A steel wire fracture risk comprehensive evaluation method based on reinforcement learning

Abstract

The application discloses a kind of steel wire fracture risk comprehensive evaluation methods based on reinforcement learning, comprising the following steps: S1, original monitoring data is collected, and standard input sequence is generated;S2, input to improved DPRNN model, construct state vector, extract risk confidence value and confidence change rate;S3, build double-confidence channel, main confidence channel records risk confidence value, and auxiliary confidence channel records confidence change rate;S4, set short-time sliding window, and construct short-time memory pool;S5, execute volatility assessment and trend determination operation;S6, execute confidence compensation mechanism;S7, iteratively update state vector, and execute confidence penalty mechanism;S8, activate the action output path of main confidence channel, and generate fracture risk prediction result.The application can realize sensitive identification and dynamic evaluation of steel wire fracture risk under the condition that confidence value fluctuates sharply and trend changes abnormally, improve the accuracy and response timeliness of fracture early warning.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence risk assessment technology, and in particular to a comprehensive assessment method for steel wire breakage risk based on reinforcement learning. Background Technology

[0002] With the development of intelligent monitoring and structural health assessment technologies, data-driven methods for predicting steel wire fracture risks have gradually gained attention in the industrial sector, especially in critical structures such as lifting equipment, bridge cables, and elevator wire ropes. Accurate identification of their fatigue state and potential fracture risks has become a core aspect of safe operation and maintenance. Existing steel wire fracture risk assessment methods often employ stress thresholds, life models, or finite state machines to construct early warning mechanisms. However, under complex operating conditions, these methods often suffer from the following problems: the fatigue stage signal fluctuations during fracture are weak, making it difficult for conventional methods to effectively perceive their changing trends, resulting in poor early risk identification capabilities; while reinforcement learning methods possess adaptive policy optimization capabilities, when faced with weak fatigue signals, the policy output is easily affected by unstable reward signals and ambiguous state judgments, leading to frequent oscillations in assessment results and high false alarm and false negative rates; in the continuous evolution interval from fatigue to fracture, the policy confidence tends to slowly shift over time. Without an effective suppression mechanism, the model will gradually deviate from the actual risk evolution path, resulting in policy drift, which affects the reliability and consistency of fracture prediction. Existing methods lack confidence dynamic adjustment and drift compensation mechanisms for such fatigue ambiguity states, making it difficult to continuously identify and stably judge early risk signals, thus limiting their practicality and deployment capabilities in high-security application scenarios.

[0003] Therefore, how to provide a comprehensive assessment method for steel wire fracture risk based on reinforcement learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a comprehensive assessment method for steel wire fracture risk based on reinforcement learning. This invention fully integrates vibration signals, strain signals, and acoustic emission signals during the operation of the steel wire to construct a standard input sequence. It extracts risk confidence values ​​and confidence change rates through an improved DPRNN model, further constructs a main confidence channel and an auxiliary confidence channel, and combines a short-time memory pool structure to achieve fatigue fuzzy stage identification and dynamic response to fracture risk. Based on the identification of risk trends, it introduces a confidence compensation mechanism and a confidence penalty mechanism, and triggers fracture risk prediction operations through abnormal mutation signals. The invention details the steel wire fracture risk assessment process for continuous monitoring scenarios, which has the advantages of high response sensitivity, high prediction accuracy, and strong anti-fluctuation capability.

[0005] A comprehensive assessment method for steel wire fracture risk based on reinforcement learning according to an embodiment of the present invention includes the following steps:

[0006] S1. Collect raw monitoring data during the operation of the steel wire, normalize and format it according to a unified time step, and generate a standard input sequence;

[0007] S2. Input the standard input sequence into the improved DPRNN model to construct the state vector corresponding to the time step, and extract the risk confidence value and confidence change rate from the state vector;

[0008] S3. Construct a dual confidence channel based on the risk confidence value and confidence change rate. The main confidence channel records the risk confidence value at each time step, and the auxiliary confidence channel records the confidence change rate of the main confidence channel at consecutive time steps.

[0009] S4. Set a short-time sliding window and construct a short-time memory pool; the short-time memory pool is used to cache the confidence values ​​output by the main confidence channel in a set number of consecutive time steps;

[0010] S5. Perform volatility assessment and trend determination operations on the short-term memory pool; if the volatility of the main confidence channel is less than the first preset threshold and shows a continuous upward trend, it is determined to be in the fatigue fuzzy stage.

[0011] S6. Execute a confidence compensation mechanism in the fatigue fuzzy stage; the confidence compensation mechanism includes adjusting the reward function of the time step, introducing a trend adjustment term into the reward function, stopping the real-time update operation of the improved DPRNN model, and performing alignment correction between the confidence value of the main confidence channel and the mean confidence value in the short-term memory pool.

[0012] S7. Continue to collect raw monitoring data and iteratively update the state vector. If the confidence change rate in the auxiliary confidence channel is greater than the second preset threshold within a set number of consecutive time steps, and no breakage risk event is detected, then execute the confidence penalty mechanism. The confidence penalty mechanism includes introducing a negative trend term into the reward function, performing attenuation processing on the confidence value of the main confidence channel, and performing weighted attenuation operation on the historical confidence values ​​in the short-term memory pool.

[0013] S8. If an abnormal mutation signal is detected within a short sliding window, the execution of the confidence penalty mechanism is stopped, the action output path of the main confidence channel is activated, and the fracture risk prediction result is generated.

[0014] Preferably, S1 specifically comprises:

[0015] Raw monitoring data of the steel wire during operation is collected; the raw monitoring data includes vibration signals, strain signals, and acoustic emission signals; the vibration signals are obtained by acquiring vibration amplitude sequences through accelerometers installed on the surface of the steel wire; the strain signals are obtained by acquiring strain variable sequences through strain gauges deployed at key nodes of the steel wire; the acoustic emission signals are obtained by acquiring acoustic emission amplitude and frequency change sequences through a distributed acoustic emission sensor array.

[0016] Missing data completion and anomaly removal are performed on vibration signals, strain signals and acoustic emission signals respectively to construct a data channel structure with a unified format.

[0017] Set a uniform time step length and alignment reference time scale, perform time synchronization processing and interpolation alignment operations; perform normalization operations on the time-synchronized signal data, normalize the vibration signal and strain signal according to the minimum and maximum ranges, normalize the acoustic emission signal according to the logarithmic amplitude, and construct a standard input sequence arranged in time order.

[0018] Preferably, S2 specifically includes:

[0019] The standard input sequence is divided into fixed-length local signal blocks to construct an input structure that includes local time intervals and overall time intervals;

[0020] The input structure is input into the improved DPRNN model, which includes a temporal dual-channel structure, a position-gated structure, a confidence generation structure, and a residual memory structure.

[0021] The temporal features within local blocks and the temporal correlation features between blocks are extracted by a temporal dual-channel structure, and the dual-channel fused feature representation is output.

[0022] The position vector, composed of the dual-channel fused feature representation and the time step index, is input into the position gating structure, and an attention weighting operation is performed to generate a weighted fusion result.

[0023] The weighted fusion result is input into the confidence generation structure. The first output path outputs the risk confidence value, and the second output path outputs the confidence change rate.

