and predicting crane wire rope damage. Background Technology

[0002] As a critical component bearing significant loads, the safety of crane wire ropes directly impacts equipment operation and personnel safety. Existing wire rope inspection technologies primarily rely on manual inspection, ultrasonic testing, and X-rays for physical damage detection. However, these methods are typically time-consuming and labor-intensive, lack real-time performance, have low accuracy, and cannot predict potential damage. Therefore, achieving real-time monitoring and damage prediction of wire ropes has become a key technical challenge for improving crane safety and operational efficiency. Summary of the Invention

[0003] This invention provides a method for identifying and predicting damage to crane wire ropes, which can solve one of the problems existing in the background art.

[0004] The present invention provides a method for identifying and predicting damage to crane wire ropes, comprising: deploying sensors along the axial direction of the wire rope to collect the original magnetic field signal of the wire rope to be tested; preprocessing the original magnetic field signal to obtain a preprocessed magnetic field signal; performing continuous wavelet transform on the preprocessed magnetic field signal to extract time-frequency features; inputting the time-frequency features into a pre-trained deep magnetic field perception learning model, wherein the deep magnetic field perception learning model processes the features and outputs damage identification data and damage prediction data; and obtaining the damage identification result and damage prediction result of the wire rope based on the damage identification data and damage prediction data.

[0005] The system comprises multiple sensors. The raw magnetic field signals collected by the multiple sensors are spliced ​​together to form raw multi-channel time-series data. The raw magnetic field signal from each sensor serves as a single-channel time-series data. The preprocessing involves processing each single-channel timing data... Filtering, denoising, normalization, and sliding window segmentation are performed separately to obtain the preprocessed time-series data for a single channel. , For sensor indexing, For the time of data collection, For the first The position coordinates of each sensor.

[0006] The time-frequency feature is a multi-channel time-frequency feature map, and the extraction steps of the multi-channel time-frequency feature map are as follows:

[0007] From the preprocessed time series data of a single channel The amplitude sequence was extracted. Frequency component sequence and phase sequence ;

[0008] The amplitude sequence of the extracted single channel Frequency component sequence and phase sequence Continuous wavelet transform (CWT) is performed on each component to extract the corresponding time-frequency features, resulting in amplitude wavelet time-frequency plots. Frequency wavelet time-frequency diagram Phase wavelet time-frequency diagram ;

[0009] Then, they are concatenated along the feature channel dimension to obtain a single-channel multi-feature time-frequency map. ;

[0010] The single-channel multi-feature time-frequency maps from multiple channels are stitched together to obtain a multi-channel time-frequency feature map. for,

[0011] In the formula, The number of channels is equal to the number of sensors.

[0012] The amplitude wavelet time-frequency diagram It is obtained by the following formula.

[0013] In the formula, As a scale, For time, represents the relative time displacement of the local time-frequency characteristics within the window. The generating function of the intelligent extraction engine. To intelligently extract the engine's scale factor, To intelligently extract engine displacement factors,

[0014] The frequency wavelet time-frequency diagram It is obtained by the following formula.

[0015] The phase wavelet time-frequency diagram It is obtained by the following formula. .

[0016] The pre-trained deep magnetic field sensing learning model includes: a cascaded feature extraction module, a temporal modeling module, a damage identification module, and a trend prediction module;

[0017] The feature extraction module is used to process the multi-channel time-frequency feature maps of the input data of the deep magnetic field sensing learning model. The characteristic response intensity is obtained by processing. With time-series input sequence The output will show that 1 and 2 are the time step indices. The time window length,

[0018] The time-series modeling module is used to receive the time-series input sequence output by the feature extraction module. The hidden state feature vector is then processed to obtain the hidden state feature vector. And output, For time step index;

[0019] The damage recognition module is used to receive the feature response intensity output by the feature extraction module. The hidden state feature vector output by the time series modeling module The data is then processed to obtain the damage identification data at time t. And according to damage identification data The calculation method involves calculating the damage identification data from historical time t−T+1 to t−1, and then concatenating these data to obtain the damage identification time series. The damage identification module outputs the damage identification data at time t. With damage identification time series ;

[0020] The trend prediction module is used to receive the damage identification time series output by the damage identification module. The data is then processed to obtain damage prediction data. , Step size;

[0021] The damage identification data With damage prediction data As the output of the deep magnetic field sensing learning model.

[0022] The feature extraction module includes: a cascaded spatiotemporal feature extraction layer, a spatiotemporal feature mapping layer, and a spatiotemporal feature pooling layer;

[0023] The spatiotemporal feature extraction layer consists of one or more convolutional layers, used to process multi-channel time-frequency feature maps. Perform convolution operations to obtain the convolution output. And output;

[0024] The first convolutional layer of the spatiotemporal feature extraction layer is also used to statistically obtain basic quantization indicators based on all time-frequency energy values ​​within the convolutional window, including: local energy mean. Energy variance Peak frequency With pulse duration and output it;

[0025] The spatiotemporal feature mapping layer includes a ReLU activation layer and a concatenation layer. The ReLU activation layer is used to receive the convolution operation output from the spatiotemporal feature extraction layer. Then, a nonlinear activation process is performed on it to obtain the spatiotemporal feature map. ;

[0026] The splicing layer is used to receive the local energy mean output by the spatiotemporal feature extraction layer. Energy variance Peak frequency With pulse duration And map it with spatiotemporal features Channel splicing is performed to obtain three-dimensional spatiotemporal frequency characteristics. ;

[0027] The spatiotemporal feature pooling layer comprises a cascaded frequency pooling layer, a spatial pooling layer, and a stitching layer. The frequency pooling layer is composed of global max pooling or global average pooling and is used to perform three-dimensional spatiotemporal frequency features. Frequency aggregation is performed to obtain a two-dimensional spatiotemporal feature vector sequence. The spatial pooling layer consists of global max pooling or global average pooling, and is used to process two-dimensional spatiotemporal feature vector sequences. Spatial aggregation is performed to obtain a one-dimensional feature vector. , For time step index, , The length of the time window;

[0028] The splicing layer is used to... The corresponding one-dimensional feature vector The sequences are concatenated to form a time-series input sequence. ;

[0029] Two-dimensional spatiotemporal feature vector sequence As characteristic response intensity ,Right now Output feature response intensity from spatiotemporal feature pooling layer and timing input sequence .

[0030] The convolutional layers in the spatiotemporal feature extraction layer are calculated according to the following formula.

[0031] In the formula, For the first Layer convolution kernel in the th Weight parameters on each channel, Multi-channel time-frequency feature map The Middle Each channel at time Input value at the location, For bias terms, The output is the result of the convolution operation.

[0032] The first convolutional layer of the spatiotemporal feature extraction layer is also used to statistically obtain the basic quantization index based on all time-frequency energy values ​​within the convolutional window.

[0033] Let the convolution window size be M×N, and the set of energy for all pixels within the convolution window be... ,

[0034] Then, the local average energy can be calculated according to the following formula. for,

[0035] In the formula, Represents any single energy element within the window.

[0036] The energy variance is calculated using the following formula. for,

[0037] And, by traversing the window energy set Take the maximum energy value The corresponding frequency scale coordinates are used as the peak frequency. ,in,

[0038] Furthermore, by statistically analyzing whether the energy within the window exceeds a preset damage threshold. The pulse duration is obtained by taking the continuous time-domain sampling length. for, .

