A precise trigger synchronous multi-channel vibration data acquisition system

Abstract

The application discloses a kind of accurate trigger synchronous multichannel vibration data acquisition system, comprising: data acquisition and pre-processing module, for reading data and generating vibration state sample and action state pair;Mutual inhibition reserve pool echo state network module, for executing dynamic feature extraction to vibration state sample;Event sensitive trigger gate module, for generating gated activation vector and timing screening is carried out to reserve pool activation state, output event response feature vector;Reverse memory drive DRQN model module, for receiving event response feature vector and executing forward recursion and reverse recursion, output Q value estimation and policy action;Error training module, for calculating the Q value error of policy action and generating training signal;Experience replay module, for constructing priority experience replay queue based on Q value error and selecting training segment to update DRQN model parameter.The application realizes the high-precision trigger and synchronous acquisition of multichannel vibration signal, improves the stability of feature and policy decision performance.

Description

Technical Field

[0001] This invention relates to the field of intelligent trigger control, and in particular to a multi-channel vibration data acquisition system for precise trigger synchronization. Background Technology

[0002] Existing multi-channel vibration signal acquisition methods generally rely on fixed threshold triggering or local triggering strategies based on single-channel characteristics. In complex vibration environments, these methods are prone to problems such as trigger delay, false triggering, and missed triggering, resulting in the inability to accurately capture key vibration events. Traditional triggering mechanisms mainly rely on manually set rules and lack temporal memory capabilities. They are unable to form stable judgment criteria for rapidly changing vibration signals and perform poorly under non-stationary noise or weak event conditions.

[0003] In the field of multi-channel vibration signal processing, time-series-oriented deep learning methods are gradually being introduced. However, existing methods typically only extract features based on forward time series, lacking backward correlation modeling for future time segments. This makes it difficult for hidden states to accurately represent the dynamic trends of vibration events. Existing echo state network structures mostly adopt a unidirectional reservoir update method, which is insufficient for handling the mutual interference and mutual excitation characteristics in multi-channel correlated vibration signals. The reservoir states are prone to drift or over-response, failing to generate stable dynamic features. In reinforcement learning-related trigger policy training, traditional experience playback mechanisms generally use random or uniform sampling methods, failing to consider the importance of distinguishing samples from different time periods, and failing to incorporate temporal context information for segment selection. This results in low policy update efficiency, slow training convergence speed, and limited improvement in policy performance.

[0004] Therefore, how to provide a multi-channel vibration data acquisition system that can accurately trigger synchronization is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a multi-channel vibration data acquisition system with precise triggering and synchronization. This invention uses a mutually inhibited reservoir echo state network to extract dynamic features and generates a triggering strategy through a reverse memory-driven DRQN ​​model to achieve high-precision triggering and synchronous acquisition of multi-channel vibration signals. This invention has the advantages of high triggering accuracy and sensitivity to weak events.

[0006] A multi-channel vibration data acquisition system for precise triggering synchronization according to an embodiment of the present invention includes: The data acquisition and preprocessing module is used to read the timing signals of the multi-channel vibration sensor and the corresponding control state information, and to perform preprocessing on the timing signals to generate vibration state samples and action state pairs. The mutual inhibition reservoir echo state network module is used to perform input mapping, mutual inhibition reservoir state update and dynamic feature extraction based on the vibration state samples generated by the data acquisition and preprocessing module, and output compressed dynamic feature representation. The event-sensitive trigger gate module is used to generate a gated activation vector based on the control state information, and to perform time-series filtering on the intermediate states output by the mutual inhibition reservoir echo state network module, and output an event response feature vector. The reverse memory-driven DRQN ​​model module is used to receive the event response feature vector output by the event-sensitive trigger gate module, perform bidirectional time-aware Q-value estimation, and output policy actions. The error training module is used to calculate the Q-value error based on the policy action output by the reverse memory-driven DRQN ​​model module and generate training signals. The experience playback module is used to construct a priority experience playback queue based on the Q-value error, select training segments based on the reverse memory driving rule, update the parameters of the DRQN ​​model, and form a multi-channel vibration recognition strategy model.