[0024] A residual memory structure is introduced into the confidence generation structure. It receives the fused feature representation of the previous time step and the fused feature representation of the current time step, performs residual connection, and then participates in the confidence output together. Finally, a state vector containing risk confidence value and confidence rate of change is constructed.

[0025] Preferably, the improved DPRNN model includes a temporal dual-channel structure, a position-gated structure, a confidence generation structure, and a residual memory structure, specifically:

[0026] The timing dual-channel structure includes multiple bidirectional gated loop units, which receive the local signal block sequence and the global signal block sequence after being divided in the standard input sequence, respectively extract the timing features within the local block and the timing correlation features between cross blocks, and generate a dual-channel fused feature representation.

[0027] The position gating structure includes a time position encoding module and a gating attention unit. The time position encoding module generates a corresponding position vector based on the time step index. The gating attention unit receives the dual-channel fused feature representation and the position vector, constructs an attention weight distribution, performs weighted adjustment on the features of each time step, and generates a fused feature result.

[0028] The confidence generation structure includes a first output path and a second output path. The first output path includes at least one fully connected network structure and outputs a risk confidence value. The second output path includes at least one fully connected network structure and outputs a confidence change rate.

[0029] The residual memory structure constructs a historical state information pathway, performs a residual connection between the fusion feature representation of the previous time step and the fusion feature representation of the current time step, and inputs the connection result into the confidence generation structure to generate a state vector containing historical evolution information.

[0030] Preferably, S3 specifically includes:

[0031] After extracting the risk confidence value and confidence change rate corresponding to each time step, a dual confidence channel structure is constructed, which includes a primary confidence channel and a secondary confidence channel.

[0032] The primary confidence channel receives the risk confidence values ​​in the time step state vector and records the risk confidence values ​​of each time step in chronological order as the primary confidence sequence.

[0033] The secondary confidence channel calculates the rate of change between adjacent time steps based on the continuous time step confidence value sequence of the primary confidence channel. The rate of change is obtained by difference and forms a confidence rate of change sequence.

[0034] At each time step, the main confidence channel outputs the risk confidence value corresponding to that time step, and the auxiliary confidence channel outputs the rate of change of the confidence value between that time step and the previous time step. The outputs of the two confidence channels together form the confidence expression matrix.

[0035] Preferably, S4 specifically comprises:

[0036] Set the length parameter of the short-time sliding window, construct the short-time sliding window according to the set number of continuous time steps, and define the range of the short-time confidence cache.

[0037] A short-time memory pool structure is constructed, which is a fixed-length cache module that sequentially stores the risk confidence values ​​output by the main confidence channel at each time step.

[0038] When the main confidence channel outputs a new risk confidence value, the short-time memory pool performs an update operation in chronological order, removing the oldest stored risk confidence value from the cache and writing the risk confidence value corresponding to the current time step into the cache, keeping the amount of cached content constant.

[0039] The short-time memory pool only receives and stores the risk confidence values ​​output by the primary confidence channel, does not receive the confidence change rate output by the secondary confidence channel, and does not perform data update operations related to the confidence change rate.

[0040] Preferably, S5 specifically includes:

[0041] A volatility assessment operation is performed on the risk confidence values ​​output by the main confidence channel stored in the short-term memory pool. The volatility assessment operation calculates the time series volatility of the confidence values ​​based on a preset statistical analysis method and compares the calculation results with a first preset threshold.

[0042] If the volatility assessment result is less than the first preset threshold, a trend determination operation is performed. The trend determination operation determines whether the risk confidence value has risen continuously within a set time period by fitting the growth direction of the confidence value time series and meets the condition that the slope is greater than the set growth threshold.

[0043] If the volatility assessment result is lower than the first preset threshold, and the trend determination result shows that the confidence value is steadily rising over time, then the main confidence channel is determined to have entered the fatigue fuzzy stage.

[0044] Preferably, S6 specifically includes:

[0045] After the main confidence channel enters the fatigue fuzzy stage, a confidence compensation mechanism is executed; the confidence compensation mechanism includes a reward function adjustment process, an improved DPRNN model update termination process, and a main confidence channel alignment correction process.

[0046] The reward function adjustment process includes constructing a trend adjustment term, which is calculated and generated based on the growth rate of risk confidence values ​​in the short-term memory pool within a set time period, and is added to the reward function as an additional quantity to increase the proportion of the influence of the risk confidence value change trend in the reward function.

[0047] The improved DPRNN model update termination process includes freezing the parameter update operation of the improved DPRNN model at a specified time step, blocking the update path between the state vector and the reward feedback, and making the improved DPRNN model maintain the state expression of the previous time step.

[0048] The alignment correction process of the main confidence channel includes extracting the risk confidence values ​​corresponding to consecutive time steps in the short-time memory pool, calculating the mean of the confidence values, calculating the difference between the risk confidence value output by the main confidence channel at a specified time step and the mean of the risk confidence values, performing weighted adjustment according to a preset correction coefficient, and generating the alignment-corrected confidence value result; the alignment-corrected confidence value result is used as the final output of the main confidence channel at the corresponding time step.

[0049] Preferably, S7 specifically includes:

[0050] The raw monitoring data during the operation of the steel wire is continuously collected and input into the improved DPRNN model. The state vector corresponding to each time step is iteratively updated, and the risk confidence value output by the main confidence channel and the confidence change rate output by the auxiliary confidence channel are extracted.

[0051] Within a set number of consecutive time steps, if the confidence change rate of the auxiliary confidence channel is greater than the second preset threshold and no breakage risk event is detected, a confidence penalty mechanism is executed; the confidence penalty mechanism includes a negative adjustment process for the reward function, a confidence decay process for the main confidence channel, and a weighted decay process for the short-term memory pool.

[0052] The negative adjustment process of the reward function includes constructing a negative trend term. The negative trend term is calculated and generated based on the magnitude of the confidence change rate in the auxiliary confidence channel being greater than the second preset threshold, and is superimposed on the original reward function to form a negative trend adjustment result, thereby reducing the output intensity of the reward function.

[0053] The confidence decay process of the main confidence channel includes performing a reduction process on the risk confidence value output by the main confidence channel at the corresponding time step. The reduction magnitude is calculated based on the difference between the confidence change rate and the second preset threshold, and the attenuated confidence value result is output in combination with the preset decay coefficient.

[0054] The short-term memory pool weighted decay process includes extracting the historical confidence value sequence stored in the short-term memory pool, setting decay weights according to the time step interval, performing weighted decay calculations on each historical confidence value, and generating a corrected confidence sequence; the corrected confidence sequence is written into the short-term memory pool to replace the original historical confidence values ​​and constitute the updated short-term confidence cache content.

[0055] Preferably, S8 specifically includes:

[0056] Anomaly detection is performed on the output sequences of the main confidence channel and the auxiliary confidence channel within a short sliding window. The anomaly detection operation includes monitoring the instantaneous change amplitude, change rate and change duration of the confidence value. If the change amplitude of the confidence value at any time step is greater than the third preset threshold and the change duration is less than the set time step range, it is determined to be an abnormal change signal.

[0057] Upon detecting an abnormal mutation signal, the execution of the confidence penalty mechanism is halted, the negative adjustment process of the reward function and the weighted decay process of the short-term memory pool are frozen, and the confidence output state of the main confidence channel is prevented from decaying further.