[0039] The time series modeling module is built based on a first Long Short-Term Memory (LSTM) network and is used to process the time series input sequence. Transform into hidden state feature vectors And output from the time series modeling module;

[0040] The damage identification module includes: a damage category identification submodule, a damage location identification submodule, and a damage severity assessment submodule connected in parallel, and a connection layer; the inputs to the damage category identification submodule and the damage severity assessment submodule are hidden state feature vectors. The input to the damage location identification submodule is the feature response intensity. The outputs of the damage category identification submodule, the damage location identification submodule, and the damage severity assessment submodule are collectively used as the input to the connection layer.

[0041] The damage category identification submodule consists of a fully connected layer and a Softmax layer. First, it hides the state feature vector. The damage category probability vector is obtained by sequentially processing through a fully connected layer and a Softmax layer, and then calculating using the following formula. for,

[0042] In the formula, Weight matrix for identifying damage categories, To identify the bias vector for the damage category, ,in , , They are respectively the first type, the second type, and the third type. The probability corresponding to each damage type, where t is time t. This represents the total number of categories of wire rope damage.

[0043] The damage location identification submodule is composed of a cascaded feature response decoupling unit and a spatial weighted centroid calculation unit, used to analyze the feature response intensity. The damage location is estimated by calculating using the following formula. for,

[0044] In the formula, For the number of sensors, Representing the The strength of the magnetic field anomaly at the spatial location of each sensor;

[0045] The damage assessment submodule is built based on a regression network and also includes a normalization submodule for processing the hidden state feature vector. The process is performed to obtain a damage assessment vector. for,

[0046] In the formula, For normalization function, Assess the weighting of the degree of damage. As a bias for assessing the degree of damage, , , , These are the damage severity assessment values ​​corresponding to the first, second, and C damage categories, respectively.

[0047] The connection layer is used to process the damage category probability vector. Damage location estimate Damage severity assessment value By stitching the data together, damage identification data at time t is obtained. Furthermore, the damage identification data from historical time t−T+1 to t−1 are spliced ​​together to obtain the damage identification time series. .

[0048] The trend prediction module includes a cascaded second Long Short-Term Memory (LSTM) network and a fully connected linear layer, which will sequence the damage identification time series. The input to the second Long Short-Term Memory network is processed to obtain the hidden state feature vector of the future damage prediction module. The hidden state feature vector The fully connected linear layer is input and processed using the following formula to obtain the damage prediction result. for,

[0049] In the formula, Indicates the future The prediction results for the time period, For the corresponding time step The weight parameters, For bias terms, For the future The probability vector of the damage category at time step, , , , In the future The probabilities of the first, second, and c-th damage types at time point c. Step size, For the future Predicted damage location at time [time]. For the future The damage prediction vector at time step. , , , For the future The first, second, and third time Predicted damage severity values ​​corresponding to different damage categories.

[0050] The step of obtaining the damage identification and prediction results for the wire rope based on damage identification and prediction data includes: selecting the damage category with the highest probability from the damage category probability vector as the identified damage category, and combining the estimated damage degree and damage location corresponding to this identified damage category to form the damage identification result for the wire rope.

[0051] In the future The damage category with the highest probability is selected from the damage category probability vector at any given time as the predicted damage category. The predicted damage level and the estimated damage location corresponding to this predicted damage category are then combined to form the damage prediction result for the wire rope.

[0052] The method further includes: deriving a remaining life assessment based on the damage prediction results, including:

[0053] First, the preset damage level failure threshold vector is: , , , These are the failure thresholds for damage categories 1, 2, and C, respectively.

[0054] Then, the future damage prediction module calculates the future damage. to Damage severity prediction sequence at time step This is combined with a preset damage severity failure threshold vector to calculate the remaining life of the wire rope under each individual damage category. for,

[0055] In the formula, Representative at the The remaining life of a wire rope under a single damage category. , The time interval corresponding to the k-th prediction step size. For the future Damage prediction vector at time step The Middle Predicted damage severity values ​​for each damage category For the first Failure threshold for each damage category Indicates the predicted value of damage level The first time the damage level or failure threshold is reached or exceeded The corresponding time interval, the remaining life of the wire rope after a single damage. For only the first The estimated time from the current moment until the wire rope reaches the scrap condition under each type of damage;

[0056] Finally, the remaining life of the wire rope under each single type of damage. The minimum value in the range is taken as the comprehensive remaining life of the wire rope, and is expressed as: .

[0057] The crane wire rope damage identification and prediction method provided by this invention extracts time-frequency features from the magnetic field signal using continuous wavelet transform to reflect the minute damage on and inside the wire rope. Then, it analyzes the time-frequency features using a pre-trained deep magnetic field perception learning model built based on deep learning and time series analysis technology. This enables real-time, highly sensitive non-destructive testing of the wire rope, identifying minute physical damage such as corrosion, wear, fracture, and local plastic deformation. Furthermore, it can predict the damage state of the wire rope at a future moment based on the historical evolution data of the identified minute damage. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0059] Figure 2 This is a schematic diagram of the deep magnetic field sensing learning model of the method of the present invention.

[0060] Figure 3 This is a schematic diagram of the feature extraction module in the deep magnetic field sensing learning model of the present invention.

[0061] Figure 4 This is a schematic diagram of the intelligent magnetic field sensing array composed of multiple sensor units, which is an embodiment of the method of the present invention.

[0062] Figure 5 This is a visual interface diagram of an embodiment of the method of the present invention. Detailed Implementation

[0063] like Figure 1 The diagram shown is a flowchart of the method of the present invention: The crane wire rope damage identification and prediction method disclosed in this invention includes:

[0064] S1. Sensors are installed along the axial direction of the wire rope to collect the original magnetic field signal of the wire rope to be tested;

[0065] The system comprises multiple sensors. The raw magnetic field signals collected by the multiple sensors are spliced ​​together to form raw multi-channel time-series data. The raw magnetic field signal from each sensor serves as a single-channel time-series data. ;

[0066] Furthermore, the sensor is The first sensor is fixedly installed along the axial direction of the steel wire rope. The location of the second sensor is , No. The location of each sensor is Each sensor continuously acquires the original magnetic field signal according to a preset sampling frequency, i.e., single-channel time-series data. for,

[0067] In the formula, The magnitude of the magnetic field strength. For signal frequency, The phase angle, The sampling time.

[0068] S2. Preprocess the original magnetic field signal to obtain the preprocessed magnetic field signal;

[0069] The preprocessing involves processing each single-channel time-series data... Filtering, denoising, normalization, and sliding window segmentation are performed separately to obtain the preprocessed time-series data for a single channel. , For sensor indexing, For the time of data collection, For the first The position coordinates of each sensor.

[0070] Specifically, a bandpass filter is used for filtering to eliminate power frequency interference, a wavelet threshold denoising method is used for denoising to eliminate environmental noise, and the magnetic field signal amplitude is mapped to the [0,1] interval through normalization to eliminate the influence of dimensions, thereby obtaining the normalized magnetic field signal of a single channel. Then, the normalized magnetic field signal of a single channel is processed by moving window segmentation.

[0071] S3. Perform continuous wavelet transform on the preprocessed magnetic field signal to extract time-frequency features;

[0072] The time-frequency feature is a multi-channel time-frequency feature map, and the extraction steps of the multi-channel time-frequency feature map are as follows:

[0073] From the preprocessed time series data of a single channel The amplitude sequence was extracted. Frequency component sequence and phase sequence ;

[0074] Among them, amplitude sequence Time series data of a single preprocessed channel For the same physical quantity, Frequency component sequence The extraction method involves extracting the time-series data from a single preprocessed channel. Perform a Fast Fourier Transform to decompose the frequency components into amplitude components corresponding to each order. Organize all effective frequency components for each time period according to the time window to form a frequency component time series; phase sequence The extraction method is to use the preprocessed time-series data of a single channel. As input, an analytic signal is constructed using Hilbert transform. The instantaneous phase is then solved by analyzing the imaginary and real parts of the analytic signal. The phase values ​​at each sampling point are output, and the results are concatenated to obtain a complete phase-time sequence. ;

[0075] The amplitude sequence of the extracted single channel Frequency component sequence and phase sequence Continuous wavelet transform (CWT) is performed on each component to extract the corresponding time-frequency features, resulting in amplitude wavelet time-frequency plots. Frequency wavelet time-frequency diagram Phase wavelet time-frequency diagram ;

[0076] The amplitude wavelet time-frequency diagram It is obtained by the following formula.

[0077] In the formula, As a scale, For time, represents the relative time displacement of the local time-frequency characteristics within the window. The generating function of the intelligent extraction engine. To intelligently extract the engine's scale factor, To intelligently extract engine displacement factors,

[0078] The frequency wavelet time-frequency diagram It is obtained by the following formula.

[0079] The phase wavelet time-frequency diagram It is obtained by the following formula. .

[0080] Then, they are concatenated along the feature channel dimension to obtain a single-channel multi-feature time-frequency map. ;

[0081] The single-channel multi-feature time-frequency maps from multiple channels are stitched together to obtain a multi-channel time-frequency feature map. for, In the formula, The number of channels is equal to the number of sensors.

[0082] S4. Input the time-frequency features into the pre-trained deep magnetic field sensing learning model, which processes them and outputs damage identification data and damage prediction data.

[0083] like Figure 2 The diagram shown is a schematic representation of the deep magnetic field sensing learning model of the present invention. The pre-trained deep magnetic field sensing learning model includes: a cascaded feature extraction module, a temporal modeling module, a damage recognition module, and a trend prediction module.

[0084] The feature extraction module is used to process the multi-channel time-frequency feature maps of the input data of the deep magnetic field sensing learning model. The characteristic response intensity is obtained by processing. With time-series input sequence The output will show that 1 and 2 are the time step indices. The time window length,

[0085] The time-series modeling module is used to receive the time-series input sequence output by the feature extraction module. The hidden state feature vector is then processed to obtain the hidden state feature vector. And output, For time step index;

[0086] The damage recognition module is used to receive the feature response intensity output by the feature extraction module. The hidden state feature vector output by the time series modeling module The data is then processed to obtain the damage identification data at time t. And according to damage identification data The calculation method involves calculating the damage identification data from historical time t−T+1 to t−1, and then concatenating these data to obtain the damage identification time series. The damage identification module outputs the damage identification data at time t. With damage identification time series ;

[0087] The trend prediction module is used to receive the damage identification time series output by the damage identification module. The data is then processed to obtain damage prediction data. , Step size;

[0088] The damage identification data With damage prediction data As the output of the deep magnetic field sensing learning model.

[0089] like Figure 3 The diagram shown illustrates the structure of the feature extraction module in the deep magnetic field sensing learning model of the present invention. The feature extraction module includes: a cascaded spatiotemporal feature extraction layer, a spatiotemporal feature mapping layer, and a spatiotemporal feature pooling layer.

[0090] The spatiotemporal feature extraction layer consists of one or more convolutional layers, used to process multi-channel time-frequency feature maps. Perform convolution operations to obtain the convolution output. And output;

[0091] Its function is to extract local spatiotemporal features by performing convolution operations on multi-channel time-frequency feature maps, and then extract spatial local correlations. The spatial local correlation refers to the relationship between the magnetic field features collected by adjacent sensors along the length of the wire rope. Since multiple sensors are arranged in fixed spatial positions within the wire rope detection area, when a wire breaks, wears, or corrodes at a certain location on the wire rope, the time-frequency features corresponding to multiple adjacent sensors usually change simultaneously, thus exhibiting local spatial correlation features.

[0092] The convolutional layers in the spatiotemporal feature extraction layer are calculated according to the following formula.

[0093] In the formula, For the first Layer convolution kernel in the th Weight parameters on each channel, Multi-channel time-frequency feature map The Middle Each channel at time Input value at the location, For bias terms, The output of the convolution operation serves as the basic linear transformation output of the network, forming the main body of subsequent feature mapping.

[0094] The first convolutional layer of the spatiotemporal feature extraction layer is also used to statistically obtain basic quantization indicators based on all time-frequency energy values ​​within the convolutional window, including: local energy mean. Energy variance Peak frequency With pulse duration and output it;

[0095] Specifically,

[0096] Let the convolution window size be M×N, and the set of energy for all pixels within the convolution window be... ,

[0097] Then, the local average energy can be calculated according to the following formula. for,

[0098] In the formula, Represents any single energy element within the window.

[0099] The energy variance is calculated using the following formula. for,

[0100] And, by traversing the window energy set Take the maximum energy value The corresponding frequency scale coordinates are used as the peak frequency. ,in,

[0101] Furthermore, by statistically analyzing whether the energy within the window exceeds a preset damage threshold. The pulse duration is obtained by taking the continuous time-domain sampling length. for, .

[0102] The spatiotemporal feature mapping layer includes a ReLU activation layer and a concatenation layer. The ReLU activation layer is used to receive the convolution operation output from the spatiotemporal feature extraction layer. Then, a nonlinear activation process is performed on it to obtain the spatiotemporal feature map. ;

[0103] The spatiotemporal feature mapping The calculation formula is: ,

[0104] The splicing layer is used to receive the local energy mean output by the spatiotemporal feature extraction layer. Energy variance Peak frequency With pulse duration And map it with spatiotemporal features Channel splicing is performed to obtain three-dimensional spatiotemporal frequency characteristics. The three-dimensional spatiotemporal frequency characteristics The calculation formula is as follows: ;

[0105] The spatiotemporal feature pooling layer comprises a cascaded frequency pooling layer, a spatial pooling layer, and a stitching layer. The frequency pooling layer is composed of global max pooling or global average pooling and is used to perform three-dimensional spatiotemporal frequency features. Frequency aggregation is performed to obtain a two-dimensional spatiotemporal feature vector sequence. The spatial pooling layer consists of global max pooling or global average pooling, and is used to process two-dimensional spatiotemporal feature vector sequences. Spatial aggregation is performed to obtain a one-dimensional feature vector. , For time step index, , The length of the time window;

[0106] The three-dimensional spatiotemporal frequency characteristics Frequency aggregation refers to the aggregation of three-dimensional spatiotemporal frequency features. Frequency aggregation, i.e., compression along frequency channel f, is expressed by the formula:

[0107] In the formula, Here, the function is the frequency aggregation function, and the expression is either global max pooling or global average pooling. This is a spatial channel index, representing the spatial location of the sensor. For time steps, For frequency channel index, This represents the total number of frequency channels.