[0058] After the confidence penalty mechanism is terminated, the action output path of the main confidence channel is activated. The action output path receives the confidence value of the main confidence channel at the time step when the abnormal mutation signal is triggered and the confidence change rate of the auxiliary confidence channel to construct the risk state input vector.

[0059] The risk state input vector is input to the fracture risk prediction unit, which performs risk assessment calculations and generates fracture risk prediction results based on the confidence fluctuation characteristics and abrupt signal characteristics in the risk state input vector.

[0060] The fracture risk prediction results include three types of output: fracture risk level, risk time index, and confidence interval range. The fracture risk level reflects the risk intensity distribution, the risk time index corresponds to the time step position of the mutation signal, and the confidence interval range marks the confidence distribution of the prediction results. The fracture risk prediction results serve as the final output data in the evaluation stage.

[0061] The beneficial effects of this invention are:

[0062] This invention addresses the challenges of heterogeneous temporal dimensions, strong characteristic fluctuations, and difficulty in identifying pre-fracture fatigue in multi-source monitoring data during steel wire operation through the joint design of an improved DPRNN model and a dual-confidence channel structure. It employs a standard input sequence construction and state vector extraction mechanism, combining a primary confidence channel and an auxiliary confidence channel to construct a risk confidence representation. A short-time memory pool structure is used to cache and analyze the confidence value sequence. In the evaluation phase, a confidence compensation mechanism and a confidence penalty mechanism are introduced to dynamically adjust the reward function and model state. In the prediction phase, the action output path is activated based on abnormal mutation signals, constructing a risk state input vector and generating fracture risk prediction results. This effectively solves the problems of smooth risk signals, high misjudgment rates, and difficulty in capturing sudden fracture signal changes during the fatigue stage. It achieves continuous monitoring, trend perception, temporal identification, and graded prediction of steel wire fracture risk, outputting three types of results: fracture risk level, risk time index, and confidence interval range. Attached Figure Description

[0063] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0064] Figure 1 This is a flowchart of a comprehensive assessment method for steel wire fracture risk based on reinforcement learning proposed in this invention;

[0065] Figure 2 This is a data flow diagram of a reinforcement learning-based comprehensive assessment method for steel wire fracture risk proposed in this invention. Detailed Implementation

[0066] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0067] refer to Figure 1-2 A comprehensive assessment method for steel wire fracture risk based on reinforcement learning includes the following steps:

[0068] S1. Collect raw monitoring data during the operation of the steel wire, normalize and format it according to a unified time step, and generate a standard input sequence;

[0069] S2. Input the standard input sequence into the improved DPRNN model to construct the state vector corresponding to the time step, and extract the risk confidence value and confidence change rate from the state vector;

[0070] S3. Construct a dual confidence channel based on the risk confidence value and confidence change rate. The main confidence channel records the risk confidence value at each time step, and the auxiliary confidence channel records the confidence change rate of the main confidence channel at consecutive time steps.

[0071] S4. Set a short-time sliding window and construct a short-time memory pool; the short-time memory pool is used to cache the confidence values ​​output by the main confidence channel in a set number of consecutive time steps;

[0072] S5. Perform volatility assessment and trend determination operations on the short-term memory pool; if the volatility of the main confidence channel is less than the first preset threshold and shows a continuous upward trend, it is determined to be in the fatigue fuzzy stage.

[0073] S6. Execute a confidence compensation mechanism in the fatigue fuzzy stage; the confidence compensation mechanism includes adjusting the reward function of the time step, introducing a trend adjustment term into the reward function, stopping the real-time update operation of the improved DPRNN model, and performing alignment correction between the confidence value of the main confidence channel and the mean confidence value in the short-term memory pool.

[0074] S7. Continue to collect raw monitoring data and iteratively update the state vector. If the confidence change rate in the auxiliary confidence channel is greater than the second preset threshold within a set number of consecutive time steps, and no breakage risk event is detected, then execute the confidence penalty mechanism. The confidence penalty mechanism includes introducing a negative trend term into the reward function, performing attenuation processing on the confidence value of the main confidence channel, and performing weighted attenuation operation on the historical confidence values ​​in the short-term memory pool.

[0075] S8. If an abnormal mutation signal is detected within a short sliding window, the execution of the confidence penalty mechanism is stopped, the action output path of the main confidence channel is activated, and the fracture risk prediction result is generated.

[0076] This implementation constructs a comprehensive risk assessment process for steel wire fracture based on reinforcement learning, proposing a multi-stage closed-loop analysis mechanism that includes standard input construction, state vector extraction, confidence channel evolution, memory pool regulation, and risk output. This mechanism possesses high risk identification sensitivity and dynamic adjustment capabilities. First, multi-source monitoring data during the steel wire's operation is collected and normalized and formatted to generate a standardized input sequence with a unified structure and time alignment, improving the consistency and effectiveness of multi-source signal fusion. The standardized input sequence is then input into an improved DPRNN model to extract state vectors containing risk confidence values ​​and confidence change rates, which helps express the risk characteristics of the steel wire's state evolution over time. Based on the state vectors, a primary and secondary dual-confidence channel structure is constructed to achieve separate channel modeling of risk confidence values ​​and confidence change rates, enhancing the discriminative power of confidence expression. A short-time sliding window is set, and a short-time memory pool is constructed to perform fixed-length caching of the primary confidence channel output, limiting the analysis scope to short-period sequences and improving the response speed to stage fluctuations and trend changes. Furthermore, by performing volatility assessment and trend determination operations, it is possible to identify a fatigued ambiguity state where the confidence value sequence is rising but with small fluctuations, thereby triggering a compensation mechanism in advance and stabilizing the risk assessment process. Going further, when the confidence change rate continues to rise but a break event is not triggered, a confidence penalty mechanism is executed to suppress and adjust the confidence expression, reducing the risk of prediction bias caused by erroneous increases in confidence values. Simultaneously, upon identifying a confidence mutation signal, the penalty mechanism is terminated and the action output path is activated to quickly generate break risk prediction results, improving the response efficiency to sudden risks.

[0077] In this embodiment, S1 specifically refers to:

[0078] Raw monitoring data of the steel wire during operation is collected. This raw monitoring data includes vibration signals, strain signals, and acoustic emission signals. The vibration signals are obtained by using an accelerometer mounted on the surface of the steel wire to capture the sequence of vibration amplitude changes over time, with a sampling frequency set to no less than 5 kHz. The strain signals are obtained by using resistance strain gauges deployed at key stress points on the steel wire to capture the sequence of strain changes over time, employing a high-sensitivity bridge-type electrical measurement method to collect minute strain changes. The acoustic emission signals are obtained by using a wide-band acoustic emission sensor array deployed at the distribution locations of the steel wire to capture the acoustic emission amplitude changes and frequency drift sequences. The sensor response frequency band covers 100 kHz to 1 MHz, capable of reflecting the microcrack propagation and energy release process.

[0079] Missing data completion processing was performed on vibration signals, strain signals, and acoustic emission signals respectively. Short-time acquisition missing segments were reconstructed using bidirectional interpolation and low-order polynomial fitting, with an interpolation window length of less than five time steps. Anomaly removal processing was performed, and an anomaly identification model based on statistical features was constructed. The sliding standard deviation method was used to identify outliers outside the confidence interval and replace them with the nearest mean. The above processing steps constructed a data channel structure with a unified format, including timestamp, standard code, unit identifier, and processing status flag fields for each type of signal.