[0108] The two-dimensional spatiotemporal feature vector sequence Spatial aggregation refers to the process of spatial aggregation of the first... The feature values ​​of all spatial locations at each time step are aggregated, as expressed by the formula:

[0109] In the formula, Let be the spatial feature aggregation function, which is one of global average pooling, global max pooling, or spatial weighted summation. For global average pooling, the spatial aggregation function is expanded as follows:

[0110] when For global max pooling, the spatial aggregation function is expanded as follows:

[0111] when For a space-weighted summation, the space aggregation function can be expanded as follows: .

[0112] The splicing layer is used to... The corresponding one-dimensional feature vector The sequences are concatenated to form a time-series input sequence. ;

[0113] Two-dimensional spatiotemporal feature vector sequence As characteristic response intensity ,Right now Output feature response intensity from spatiotemporal feature pooling layer and timing input sequence .

[0114] The time series modeling module is built based on a first Long Short-Term Memory (LSTM) network and is used to process the time series input sequence. Transform into hidden state feature vectors And output from the time series modeling module;

[0115] The first Long Short-Term Memory (LSTM) network includes cascaded forget gates, input gates, cell state update gates, and output gates. Through the gating mechanism of the forget gate, input gate, cell state update gate, and output gate, it learns the evolutionary dependencies of damage features over time. The standard computation flow of the first LSM network is as follows.

[0116] Input time sequence Input the forget gate of the time series modeling module.

[0117] The forget gate output gating value is calculated using the following formula. for,

[0118] In the formula, For the first The forgetting gate output gating value at each time step, after being mapped by the sigmoid activation function, ranges from (0,1). Its value controls the proportion of cell state retained from the previous time step to the current time step; a value closer to 0 indicates more forgotten historical information, while a value closer to 1 indicates more retained historical information. The sigmoid activation function is used to map the gated output to the (0,1) interval. Here is the weight matrix for the forget gate. for The hidden state vector at each time step, For the first The time-series input sequence elements at each time step, This is an operation that concatenates the hidden state from the previous time step with the input vector from the current time step. Let be the bias vector of the forget gate.

[0119] The input gate value is calculated using the following formula. for,

[0120] In the formula, For the first The input gating value at each time step, after being mapped by the sigmoid activation function, takes a value in the range (0,1). This value is used to control the update ratio of the input information at the current time step written to the cell state. For the input gate weights, For input gate bias,

[0121] The candidate cell state value is then calculated using the following formula. for,

[0122] In the formula, For the first The candidate cell states at each time step are generated using the tanh activation function, with values ​​ranging from -1 to 1, representing the new information to be updated and written into the cell state at the current time step. The hyperbolic tangent activation function is used to map candidate state values ​​to the interval (-1, 1). Let be the candidate cell state weight matrix. This is the candidate cell state bias vector.

[0123] Cell state updates are calculated using the following formula to obtain the current cell state. for,

[0124] In the formula, For element-wise multiplication, For a moment The state of the cells,

[0125] The output gate value is calculated using the following formula. for,

[0126] In the formula, For the first The output gating value at each time step, after being mapped by the sigmoid activation function, takes a value in the range (0,1), and is used to control the proportion of information output from the cell state to the hidden state. This is the output gate weight matrix. This is the output gate bias vector.

[0127] The hidden state is obtained by calculation using the following formula. for,

[0128] In the formula, For time steps The hidden state,

[0129] Take the first Long Short-Term Memory (LSTM) network at the last time step Hidden state output As the final Hidden state feature vector at time step ,Right now ,

[0130] Output from the time series modeling module Hidden state feature vector at time step ,

[0131] The damage identification module includes: a damage category identification submodule, a damage location identification submodule, and a damage severity assessment submodule connected in parallel, and a connection layer; the inputs to the damage category identification submodule and the damage severity assessment submodule are hidden state feature vectors. The input to the damage location identification submodule is the feature response intensity. The outputs of the damage category identification submodule, the damage location identification submodule, and the damage severity assessment submodule are collectively used as the input to the connection layer.

[0132] The damage category identification submodule consists of a fully connected layer and a Softmax layer. First, it hides the state feature vector. The damage category probability vector is obtained by sequentially processing through a fully connected layer and a Softmax layer, and then calculating using the following formula. for,

[0133] In the formula, Weight matrix for identifying damage categories, To identify the bias vector for the damage category, ,in , , Let be the probabilities corresponding to the first, second, and c-th damage categories, respectively, and t be the time interval. This represents the total number of categories of wire rope damage.

[0134] The damage location identification submodule is composed of a cascaded feature response decoupling unit and a spatial weighted centroid calculation unit, used to analyze the feature response intensity. The damage location is estimated by calculating using the following formula. for,

[0135] In the formula, For the number of sensors, Representing the The strength of the magnetic field anomaly at the spatial location of each sensor;

[0136] Specifically, the feature response decoupling unit is used to separate the independent magnetic field feature response intensities of each sensor from the fused spatiotemporal features, and the spatial weighted centroid calculation unit realizes accurate axial damage localization of the wire rope based on the decoupled multi-channel response intensities. The two units work together to calculate the feature response intensities. Process it.

[0137] The damage assessment submodule is built based on a regression network and also includes a normalization submodule for processing the hidden state feature vector. The process is performed to obtain a damage assessment vector. for,

[0138] In the formula, For normalization function, Assess the weighting of the degree of damage. As a bias for assessing the degree of damage, , , , These are the damage severity assessment values ​​corresponding to the first, second, and C damage categories, respectively. , , The range of values ​​for is [0,1]. or or A value of 0 indicates that the wire rope is undamaged in damage category 1, damage category 2, or damage category C. or or A value greater than 0 and less than 0.3 indicates that the wire rope is in a state of minor damage in damage category 1, damage category 2, or damage category C. or or A value greater than 0.3 and less than 0.7 indicates that the wire rope is in a moderate damage state in damage category 1, damage category 2, or damage category C. or or A value greater than 0.7 indicates that the wire rope is severely damaged in damage category 1, 2, or C. or or A value of 1 indicates that the wire rope is in the most severe damage state in damage category 1, damage category 2, or damage category C.

[0139] The connection layer is used to process the damage category probability vector. Damage location estimate Damage severity assessment value By stitching the data together, damage identification data at time t is obtained. Furthermore, the damage identification data from historical time t−T+1 to t−1 are spliced ​​together to obtain the damage identification time series. .

[0140] The weights and bias vectors in the damage recognition module are network parameters obtained through supervised learning iterative optimization of the deep magnetic field sensing learning architecture during the model training phase. The specific acquisition process is as follows:

[0141] 1. Constructing a training dataset: Collect wire rope magnetic flux leakage detection samples with various working conditions such as wire breakage, corrosion, wear, plastic deformation, and health status. Process the samples with an intelligent feature extraction engine to obtain corresponding feature vector samples, and label each sample with the true damage category label.

[0142] 2. Iterative optimization to solve parameters: Taking the cross-entropy loss function as the optimization objective, the labeled feature vector samples are input into the network, and the backpropagation algorithm (with optimizers such as Adam or SGD) is used to iteratively update the weights Wc and biases bc, continuously reducing the error between the model's predicted probability distribution and the true label until the model converges.

[0143] 3. Fixed parameters for inference: After training, the converged Wc and bc are saved as fixed parameters for the damage category identification module; during the inference phase of actual online monitoring, the feature vector h at the current time is directly used. (t) Substitute the values ​​into the formula to calculate the probability results for each damage category, and you can get the output.

[0144] The trend prediction module includes a cascaded second Long Short-Term Memory (LSTM) network and a fully connected linear layer. The processing of the second LSTM network is the same as that of the first LSTM network. The second LSTM network iterates through input gates, forget gates, and output gates, updating the cell state and latent state time-by-time, automatically retaining long-term damage evolution patterns, and serializing the damage identification time series. The input to the second Long Short-Term Memory network is processed to obtain the hidden state feature vector of the future damage prediction module. The hidden state feature vector The fully connected linear layer is input and processed using the following formula to obtain the damage prediction result. for,

[0145] In the formula, Indicates the future The prediction results for the time period, For the corresponding time step The weight parameters, For bias terms, For the future The probability vector of the damage category at time step, , , , In the future The probabilities of the first, second, and c-th damage types at time point c. Step size, For the future Predicted damage location at time [time]. For the future The damage prediction vector at time step. , , , For the future The first, second, and third time Predicted damage severity values ​​corresponding to different damage categories. , , It is a normalized continuous scalar with a value range of [0,1]. or or When the value is 0, it represents the future. The wire rope is in an undamaged state at any given time under damage category 1, damage category 2, or damage category C. or or A value greater than 0 and less than 0.3 represents the future. The wire rope is in a state of minor damage under damage category 1, 2, or C. or or A value greater than 0.3 and less than 0.7 represents the future. The wire rope is in a severely damaged state at any given time, whether it falls under Damage Category 1, Damage Category 2, or Damage Category C. or or When the value is 1, it represents the future. The wire rope is in the most severe damage state in either the first damage category, the second damage category, or the C damage category.

[0146] The weights and bias vectors in the trend prediction module are network parameters obtained through supervised learning iterative optimization of the deep magnetic field sensing learning architecture during the model training phase. The specific acquisition process is as follows:

[0147] 1. Construct a training dataset: Collect historical operational data of the wire rope throughout its entire lifecycle to form a sequence of "damage identification results at T consecutive time points" corresponding to "future..." The labeled sample set of "real damage state labels at any time" is used as the training set for the prediction model.

[0148] 2. Set optimization objective: Use the error between the model's predicted output and the actual damage label (such as mean squared error, MSE) as the loss function, and minimize the prediction error as the training objective.

[0149] 3. Iteratively update weight parameters: Using the backpropagation algorithm, combined with gradient descent optimizers such as Adam and SGD, iteratively update the weights corresponding to each time step. With bias The parameters are continuously adjusted to reduce the prediction bias until the model loss converges.

[0150] In practical implementation, the second Long Short-Term Memory (LSTM) network can be replaced with other time-series prediction networks depending on the computing power, data scale, and prediction accuracy requirements: When facing situations with low computing power, limited sample size, and basic prediction accuracy requirements, GRU gated recurrent units or single-layer RNNs are selected. These networks have fewer parameters, faster training speed, and relatively low hardware computing power requirements, making them suitable for on-site edge embedded devices; When facing situations with moderate computing power, data scale, and conventional prediction accuracy, TCN temporal convolutional networks or Bi-GRUs are selected. These networks can balance temporal dependence with computational speed; When facing situations with high computing power requirements, massive amounts of data, and high prediction accuracy requirements, Transformer temporal models or Bi-LSTM+Attention networks are selected. These networks can finely distinguish the evolutionary patterns of multiple loss couplings to achieve fine identification and accurate prediction of wire rope damage, making them suitable for long-term safety inspection scenarios of large lifting equipment.

[0151] The process of obtaining the damage identification and prediction results for the wire rope based on damage identification and prediction data includes: selecting the damage category with the highest probability from the damage category probability vector as the identified damage category; and combining the estimated damage degree and damage location corresponding to this identified damage category to form the damage identification result for the wire rope. Specifically, let the probability of the highest damage category be... Then, the first type of damage is determined as the damage category, combined with the estimated damage location. Damage severity assessment value corresponding to the first type of damage The damage identification results of the wire rope are expressed as follows: at the location The probability of the first type of injury occurring at the location is The degree of damage is .

[0152] In the future The damage category with the highest probability in the damage category probability vector at any given time is selected as the predicted damage category. This predicted damage category is then combined with the predicted damage severity and estimated damage location to form the overall damage prediction result for the wire rope. Specifically, let the future... The probability of the highest damage category at time is Then, the first type of damage is determined as the predicted damage type, combined with the predicted damage location value. Predicted damage severity value corresponding to the first damage category The damage prediction results for the wire rope are expressed as follows: future Always in position The probability of the first type of injury occurring at the location is The degree of damage is .

[0153] S5. Based on the damage identification data and damage prediction data, obtain the damage identification results and damage prediction results for the wire rope, including:

[0154] First, the preset damage level failure threshold vector is: , , , These are the failure thresholds for damage categories 1, 2, and C, respectively.

[0155] The preset damage level failure threshold vector is based on the wire rope industry operation and maintenance standards and on-site scrapping specifications - "Maintenance, Upkeep, Inspection and Scrapping of Crane Wire Ropes" (…). https: / / img.antpedia.com / standard / files / pdfs_ ora / 20240720 / GB_T%205972-2023.pdf )set up.

[0156] Then, the future damage prediction module calculates the future damage. to Damage severity prediction sequence at time step This is combined with a preset damage severity failure threshold vector to calculate the remaining life of the wire rope under each individual damage category. for,

[0157] In the formula, Representative at the The remaining life of a wire rope under a single damage category. , The time interval corresponding to the k-th prediction step size. For the future Damage prediction vector at time step The Middle Predicted damage severity values ​​for each damage category For the first Failure threshold for each damage category Indicates the predicted value of damage level The first time the damage level or failure threshold is reached or exceeded The corresponding time interval, the remaining life of the wire rope after a single damage. For only the first The estimated time from the current moment until the wire rope reaches the scrap condition under each type of damage;

[0158] Finally, the remaining life of the wire rope under each single type of damage. The minimum value in the range is taken as the comprehensive remaining life of the wire rope, and is expressed as: .

[0159] S6. Visualize the results of wire rope damage identification, damage prediction, and remaining life assessment, enabling operators to easily monitor various performance indicators of the wire rope and react quickly.

[0160] The method of the present invention can perform high-sensitivity non-destructive testing on steel wire ropes in real time, identify minute physical damage to the steel wire ropes, such as corrosion, wear, breakage, and local plastic deformation, and can predict the damage state of the steel wire ropes at a future time based on the historical evolution data of the identified minute damages.