[0080] A uniform time step length is set, which is based on the sampling frequency and the signal change rate. The suggested step length is determined by fitting the distance between local extrema of the derivative of the signal change in the whole sequence. Finally, the signal alignment reference time scale is selected as the starting time of the vibration signal.

[0081] The time synchronization processing operation is performed by using a signal resampling and window function alignment mechanism to uniformly resample the three types of signals into a sampling point sequence with the same time resolution, and construct an alignment index set that matches the starting point and step size. Based on the alignment index set, the interpolation alignment operation is performed. Piecewise linear interpolation is used for vibration signals and strain signals, and spline interpolation is used for acoustic emission signals to reconstruct the characteristic change range, thus making up for the time misalignment difference caused by sampling offset.

[0082] Normalization is performed on the time-synchronized signal data. The vibration signal and strain signal are constructed with a linear mapping function based on their maximum and minimum values ​​in the whole sequence. Minimum-maximum normalization is performed to ensure that their value range is mapped to [0, 1]. Due to the difference in magnitude, the acoustic emission signal is processed by logarithmic amplitude normalization. The original amplitude is mapped to the standard range after taking the natural logarithm. This enhances the model's ability to distinguish high-amplitude acoustic emission events.

[0083] Finally, a standard input sequence is constructed, which is organized according to a time step structure. Each time step includes a normalized vibration feature vector, strain feature vector, and acoustic emission feature vector, which are uniformly arranged into a multi-channel splicing structure.

[0084] In this embodiment, S2 specifically refers to:

[0085] The standard input sequence is divided into fixed-length local signal blocks. Each local signal block contains multi-channel data of several consecutive time steps. A set of signal blocks covering the complete time series is constructed according to a set sliding interval. When constructing the signal block set, the data of the global time period in the complete input sequence is extracted at the same time, and the context information of cross-block association is preserved. The local time period signal blocks and the global time period data together constitute the input structure. The input structure has local time series sensitivity and global time series dependency characteristics.

[0086] The input structure is fed into the improved DPRNN model, which includes a temporal dual-channel structure, a position-gated structure, a confidence generation structure, and a residual memory structure. The improved DPRNN model performs sequence modeling tasks with time steps as the main axis, constructing intra-block paths based on local feature learning and inter-block paths based on global trend association, and embedding confidence value evolution modeling and historical state fusion mechanisms.

[0087] A temporal dual-channel structure is used to extract temporal features within local blocks and temporal correlation features between blocks. The temporal dual-channel structure consists of multiple bidirectional gated recurrent units. The local path processes continuous time steps in each local signal block to extract fine-grained dynamic patterns; the global path processes connection sequences between signal blocks to extract macro trends and periodic correlations. The feature vectors output by the two paths are fused through a concatenation operation to output a dual-channel fused feature representation.

[0088] The dual-channel fused feature representation and the position vector composed of the time step index are combined and input into the position gating structure. The time step index is generated based on the position information of the standard input sequence and mapped to the position encoding vector. The position gating structure includes a position-aware attention module and a gating adjustment unit. The position-aware attention module calculates the correlation weight between different time steps based on the attention mechanism and adjusts the focus of attention in the feature representation through a weighted method to generate an information representation vector that fuses different temporal semantics.

[0089] The weighted fusion result is input into the confidence generation structure, which includes two output paths. The first output path is the risk confidence value path, which consists of at least one fully connected neural network layer. It inputs the weighted fusion result and outputs the confidence value. The second output path is the confidence change rate path, which has the same structure as the first path and outputs the confidence change trend between adjacent time steps. The two paths are output in parallel, which respectively constitute the risk confidence value and confidence change rate of the time step.

[0090] A residual memory structure is introduced into the confidence generation structure. The residual memory structure connects the dual-channel fusion feature representation of the historical time step with the fusion feature representation of the current time step, and performs a residual connection operation. The residual connection result serves as one of the inputs to the confidence generation structure, guiding the model to learn the state transitions and historical remnants during the confidence value evolution process. The final output state vector includes the risk confidence value and confidence change rate of the current time step, which serves as the basis for subsequent dual-confidence channel construction and risk state evolution analysis.

[0091] This implementation method constructs an improved DPRNN model, integrating local and global dual-channel learning paths, time-location gating mechanisms, and confidence evolution modeling mechanisms. This enables the capture of the temporal sensitivity of wire fracture risk and the enhanced identification of trends, thereby improving the accuracy and interpretability of the state vector in expressing the evolutionary characteristics of fracture risk.

[0092] In this embodiment, the improved DPRNN model includes a temporal dual-channel structure, a position-gated structure, a confidence generation structure, and a residual memory structure, specifically:

[0093] The temporal dual-channel structure comprises two parts: an intra-block path and an inter-block path. The intra-block path includes multiple bidirectional gated loop units, which input a local signal block sequence and perform forward and backward temporal feature extraction on the continuous time step data within each signal block to extract local short-term dynamic change features. The inter-block path includes multiple bidirectional gated loop units, which input a global signal sequence spanning local blocks and extract the long-term correlation and evolution trend between different signal blocks to generate a global feature representation. The output results of the local path and the global path are fused by concatenation to form a dual-channel fused feature representation. The fused feature representation retains both local temporal change details and global trend correlations, serving as input for subsequent modules.

[0094] The position gating structure includes a temporal position encoding module and a gating attention unit. The temporal position encoding module receives the time index of each time step in the standard input sequence, encodes the time index using sine and cosine functions, and generates a position encoding vector. The encoding method is consistent with the temporal position encoding in the Transformer structure. The gating attention unit includes an attention score calculation module and a gating adjustment module. The attention score calculation module performs correlation calculation between the dual-channel fused feature representation and the position vector to construct an attention weight distribution. The gating adjustment module performs a weighted summation operation on the dual-channel fused features according to the attention weight distribution and introduces a gating factor to adjust the fusion intensity at each time step, finally outputting the fused feature result.

[0095] The confidence generation structure includes a first output path and a second output path. The first output path receives the fused feature result and contains at least one layer of fully connected neural network and activation function combination structure, outputting the risk confidence value corresponding to each time step. The second output path structure is consistent with the first output path, receiving the fused feature result and outputting the confidence change rate. During training, the risk confidence value output uses the mean squared error loss function to fit the labeled risk probability, and the confidence change rate output uses time series difference labels for supervised learning, constructing a confidence change rate supervision signal through the time step difference function.

[0096] The residual memory structure constructs a historical state information pathway, extracts the dual-channel fusion feature representation of the previous time step from the cache of the previous time step, and performs a residual connection operation with the dual-channel fusion feature representation of the current time step; the residual connection performs linear weighted fusion by subtracting vectors and introducing residual coefficients, and the result serves as an additional input to the confidence generation structure; the residual memory structure provides a temporal continuity reference in the modeling of risk state change processes, and improves the sensitivity of the state vector to trend changes and the ability to retain historical evolution information.

[0097] This implementation enhances the temporal modeling capability of the improved DPRNN model by constructing a temporal dual-channel structure and a position-gated structure. It also improves the stability and trend expression capability of the risk confidence output by combining a confidence generation structure and a residual memory structure, thereby achieving accurate expression and dynamic early warning analysis of wire breakage risk during continuous operation.