[0161] The effects of the method of the present invention will be illustrated below with reference to an embodiment:

[0162] Taking a single wire rope of a typical crane as the monitoring object, this paper elaborates on the specific implementation process of a crane wire rope damage identification and prediction method.

[0163] Step 1: Deploy the sensor array and collect the raw magnetic field signal

[0164] like Figure 4 The diagram shows a sensor array layout consisting of multiple sensor units, representing an embodiment of the method of the present invention. First, in the critical load-bearing section of the crane wire rope, a sensor array consisting of N magnetic field sensors is arranged at equal intervals along the wire rope axis. In this embodiment, N=5. Starting from the initial end face of the wire rope, the five sensors are arranged at equal intervals of 1m along the x-axis. The initial end face of the critical load-bearing section of the wire rope is taken as the origin of a one-dimensional coordinate system (x=0); the direction extending along the wire rope axis away from the initial end face is taken as the positive x-axis direction. Therefore, the position coordinates of the first sensor are... The position coordinates of the second sensor are The position coordinates of the third sensor are The position coordinates of the 4th sensor are The position coordinates of the 5th sensor are Each sensor unit synchronously acquires the magnetic field signal at its location at a sampling frequency of 1kHz, forming raw multi-channel timing data. Each sensor signal corresponds to a single-channel timing data. The key load-bearing sections of the wire rope include: the hoisting rope section extending from the end of the boom, as well as the drum-pulley transition section and pulley winding section, which also fall into this category. The starting end refers to the end of the wire rope extending from the end of the boom.

[0165] Each sensor unit acquires the magnetic field signal at its location in real time. For the i-th sensor, the acquired magnetic field signal S... i (t) can be expressed as:

[0166] Among them, A i (t) represents the magnetic field strength amplitude ω i For signal frequency, Փ i The phase angle is used. The system synchronously acquires these 5 signals at a sampling frequency of 1kHz to obtain the raw multi-channel timing data. This data not only contains leakage magnetic information reflecting local defects in the wire rope (such as broken wires, corrosion, wear, and plastic deformation), but also stress-induced magnetization change information reflecting its stress state (tension, compression, and bending).

[0167] Table 1. Examples of typical data collected by each sensor

[0168]

[0169] These high-dimensional data, including spatial location, magnetic field fluctuation intensity, frequency response characteristics, and their temporal variation trends, form the input basis for subsequent analysis.

[0170] Step 2: Signal preprocessing and multi-channel time-frequency feature extraction

[0171] The raw signals collected by all sensors are preprocessed and time-frequency features are constructed sequentially. The specific process is as follows:

[0172] Preprocessing: For each signal S i(t) Filtering, denoising, normalization, and sliding window segmentation are performed sequentially. Specifically, a bandpass filter is used for filtering to eliminate power frequency interference; wavelet threshold denoising is used to eliminate environmental noise; normalization is performed to map the magnetic field signal amplitude to the [0,1] interval to eliminate the influence of dimensions, resulting in the normalized single-channel magnetic field signal; then, the normalized single-channel magnetic field signal is segmented by sliding window, with a window length of 512 sampling points and a sliding step size of 256 sampling points, dividing the long-time sequence signal into multiple short-time analysis segments.

[0173] Step 3: Extract the amplitude sequence, frequency component sequence, and phase sequence from the preprocessed single-channel time-series data:

[0174] The amplitude sequence and the preprocessed time series data are the same physical quantity;

[0175] The frequency component sequence is obtained by performing a fast Fourier transform on the signal within the window to decompose the amplitude components corresponding to each frequency order, and then organizing all effective frequency components in each time period according to the time window.

[0176] The phase sequence is constructed by Hilbert transform to obtain an analytic signal. The instantaneous phase is solved by the imaginary and real parts of the analytic signal. The phase value at each sampling point is output and then spliced ​​together to obtain a complete phase time sequence.

[0177] Continuous Wavelet Transform and Feature Stitching: Continuous wavelet transform (CWT) is performed on the amplitude, frequency, and phase sequences respectively, using the Morlet mother wavelet to generate amplitude, frequency, and phase wavelet time-frequency maps. First, the three time-frequency maps from a single sensor are stitched along the feature channel dimension to obtain a single-channel multi-feature time-frequency map. Then, the single-channel time-frequency maps from the five sensors are stitched together to form a dimension... The multi-channel time-frequency feature map is generated, where H is the length of the frequency axis and W is the length of the time axis. This multi-channel time-frequency feature map serves as the input to the subsequent deep magnetic field sensing learning model.

[0178] Step 4: Calculation of the deep magnetic field sensing learning model

[0179] The multi-channel time-frequency feature map is input into the pre-trained deep magnetic field sensing learning model, and the calculation is completed sequentially through the feature extraction module, the time series modeling module, the damage recognition module, and the trend prediction module.

[0180] 1) Feature extraction module operation

[0181] The feature extraction module processes features sequentially through a spatiotemporal feature extraction layer, a spatiotemporal feature mapping layer, and a spatiotemporal feature pooling layer:

[0182] Spatiotemporal feature extraction layer:

[0183] It consists of multiple cascaded convolutional layers, performing convolution operations on multi-channel time-frequency feature maps to extract local spatiotemporal correlation features. For the first... The operation formula for a convolutional layer is:

[0184] The first convolutional layer simultaneously calculates four basic quantization metrics: assuming the convolutional window size is 3x3, the set of pixel energy within the window is... The mean and variance of local energy are calculated respectively; the frequency scale corresponding to the maximum energy value is taken by traversing the energy set to obtain the peak frequency; the continuous time domain sampling length of the energy within the statistical window that is greater than the damage judgment threshold is used to obtain the pulse duration.

[0185] Spatiotemporal feature mapping layer:

[0186] The convolutional output is activated by the ReLU function to obtain a spatiotemporal feature map; the spatiotemporal feature map is then concatenated with the above four quantization indicators along the channel dimension through a concatenation layer to obtain three-dimensional spatiotemporal frequency features.

[0187] Spatiotemporal feature pooling layer:

[0188] First, global average pooling is used to aggregate frequencies, compressing the frequency dimension to obtain a two-dimensional spatiotemporal feature vector sequence (i.e., feature response intensity). Then, global average pooling is used to aggregate spatial features, fusing the spatial features of all sensors and outputting a one-dimensional feature vector corresponding to each time step. Finally, the features of each time step are concatenated to form a time-series input sequence of length T.

[0189] (2) Timing modeling module operation

[0190] The temporal input sequence is fed into the first Long Short-Term Memory (LSTM) network. The network sequentially updates the cell state and hidden state through gated iterative operations of forget gate, input gate, cell state update, and output gate, and automatically learns the evolutionary dependency of damage features in the time dimension. The hidden state output of the last time step is taken as the hidden state feature vector h(t) at the current time step.

[0191] (3) Damage recognition module calculation

[0192] The damage identification module consists of three parallel sub-modules: damage type identification, damage location identification, and damage severity assessment, as well as a connection layer.

[0193] Damage type identification: The damage type identification submodule consists of a fully connected layer and a Softmax layer. The hidden state feature vector h(t) is substituted into the formula. The probability vectors for four types of damage are calculated. The output results in this embodiment are: corrosion probability 32%, wear probability 18%, wire breakage probability 5%, and plastic deformation probability 8%.

[0194] Damage location identification submodule: Composed of a cascaded feature response decoupling unit and a spatial weighted centroid calculation unit, it calculates the axial position of the damage using the feature response intensity as input and the weighted centroid formula.

[0195] In the formula Let x be the magnetic field anomaly response intensity corresponding to the i-th sensor. In this embodiment, the estimated damage location is calculated to be x = 2.1m.

[0196] Damage assessment submodule: Based on a regression network, it outputs a normalized damage assessment vector corresponding to the four types of damage, with values ​​ranging from [0,1]. In this embodiment, the corrosion damage level is 0.2, the wear damage level is 0.11, the wire breakage damage level is 0.04, and the plastic deformation damage level is 0.06.

[0197] Finally, the connection layer concatenates the damage type probability, damage location, and damage degree to obtain the damage identification data at the current moment; it further concatenates the damage identification data from the past T moments to form a damage identification time series.

[0198] Prediction module operation

[0199] The time series of damage identification results is fed into the second Long Short-Term Memory (LSTM) network in the trend prediction module. The network mines the long-term evolution patterns of the damage and outputs the predicted hidden state. Then, it is mapped through a fully connected linear layer to obtain the future step size. The corresponding damage prediction data. This embodiment calculates the future damage prediction data. The probability of corrosion at any given time is 41%, the probability of wear is 22%, the probability of wire breakage is 14%, and the probability of plastic deformation is 23%. The predicted damage location is x=2.26m, the predicted damage degree of corrosion is 0.28, the predicted damage degree of wear is 0.16, the predicted damage degree of wire breakage is 0.05, and the predicted damage degree of plastic deformation is 0.08.

[0200] Step 5: Damage Identification Results and Damage Prediction Results

[0201] Based on the damage identification data at the current moment in this embodiment, the damage identification result is that the probability of rust damage at the location coordinate x=2.1m is 32%, and the degree of rust damage is a minor damage state.

[0202] According to the future step size in this embodiment Based on the corresponding damage prediction data, the damage prediction result for the wire rope is as follows: Future The probability of rust damage occurring at a location of 2.26m is 41%, and the degree of damage is minor.

[0203] Step Six: Remaining Life Assessment

[0204] According to the "Maintenance, Upkeep, Inspection and Scrapping of Crane Wire Ropes" (GB / T5972-2023), the failure threshold for damage level is set as follows: By extrapolating the evolution trend of damage severity, the time interval corresponding to the first time the damage severity reaches the failure threshold is taken as the remaining life of a single type of damage. Considering the coupled effects of the four types of damage, the shortest life is taken as the comprehensive remaining life. In this embodiment, it is calculated that the remaining life of corrosion damage at this location is approximately 128 days, the remaining life of wear damage is approximately 112 days, and the final comprehensive remaining life of the wire rope is 112 days.

[0205] Step 7: Visual Presentation

[0206] The real-time damage identification results, damage prediction results, early warning conclusions and remaining life assessment results of wire ropes are integrated into a unified visualization interface, which displays the real-time health status, damage evolution trend curve, risk warning prompts and maintenance suggestions, supporting maintenance personnel to make rapid decisions.

[0207] like Figure 5 The diagram shown is a visual interface illustration of an embodiment of the method of the present invention. Operators can intuitively monitor various performance indicators of the wire rope through a remote monitoring platform. The interface uses a schematic wire rope model or dashboard format, with colors (green, yellow, red) to indicate the risk level of different sections, facilitating rapid response from operators to take preventative or remedial measures, thereby improving the safety of crane operation and the scientific nature of maintenance.

Claims

1. A method for identifying and predicting damage to crane wire ropes, characterized in that, include: Sensors are deployed along the axial direction of the wire rope to collect the original magnetic field signal of the wire rope under test; The original magnetic field signal is preprocessed to obtain the preprocessed magnetic field signal; The preprocessed magnetic field signal is subjected to continuous wavelet transform to extract time-frequency features; the time-frequency features are input into a pre-trained deep magnetic field perception learning model, which processes them and outputs damage identification data and damage prediction data; based on the damage identification data and damage prediction data, the damage identification results and damage prediction results of the wire rope are obtained.

2. The method according to claim 1, characterized in that, The system comprises multiple sensors. The raw magnetic field signals collected by the multiple sensors are spliced ​​together to form raw multi-channel time-series data. The raw magnetic field signal from each sensor serves as a single-channel time-series data. The preprocessing involves processing each single-channel timing data... Filtering, denoising, normalization, and sliding window segmentation are performed separately to obtain the preprocessed time-series data for a single channel. , For sensor indexing, For the time of data collection, For the first The position coordinates of each sensor.

3. The method according to claim 2, characterized in that, The time-frequency feature is a multi-channel time-frequency feature map, and the extraction steps of the multi-channel time-frequency feature map are as follows: From the preprocessed time series data of a single channel The amplitude sequence was extracted. Frequency component sequence and phase sequence ; The amplitude sequence of the extracted single channel Frequency component sequence and phase sequence Continuous wavelet transform (CWT) is performed on each component to extract the corresponding time-frequency features, resulting in amplitude wavelet time-frequency plots. Frequency wavelet time-frequency diagram Phase wavelet time-frequency diagram ; Then, they are concatenated along the feature channel dimension to obtain a single-channel multi-feature time-frequency map. ; The single-channel multi-feature time-frequency maps from multiple channels are stitched together to obtain a multi-channel time-frequency feature map. for, In the formula, The number of channels is equal to the number of sensors.

4. The method according to claim 3, characterized in that, The amplitude wavelet time-frequency diagram It is obtained by the following formula. In the formula, As a scale, For time, represents the relative time displacement of the local time-frequency characteristics within the window. The generating function of the intelligent extraction engine. To intelligently extract the engine's scale factor, To intelligently extract engine displacement factors, The frequency wavelet time-frequency diagram It is obtained by the following formula. The phase wavelet time-frequency diagram It is obtained by the following formula. 。 5. The method according to claim 4, characterized in that, The pre-trained deep magnetic field sensing learning model includes: a cascaded feature extraction module, a temporal modeling module, a damage identification module, and a trend prediction module; The feature extraction module is used to process the multi-channel time-frequency feature maps of the input data of the deep magnetic field sensing learning model. The characteristic response intensity is obtained by processing. With time-series input sequence The output will show that 1 and 2 are the time step indices. The time window length, The time-series modeling module is used to receive the time-series input sequence output by the feature extraction module. The hidden state feature vector is then processed to obtain the hidden state feature vector. And output, For time step index; The damage recognition module is used to receive the feature response intensity output by the feature extraction module. The hidden state feature vector output by the time series modeling module The data is then processed to obtain the damage identification data at time t. And according to damage identification data The calculation method involves calculating the damage identification data from historical time t−T+1 to t−1, and then concatenating these data to obtain the damage identification time series. The damage identification module outputs the damage identification data at time t. With damage identification time series ; The trend prediction module is used to receive the damage identification time series output by the damage identification module. The data is then processed to obtain damage prediction data. , Step size; The damage identification data With damage prediction data As the output of the deep magnetic field sensing learning model.