[0098] In this embodiment, S3 specifically refers to:

[0099] After extracting the risk confidence value and confidence change rate corresponding to each time step, a dual confidence channel structure is constructed. The dual confidence channel structure includes a main confidence channel and an auxiliary confidence channel. The dual confidence channel structure realizes the multidimensional representation of confidence information by separating the static confidence level and the dynamic confidence change trend.

[0100] The main confidence channel receives the risk confidence values ​​from the time step state vectors and arranges the risk confidence values ​​of all time steps in chronological order to form the main confidence sequence. The main confidence sequence stores the confidence strength values ​​with time steps as indices, recording the overall evolution trajectory of the risk level during operation. A time buffer is set in the main confidence channel to store a set number of historical confidence values, forming a continuous time window to support subsequent sliding window analysis and short-term trend estimation.

[0101] The secondary confidence channel takes the primary confidence sequence as input and performs a rate of change calculation based on the confidence values ​​between adjacent time steps. The rate of change is obtained through a difference function, which uses a sliding first-order difference method to calculate the confidence difference between time step t and time step t−1, and then uses the time interval as a divisor to obtain the normalized rate of change. The rate of change sequence is processed by a smoothing function to remove high-frequency fluctuation components, forming a stable confidence rate of change sequence. During the model training phase, the rate of change can be fitted to the actual risk change trend through a time series fitting function to ensure the dynamic sensitivity and trend accuracy of the confidence rate of change.

[0102] At each time step, the main confidence channel outputs the risk confidence value corresponding to that time step, and the auxiliary confidence channel outputs the rate of change of the confidence value between that time step and the previous time step. The outputs of the two channels together constitute the confidence expression matrix. The confidence expression matrix is ​​a two-dimensional time series structure, with the first dimension corresponding to the time step index and the second dimension corresponding to the confidence strength and trend. The confidence expression matrix serves as the input basis for confidence dynamic analysis and trend modeling, providing quantitative support for subsequent fatigue fuzzy stage identification and confidence compensation mechanism.

[0103] This implementation constructs a dual-confidence channel structure to separate and express the static distribution and dynamic trend of risk confidence values, thereby achieving a dual-layer characterization of confidence information and a coupled trend representation, enhancing the model's time sensitivity and anomaly detection accuracy in continuous risk status monitoring.

[0104] In this embodiment, S4 specifically refers to:

[0105] Set the length parameter of the short-term sliding window to a fixed integer value, and define the span of the sliding interval in units of time steps; construct a short-term sliding window according to a set number of consecutive time steps, and continuously cache the risk confidence values ​​of the main confidence channel to form a fixed-length time series buffer; discard the confidence value of the earliest time step and include the confidence value of the current time step as the sliding window moves forward one step.

[0106] A short-term memory pool structure is constructed, which is a fixed-length cache module with a cache capacity consistent with the sliding window length. During the initialization phase, an empty value structure is filled to maintain the integrity of the time step alignment structure. The short-term memory pool adopts a circular buffer mechanism, which efficiently writes and updates confidence value records through pointer rotation.

[0107] When the main confidence channel outputs a new risk confidence value, the short-term memory pool automatically replaces the earliest confidence value record according to the time step index order and inserts the confidence value corresponding to the current time step at the end; the update operation is implemented through a queue structure to achieve dual execution of dequeue and enqueue operations, maintaining the temporal continuity and quantity stability of the data stored in the short-term memory pool;

[0108] The short-term memory pool only receives and stores the risk confidence values ​​output by the main confidence channel, and does not receive the confidence change rate output by the secondary confidence channel; the confidence change rate is excluded from the cache structure to avoid introducing interference factors with drastic fluctuations into the cache structure, thereby improving the stability and accuracy of subsequent volatility assessment.

[0109] The short-time memory pool structure supports direct reading of historical confidence value sequences, moving average calculation, local fluctuation amplitude evaluation, and trend fitting operations; the confidence sequence output by the short-time memory pool will serve as the core input structure for trend determination and stage identification.

[0110] This implementation achieves a temporal continuous caching and efficient update mechanism for risk confidence values ​​by setting a sliding window and constructing a short-time memory pool structure. It provides timely and structurally complete data support for state identification and trend evolution modeling in the fatigue fuzzy stage, and improves the perception sensitivity and system response advance in the early stage of fracture risk evolution.

[0111] In this embodiment, S5 specifically refers to:

[0112] A volatility assessment operation is performed on the time series composed of risk confidence values ​​of consecutive time steps in the short-term memory pool. The volatility assessment operation includes: calculating the mean and standard deviation of the risk confidence value series, constructing a confidence volatility index, wherein the confidence volatility index is equal to the ratio of the standard deviation to the mean; comparing the volatility index with a set first preset threshold, and if it is less than the first preset threshold, the confidence value series is determined to be in a low volatility state.

[0113] After the volatility assessment results meet the low volatility condition, the trend determination operation is performed. The trend determination operation includes: constructing a time series fitting function, using the least squares method to linearly fit the risk confidence value series, calculating the fitting slope k and comparing it with the set growth threshold. If k is greater than the growth threshold and the sum of squared fitting residuals is less than the upper limit of stability tolerance, then the confidence value series is determined to have a stable upward trend.

[0114] The determination of a stable upward trend further introduces a time continuity constraint. The continuity constraint includes: within a set number of time steps, the risk confidence value at each step is not less than the minimum growth difference δ compared to the previous time step, and there is no reverse downward sudden change; if the continuity condition is met, the trend fitting slope is greater than the threshold, and the volatility is lower than the first preset threshold, then the stage in which the main confidence channel is located is determined to be the fatigue fuzzy stage.

[0115] The fatigue ambiguity stage marks the precursor stage of the risk state evolving from potential accumulation to sudden change. During this stage, the confidence value steadily increases but has not yet shown drastic disturbances. The above judgment strategy can identify the ambiguity-induced stage of fracture risk in advance.

[0116] This implementation method constructs a fatigue fuzzy stage identification mechanism with high robustness and high sensitivity by combining volatility assessment and trend determination operations, which significantly improves the early response capability and intelligent identification accuracy to the potential evolution trend of wire fracture risk.

[0117] In this embodiment, S6 specifically refers to:

[0118] After the main confidence channel enters the fatigue fuzzy stage, a confidence compensation mechanism is constructed. The confidence compensation mechanism includes a reward function adjustment process, an improved DPRNN model update termination process, and a main confidence channel alignment correction process.

[0119] In the reward function adjustment process, the risk confidence value sequence within a specified time period is extracted from the short-time memory pool; a growth rate sequence is constructed using the difference method, the difference between confidence values ​​at adjacent time steps is calculated, and the difference sequence is weighted and averaged to obtain the average growth rate of risk confidence values; this growth rate is input into the trend adjustment term construction function, and a trend adjustment term is generated through a nonlinear mapping function. The trend adjustment term is then superimposed as an additional term onto the original reward function to enhance the sensitivity of the reward function to the trend of confidence value changes.

[0120] In the improved DPRNN model update termination process, a freeze flag is set. When the fatigue fuzzy stage is determined to be triggered, the freeze flag is activated. The freeze flag acts on the model parameter update path, blocking the error backpropagation path after the model forward propagation, so that the parameters no longer undergo gradient updates. The state vector output by the improved DPRNN model in the freeze stage only depends on the historical state representation and does not introduce the error information of the new time step, thereby maintaining the stability of the state representation.

[0121] In the alignment and correction process of the main confidence channel, the risk confidence values ​​cached in the short-term memory pool are statistically analyzed, and the mean is calculated to obtain the reference confidence mean. The risk confidence values ​​output by the main confidence channel at the freeze time step are extracted, and the difference between the two is calculated. The alignment weight coefficient is set, the alignment and correction confidence value is calculated, and the confidence value is used as the final output of the main confidence channel at the specified time step, replacing the original confidence value.

[0122] In this embodiment, S7 specifically refers to:

[0123] The system continuously collects raw monitoring data during the operation of the steel wire; inputs the raw monitoring data into the improved DPRNN model to update the state vector corresponding to each time step; extracts the risk confidence value output by the main confidence channel and the confidence change rate output by the auxiliary confidence channel from the state vector; records the confidence change rate sequence and fracture risk event detection results within a set number of consecutive time steps; when no fracture risk event is detected in all consecutive time steps, and the confidence change rate of the auxiliary confidence channel exceeds the second preset threshold in at least one time step, a confidence penalty mechanism is executed.

[0124] The confidence penalty mechanism includes a negative adjustment process for the reward function, a confidence decay process for the main confidence channel, and a weighted decay process for the short-term memory pool. In the negative adjustment process for the reward function, a negative trend term is constructed. The confidence change rate of each time step in a set number of consecutive time steps is subtracted from a second preset threshold to form a change rate exceeding the threshold difference sequence. The difference sequence is input into the weighting function. The negative trend term is output. The negative trend term is superimposed on the original reward function to form the reward value after negative trend adjustment, thereby reducing the reinforcement degree of the reward function.

[0125] In the confidence attenuation process of the main confidence channel, the confidence value output by the main confidence channel at the current time step is attenuated; the difference between the confidence change rate at the current time step and the second preset threshold is calculated; the attenuation coefficient is set, the confidence value attenuation function is constructed, and the attenuated confidence value is output to replace the original main confidence channel output value.

[0126] In the short-term memory pool weighted decay process, the confidence values ​​of all historical time steps in the short-term memory pool are extracted to form a confidence history sequence; a time decay weight sequence is set, decreasing in order of time proximity; weighting is performed on each element in the historical confidence value sequence to form a confidence value sequence after decay adjustment; the new confidence value sequence is written into the short-term memory pool to overwrite the original cached content, forming the updated confidence cache state.

[0127] In this embodiment, S8 specifically refers to:

[0128] Anomaly detection is performed on the output sequence of the main confidence channel and the auxiliary confidence channel within a short sliding window. The anomaly detection operation is based on a preset set of detection indicators, which includes the instantaneous change amplitude, change rate, and change duration of the risk confidence value. A change amplitude threshold, a change rate threshold, and a maximum change duration are set as the judgment criteria. If the change amplitude of the risk confidence value at any time step exceeds the change amplitude threshold and the corresponding change duration is less than the maximum change duration, then an abnormal change signal is determined to exist at that time step.

[0129] Upon detecting an abnormal mutation signal, immediately halt the execution of the confidence penalty mechanism; freeze the dynamic adjustment of the negative trend term in the reward function; suspend the decay calculation process of the confidence values ​​stored in the short-term memory pool, and maintain the confidence output result of the main confidence channel at the trigger time step unchanged;

[0130] After the confidence penalty mechanism is terminated, the action output path of the main confidence channel is activated; the action output path receives the risk confidence value corresponding to the main confidence channel at the time of the mutation and the confidence change rate output by the auxiliary confidence channel to construct the risk state input content.

[0131] Input the risk status information into the fracture risk prediction unit; the fracture risk prediction unit performs fracture risk assessment calculations based on the input confidence change characteristics and fluctuation trend characteristics, and outputs prediction results including fracture risk level, risk time index and confidence interval range;

[0132] The fracture risk level indicates the risk level corresponding to the current confidence mutation event; the risk time index corresponds to the time step position triggered by the abnormal mutation signal; the confidence interval range indicates the range of credibility of the prediction output results; the fracture risk prediction results serve as the final output of the comprehensive assessment process for steel wire fracture risk.

[0133] Example 1:

[0134] To verify the feasibility of this invention in practice, it was applied to a steel wire rope monitoring system for a large port crane. The system's steel wire ropes are approximately 120 meters long, with a single rope bearing a weight exceeding 80 tons. It performs over 400 lifting and lowering operations daily, operating under complex conditions of high stress, strong vibration, frequent bending, and a humid environment for extended periods. Traditional methods for assessing steel wire breakage risk rely primarily on manual inspection and threshold alarms, issuing an alarm when stress or vibration values ​​exceed set ranges. However, this method suffers from problems such as delayed response, fixed thresholds, and inability to identify progressive fatigue, often leading to frequent false alarms or missed detections, failing to provide maintenance personnel with sufficient advance warning time.

[0135] In this implementation scenario, the system is equipped with a multi-source sensor array, including accelerometers, strain gauges, and acoustic emission sensors distributed along the steel wire, for real-time acquisition of vibration, strain, and acoustic emission signals. The acquired signals are normalized and time-synchronized to form a standard input sequence, which is then input into an improved DPRNN model for feature extraction and risk state vector generation. During operation, the model continuously optimizes its internal reward mechanism through reinforcement learning, making the identification of confidence rate of change, fluctuation trends, and fatigue ambiguity stages more accurate. The main confidence channel outputs a risk confidence value to determine the risk level, while the auxiliary confidence channel outputs a confidence rate of change to detect abrupt changes and unstable trends. When a sustained increase in risk confidence with low volatility is detected, the system automatically enters a confidence compensation phase to prevent the model from falsely triggering excessive risk warnings. Conversely, when the confidence rate of change increases abnormally and a sudden change signal appears within a short sliding window, the system automatically activates a confidence penalty mechanism to correct the model's judgment.

[0136] During the three-month trial period, three groups of monitoring objects were selected for comparison: prototypes A and B were assessed for fracture risk using the method of this invention, while prototype C was assessed using the traditional fixed threshold method. All three groups of equipment operated under identical conditions, with the same sensor layout, sampling cycle, and data synchronization standards, and a monitoring frequency of once per minute. During the trial, the number of risk detections, false alarms and missed alarms, the accuracy of risk level assessment, and the pre-fracture warning time were statistically analyzed for each prototype.

[0137] Table 1 Comparison of Steel Wire Fracture Risk Detection Performance

[0138] Prototype testing Total monitoring duration (hours) Actual number of fracture events False alarm count Number of missed reports Accuracy rate of risk level assessment (%) Average advance warning time (minutes) Prototype A (Method of this invention) 540 3 2 0 96.8 15.3 Prototype B (Method of this invention) 538 2 1 0 95.7 14.7 Prototype C (Traditional Method) 541 3 10 2 78.2 4.1

[0139] As shown in Table 1, prototypes A and B using the method of this invention exhibited almost no missed alarms within three months, with a significant reduction in false alarms and an average early warning time exceeding 14 minutes. In contrast, the average warning time of the traditional method was only 4 minutes, indicating a significant lag. The accuracy of risk level judgment was also significantly improved. This method can identify progressive fatigue trends, abnormal abrupt changes, and potential fracture risks in real time, forming a continuous and reliable early warning curve. Traditional methods rely on a single stress threshold, making it difficult to distinguish between high-stress but stable normal working conditions and sudden abnormal fluctuations, resulting in a false alarm rate more than three times higher. Especially in the high-temperature and humid environment of the second month, the stable identification capability of this method remained above 95%, demonstrating good anti-interference and adaptive learning characteristics. This indicates that the method of this invention can effectively solve the problems of untimely risk identification, high false alarm rate, and insufficient response to complex environmental changes in traditional methods.