6. The method according to claim 5, characterized in that, The feature extraction module includes: a cascaded spatiotemporal feature extraction layer, a spatiotemporal feature mapping layer, and a spatiotemporal feature pooling layer; The spatiotemporal feature extraction layer consists of one or more convolutional layers, used to process multi-channel time-frequency feature maps. Perform convolution operations to obtain the convolution output. And output; The first convolutional layer of the spatiotemporal feature extraction layer is also used to statistically obtain basic quantization indicators based on all time-frequency energy values ​​within the convolutional window, including: local energy mean. Energy variance Peak frequency With pulse duration and output it; The spatiotemporal feature mapping layer includes a ReLU activation layer and a concatenation layer. The ReLU activation layer is used to receive the convolution operation output from the spatiotemporal feature extraction layer. Then, a nonlinear activation process is performed on it to obtain the spatiotemporal feature map. ; The splicing layer is used to receive the local energy mean output by the spatiotemporal feature extraction layer. Energy variance Peak frequency With pulse duration And map it with spatiotemporal features Channel splicing is performed to obtain three-dimensional spatiotemporal frequency characteristics. ; The spatiotemporal feature pooling layer comprises a cascaded frequency pooling layer, a spatial pooling layer, and a stitching layer. The frequency pooling layer is composed of global max pooling or global average pooling and is used to perform three-dimensional spatiotemporal frequency features. Frequency aggregation is performed to obtain a two-dimensional spatiotemporal feature vector sequence. The spatial pooling layer consists of global max pooling or global average pooling, and is used to process two-dimensional spatiotemporal feature vector sequences. Spatial aggregation is performed to obtain a one-dimensional feature vector. , For time step index, , The length of the time window; The splicing layer is used to... The corresponding one-dimensional feature vector The sequences are concatenated to form a time-series input sequence. ; Two-dimensional spatiotemporal feature vector sequence As characteristic response intensity ,Right now Output feature response intensity from spatiotemporal feature pooling layer and timing input sequence .

7. The method according to claim 6, characterized in that, The convolutional layers in the spatiotemporal feature extraction layer are calculated according to the following formula. In the formula, For the first Layer convolution kernel in the th Weight parameters on each channel, Multi-channel time-frequency feature map The Middle Each channel at time Input value at the location, For bias terms, The output is the result of the convolution operation. The first convolutional layer of the spatiotemporal feature extraction layer is also used to statistically obtain basic quantization indicators based on all time-frequency energy values ​​within the convolutional window. Let the convolution window size be M×N, and the set of energy for all pixels within the convolution window be... , Then, the local average energy can be calculated according to the following formula. for, In the formula, Represents any single energy element within the window. The energy variance is calculated using the following formula. for, And, by traversing the window energy set Take the maximum energy value The corresponding frequency scale coordinates are used as the peak frequency. ,in, Furthermore, by statistically analyzing whether the energy within the window exceeds a preset damage threshold. The pulse duration is obtained by taking the continuous time-domain sampling length. for, 。 8. The method according to claim 7, characterized in that, The time series modeling module is built based on a first Long Short-Term Memory (LSTM) network and is used to process the time series input sequence. Transform into hidden state feature vectors And output from the time series modeling module; The damage identification module includes: a damage category identification submodule, a damage location identification submodule, and a damage severity assessment submodule connected in parallel, as well as a connection layer; The input to the damage category identification submodule and the damage severity assessment submodule is the hidden state feature vector. The input to the damage location identification submodule is the feature response intensity. The outputs of the damage category identification submodule, the damage location identification submodule, and the damage severity assessment submodule are collectively used as the input to the connection layer. The damage category identification submodule consists of a fully connected layer and a Softmax layer. First, it hides the state feature vector. The damage category probability vector is obtained by sequentially processing through a fully connected layer and a Softmax layer, and then calculating using the following formula. for, In the formula, Weight matrix for identifying damage categories, To identify the bias vector for the damage category, ,in , , They are respectively the first type, the second type, and the third type. The probability corresponding to each damage type, where t is time t. This represents the total number of categories of wire rope damage. The damage location identification submodule is composed of a cascaded feature response decoupling unit and a spatial weighted centroid calculation unit, used to analyze the feature response intensity. The damage location is estimated by calculating using the following formula. for, In the formula, For the number of sensors, Representing the The strength of the magnetic field anomaly at the spatial location of each sensor; The damage assessment submodule is built based on a regression network and also includes a normalization submodule for processing the hidden state feature vector. The process is performed to obtain a damage assessment vector. for, In the formula, For normalization function, Assess the weighting of the degree of damage. As a bias for assessing the degree of damage, , , , These are the damage severity assessment values ​​corresponding to the first, second, and C damage categories, respectively. The connection layer is used to process the damage category probability vector. Damage location estimate Damage severity assessment value By stitching the data together, damage identification data at time t is obtained. Furthermore, the damage identification data from historical time t−T+1 to t−1 are spliced ​​together to obtain the damage identification time series. .

9. The method according to claim 8, characterized in that, The trend prediction module includes a cascaded second Long Short-Term Memory (LSTM) network and a fully connected linear layer, which will sequence the damage identification time series. The input to the second Long Short-Term Memory network is processed to obtain the hidden state feature vector of the future damage prediction module. The hidden state feature vector The fully connected linear layer is input and processed using the following formula to obtain the damage prediction result. for, In the formula, Indicates the future The prediction results at that time For the corresponding time step The weight parameters, For bias terms, For the future The probability vector of the damage category at time step, , , , In the future The probabilities of the first, second, and c-th damage types at time point c. Step size, For the future Predicted damage location at time [time]. For the future The damage prediction vector at time step. , , , For the future The first, second, and third time Predicted damage severity values ​​corresponding to different damage categories. The step of obtaining the damage identification and prediction results for the wire rope based on damage identification and prediction data includes: selecting the damage category with the highest probability from the damage category probability vector as the identified damage category, and combining the estimated damage degree and damage location corresponding to this identified damage category to form the damage identification result for the wire rope. In the future The damage category with the highest probability is selected from the damage category probability vector at any given time as the predicted damage category. The predicted damage level and the estimated damage location corresponding to this predicted damage category are then combined to form the damage prediction result for the wire rope.

10. The method according to claim 9, characterized in that, The method further includes: deriving a remaining life assessment based on the damage prediction results, including: First, the preset damage level failure threshold vector is: , , , These are the failure thresholds for damage categories 1, 2, and C, respectively. Then, the future damage prediction module calculates the future damage. to Damage severity prediction sequence at time step This is combined with a preset damage severity failure threshold vector to calculate the remaining life of the wire rope under each individual damage category. for, In the formula, Representative at the The remaining life of a wire rope under a single damage category. , The time interval corresponding to the k-th prediction step size. For the future Damage prediction vector at time step The Middle Predicted damage severity values ​​for each damage category For the first Failure threshold for each damage category Indicates the predicted value of damage level The first time the damage level or failure threshold is reached or exceeded The corresponding time interval, the remaining life of the wire rope after a single damage. For only the first The estimated time from the current moment until the wire rope reaches the scrap condition under each type of damage; Finally, the single-damage remaining life of the wire rope under each single-category damage condition. The minimum value in the range is taken as the comprehensive remaining life of the wire rope, and is expressed as: .

Images