[0140] To further verify the contribution of the confidence mechanism in fracture prediction, the accuracy of fracture risk prediction and the corresponding risk level distribution of each prototype at different stages were statistically analyzed. The results are shown in the table below:

[0141] Table 2 Comparison of fracture risk prediction effects before and after the introduction of the confidence mechanism

[0142] Time phase Accuracy without confidence mechanism (%) Accuracy (%) with confidence mechanism Accuracy improvement (%) False alarm rate decrease percentage (%) Month 1 83.2 89.6 6.4 15.3 2nd month 82.5 91.1 8.6 18.2 3rd month 84.1 92.4 8.3 17.9 average value 83.3 91.0 7.7 17.1

[0143] As shown in Table 2, with the introduction of confidence compensation and confidence penalty mechanisms, the overall accuracy of fracture risk prediction increased from an average of 83.3% to 91.0%, an improvement of 7.7%, while the false alarm rate decreased by an average of 17.1%. This result demonstrates that the reinforcement learning strategy combining dual confidence channels with dynamic correction via a short-term memory pool can more effectively identify patterns of change in risk confidence values ​​and significantly reduce misjudgments caused by noise disturbances or short-term anomalies. When the system is in the fatigue fuzzy stage, the addition of a trend adjustment term enhances the model's ability to perceive upward trends, while the penalty mechanism rapidly suppresses confidence bias in anomaly detection, making the prediction results more stable and reliable. Especially in the continuous high humidity testing environment of the second month, the accuracy of traditional methods decreased significantly, while the accuracy of this method remained above 90%, demonstrating excellent robustness and generalization ability.

[0144] Based on three months of operational data, this embodiment demonstrates its ability to achieve a closed-loop optimization process involving multi-signal collaborative sensing, state self-learning, and dynamic confidence correction. It not only significantly outperforms traditional methods in terms of detection accuracy and response time but also effectively reduces maintenance costs and safety hazards caused by false alarms and missed alarms. This method has broad engineering application value in fields such as heavy-duty lifting, cableway transportation, and high-strength traction, providing intelligent and sustainable technical support for the health monitoring and life prediction of steel wire structures.

[0145] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A comprehensive assessment method for steel wire fracture risk based on reinforcement learning, characterized in that, Includes the following steps: S1. Collect raw monitoring data during the operation of the steel wire, normalize and format it according to a unified time step, and generate a standard input sequence; S2. Input the standard input sequence into the improved DPRNN model to construct the state vector corresponding to the time step, and extract the risk confidence value and confidence change rate from the state vector; The improved DPRNN model includes a temporal dual-channel structure, a position-gated structure, a confidence generation structure, and a residual memory structure, specifically: The timing dual-channel structure includes multiple bidirectional gated loop units, which receive the local signal block sequence and the global signal block sequence after being divided in the standard input sequence, respectively extract the timing features within the local block and the timing correlation features between cross blocks, and generate a dual-channel fused feature representation. The position gating structure includes a time position encoding module and a gating attention unit. The time position encoding module generates a corresponding position vector based on the time step index. The gating attention unit receives the dual-channel fused feature representation and the position vector, constructs an attention weight distribution, performs weighted adjustment on the features of each time step, and generates a fused feature result. The confidence generation structure includes a first output path and a second output path. The first output path includes at least one fully connected network structure and outputs a risk confidence value. The second output path includes at least one fully connected network structure and outputs a confidence change rate. The residual memory structure constructs a historical state information pathway, performs a residual connection between the fusion feature representation of the previous time step and the fusion feature representation of the current time step, and inputs the connection result into the confidence generation structure to generate a state vector containing historical evolution information. S3. Construct a dual confidence channel based on the risk confidence value and confidence change rate. The main confidence channel records the risk confidence value at each time step, and the auxiliary confidence channel records the confidence change rate of the main confidence channel at consecutive time steps. S4. Set a short-term sliding window and build a short-term memory pool; The short-time memory pool is used to cache the confidence values ​​output by the main confidence channel in a set number of consecutive time steps; S5. Perform volatility assessment and trend determination operations on the short-term memory pool; If the volatility of the primary confidence channel is less than the first preset threshold and shows a continuous upward trend, it is determined to be in the fatigue fuzzy stage. S6. Execute a confidence compensation mechanism in the fatigue fuzzy stage; the confidence compensation mechanism includes adjusting the reward function of the time step, introducing a trend adjustment term into the reward function, stopping the real-time update operation of the improved DPRNN model, and performing alignment correction between the confidence value of the main confidence channel and the mean confidence value in the short-term memory pool. S7. Continue to collect raw monitoring data and iteratively update the state vector. If the confidence change rate in the auxiliary confidence channel is greater than the second preset threshold within a set number of consecutive time steps, and no breakage risk event is detected, then execute the confidence penalty mechanism. The confidence penalty mechanism includes introducing a negative trend term into the reward function, performing attenuation processing on the confidence value of the main confidence channel, and performing weighted attenuation operation on the historical confidence values ​​in the short-term memory pool. S8. If an abnormal mutation signal is detected within a short sliding window, the execution of the confidence penalty mechanism is stopped, the action output path of the main confidence channel is activated, and the fracture risk prediction result is generated.

2. The method for comprehensive assessment of steel wire fracture risk based on reinforcement learning according to claim 1, characterized in that, Specifically, S1 is: Raw monitoring data of the steel wire during operation is collected; the raw monitoring data includes vibration signals, strain signals, and acoustic emission signals; the vibration signals are obtained by acquiring vibration amplitude sequences through accelerometers installed on the surface of the steel wire; the strain signals are obtained by acquiring strain variable sequences through strain gauges deployed at key nodes of the steel wire; the acoustic emission signals are obtained by acquiring acoustic emission amplitude and frequency change sequences through a distributed acoustic emission sensor array. Missing data completion and anomaly removal are performed on vibration signals, strain signals and acoustic emission signals respectively to construct a data channel structure with a unified format. Set a uniform time step length and alignment reference time scale, perform time synchronization processing and interpolation alignment operations; perform normalization operations on the time-synchronized signal data, normalize the vibration signal and strain signal according to the minimum and maximum ranges, normalize the acoustic emission signal according to the logarithmic amplitude, and construct a standard input sequence arranged in time order.

3. The method for comprehensive assessment of steel wire fracture risk based on reinforcement learning according to claim 2, characterized in that, Specifically, S2 is: The standard input sequence is divided into fixed-length local signal blocks to construct an input structure that includes local time intervals and overall time intervals; The input structure is input into the improved DPRNN model, which includes a temporal dual-channel structure, a position-gated structure, a confidence generation structure, and a residual memory structure. The temporal features within local blocks and the temporal correlation features between blocks are extracted by a temporal dual-channel structure, and the dual-channel fused feature representation is output. The position vector, composed of the dual-channel fused feature representation and the time step index, is input into the position gating structure, and an attention weighting operation is performed to generate a weighted fusion result. The weighted fusion result is input into the confidence generation structure. The first output path outputs the risk confidence value, and the second output path outputs the confidence change rate. A residual memory structure is introduced into the confidence generation structure. It receives the fused feature representation of the previous time step and the fused feature representation of the current time step, performs residual connection, and then participates in the confidence output together. Finally, a state vector containing risk confidence value and confidence rate of change is constructed.

4. The comprehensive assessment method for steel wire fracture risk based on reinforcement learning according to claim 3, characterized in that, Specifically, S3 is: After extracting the risk confidence value and confidence change rate corresponding to each time step, a dual confidence channel structure is constructed, which includes a primary confidence channel and a secondary confidence channel. The primary confidence channel receives the risk confidence values ​​in the time step state vector and records the risk confidence values ​​of each time step in chronological order as the primary confidence sequence. The secondary confidence channel calculates the rate of change between adjacent time steps based on the continuous time step confidence value sequence of the primary confidence channel. The rate of change is obtained by difference and forms a confidence rate of change sequence. At each time step, the main confidence channel outputs the risk confidence value corresponding to that time step, and the auxiliary confidence channel outputs the rate of change of the confidence value between that time step and the previous time step. The outputs of the two confidence channels together form the confidence expression matrix.

5. The comprehensive assessment method for steel wire fracture risk based on reinforcement learning according to claim 4, characterized in that, Specifically, S4 is: Set the length parameter of the short-time sliding window, construct the short-time sliding window according to the set number of continuous time steps, and define the range of the short-time confidence cache. A short-time memory pool structure is constructed, which is a fixed-length cache module that sequentially stores the risk confidence values ​​output by the main confidence channel at each time step. When the main confidence channel outputs a new risk confidence value, the short-time memory pool performs an update operation in chronological order, removing the oldest stored risk confidence value from the cache and writing the risk confidence value corresponding to the current time step into the cache, keeping the amount of cached content constant. The short-time memory pool only receives and stores the risk confidence values ​​output by the primary confidence channel, does not receive the confidence change rate output by the secondary confidence channel, and does not perform data update operations related to the confidence change rate.

6. The method for comprehensive assessment of steel wire fracture risk based on reinforcement learning according to claim 5, characterized in that, Specifically, S5 is: A volatility assessment operation is performed on the risk confidence values ​​output by the main confidence channel stored in the short-term memory pool. The volatility assessment operation calculates the time series volatility of the confidence values ​​based on a preset statistical analysis method and compares the calculation results with a first preset threshold. If the volatility assessment result is less than the first preset threshold, a trend determination operation is performed. The trend determination operation determines whether the risk confidence value has risen continuously within a set time period by fitting the growth direction of the confidence value time series and meets the condition that the slope is greater than the set growth threshold. If the volatility assessment result is lower than the first preset threshold, and the trend determination result shows that the confidence value is steadily rising over time, then the main confidence channel is determined to have entered the fatigue fuzzy stage.

7. The method for comprehensive assessment of steel wire fracture risk based on reinforcement learning according to claim 6, characterized in that, Specifically, S6 is: After the main confidence channel enters the fatigue fuzzy stage, a confidence compensation mechanism is executed; the confidence compensation mechanism includes a reward function adjustment process, an improved DPRNN model update termination process, and a main confidence channel alignment correction process. The reward function adjustment process includes constructing a trend adjustment term, which is calculated and generated based on the growth rate of risk confidence values ​​in the short-term memory pool within a set time period, and is added to the reward function as an additional quantity to increase the proportion of the influence of the risk confidence value change trend in the reward function. The improved DPRNN model update termination process includes freezing the parameter update operation of the improved DPRNN model at a specified time step, blocking the update path between the state vector and the reward feedback, and making the improved DPRNN model maintain the state expression of the previous time step. The alignment correction process of the main confidence channel includes extracting the risk confidence values ​​corresponding to consecutive time steps in the short-time memory pool, calculating the mean of the confidence values, calculating the difference between the risk confidence values ​​output by the main confidence channel at a specified time step and the mean of the risk confidence values, performing weighted adjustment according to a preset correction coefficient, and generating the alignment-corrected confidence value result. The aligned and corrected confidence values ​​are used as the final output of the main confidence channel at the corresponding time step.

8. The method for comprehensive assessment of steel wire fracture risk based on reinforcement learning according to claim 7, characterized in that, Specifically, S7 is: The raw monitoring data during the operation of the steel wire is continuously collected and input into the improved DPRNN model. The state vector corresponding to each time step is iteratively updated, and the risk confidence value output by the main confidence channel and the confidence change rate output by the auxiliary confidence channel are extracted. Within a set number of consecutive time steps, if the confidence change rate of the auxiliary confidence channel is greater than the second preset threshold and no breakage risk event is detected, a confidence penalty mechanism is executed; the confidence penalty mechanism includes a negative adjustment process for the reward function, a confidence decay process for the main confidence channel, and a weighted decay process for the short-term memory pool. The negative adjustment process of the reward function includes constructing a negative trend term. The negative trend term is calculated and generated based on the magnitude of the confidence change rate in the auxiliary confidence channel being greater than the second preset threshold, and is superimposed on the original reward function to form a negative trend adjustment result, thereby reducing the output intensity of the reward function. The confidence decay process of the main confidence channel includes performing a reduction process on the risk confidence value output by the main confidence channel at the corresponding time step. The reduction magnitude is calculated based on the difference between the confidence change rate and the second preset threshold, and the attenuated confidence value result is output in combination with the preset decay coefficient. The short-term memory pool weighted decay process includes extracting the historical confidence value sequence stored in the short-term memory pool, setting decay weights according to the time step interval, performing weighted decay calculations on each historical confidence value, and generating a corrected confidence sequence; the corrected confidence sequence is written into the short-term memory pool to replace the original historical confidence values ​​and constitute the updated short-term confidence cache content.

9. The method for comprehensive assessment of steel wire fracture risk based on reinforcement learning according to claim 8, characterized in that, Specifically, S8 is: Anomaly detection is performed on the output sequences of the main confidence channel and the auxiliary confidence channel within a short sliding window. The anomaly detection operation includes monitoring the instantaneous change amplitude, change rate and change duration of the confidence value. If the change amplitude of the confidence value at any time step is greater than the third preset threshold and the change duration is less than the set time step range, it is determined to be an abnormal change signal. Upon detecting an abnormal mutation signal, the execution of the confidence penalty mechanism is halted, the negative adjustment process of the reward function and the weighted decay process of the short-term memory pool are frozen, and the confidence output state of the main confidence channel is prevented from decaying further. After the confidence penalty mechanism is terminated, the action output path of the main confidence channel is activated. The action output path receives the confidence value of the main confidence channel at the time step when the abnormal mutation signal is triggered and the confidence change rate of the auxiliary confidence channel to construct the risk state input vector. The risk state input vector is input to the fracture risk prediction unit, which performs risk assessment calculations and generates fracture risk prediction results based on the confidence fluctuation characteristics and abrupt signal characteristics in the risk state input vector. The fracture risk prediction results include three types of output: fracture risk level, risk time index, and confidence interval range. The fracture risk level reflects the risk intensity distribution, the risk time index corresponds to the time step position of the mutation signal, and the confidence interval range marks the confidence distribution of the prediction results. The fracture risk prediction results serve as the final output data in the evaluation stage.

Images