[0007] Optionally, modules can be integrated using the following methods: S1. Acquire timing signals from multi-channel vibration sensors and corresponding control status information, perform preprocessing, and construct vibration status samples and action status pairs. S2. Construct an echo state network with a mutually inhibited reserve pool structure, initialize the input weight matrix, the mutually inhibited connection matrix, the output weight matrix and the state update parameters, perform input mapping and reserve state evolution on the vibration state samples, and extract compressed dynamic feature representations. S3. Construct an event-sensitive trigger gate mechanism, generate a gated activation vector based on the control state information, perform temporal filtering on the intermediate states of the echo state network, and extract event response feature vectors. S4. Input the event response feature vector into the DRQN ​​model with reverse memory driving mechanism, construct a bidirectional time-aware Q-value estimation structure, and initialize the Q-value function and policy output layer. S5. Input training is performed on the action-state pairs, forward inference and backpropagation of error are executed, and the Q-value error of the policy output is calculated. S6. Construct a priority experience replay queue based on Q-value error, select training segments using reverse memory-driven rules, and update the parameters of the DRQN ​​model; S7. Repeat steps S2 to S6 until the Q-value function converges, and output the multi-channel vibration recognition strategy model.

[0008] Optionally, S2 specifically includes: S21. Establish a mutually inhibited reservoir echo state network, set the input weight matrix, the mutual inhibition connection matrix, the output weight matrix and the state update parameters, and send the vibration state samples into the mutually inhibited reservoir echo state network in chronological order. S22. Perform matrix multiplication on the vibration state sample at each moment, generate an input mapping vector using the input weight matrix, and send the input mapping vector into the reserve pool state update unit. S23. Read the previous state from the reserve state storage unit, perform matrix multiplication and vector addition operations, and combine the previous state with the input mapping vector to form the unsuppressed state update quantity. S24. Perform a matrix multiplication operation on the unsuppressed state update quantity using the mutual inhibition connection matrix to generate the mutual inhibition adjustment quantity, and perform a vector subtraction operation on the mutual inhibition adjustment quantity and the unsuppressed state update quantity to form the mutual inhibition processing intermediate state. S25. Input the intermediate state of mutual inhibition processing into the nonlinear activation unit, perform nonlinear activation operation unit by unit, generate the current reserve pool activation state, and write the current reserve pool activation state into the reserve state storage unit. S26. Perform matrix multiplication on the current reservoir activation state and the output weight matrix to generate the compressed dynamic feature representation for the current moment.

[0009] Optionally, S3 specifically includes: S31. Set up a gate activation vector generation unit in the event-sensitive trigger gate module, and input the control state information to the gate activation vector generation unit in chronological order. S32. Perform numerical rate of change calculation on the control state information, generate a rate of change vector, and input the rate of change vector into the gated function unit; S33. In the gated function unit, a gated activation vector is generated based on the rate of change vector. A threshold comparison operation is performed on each value in the rate of change vector. When the value is greater than the preset threshold, a first activation value is written into the gated activation vector. When the value is not greater than the preset threshold, a second activation value is written into the gated activation vector. The first activation value is greater than the second activation value. S34. Perform element-wise multiplication on the gating activation vector and the reservoir activation state generated by the mutual inhibition reservoir echo state network module to generate the gating filtering state. S35. Perform a timing combination operation on the gating and filtering states to combine the gating and filtering states of consecutive time steps into an event response feature vector.

[0010] Optionally, S4 specifically includes: S41. In the reverse memory-driven DRQN ​​model module, set up a forward recursive unit, input the event response feature vector into the forward recursive unit in chronological order, perform forward recursive operation on the event response feature vector step by step, and generate a forward hidden state sequence with fixed dimension. S42. In the reverse memory-driven DRQN ​​model module, set up a reverse recursion unit to extract a continuous hidden state of a preset length from the forward hidden state sequence. Input the extracted hidden state sequence into the reverse recursion unit in reverse time order and perform reverse recursion operation on the hidden state sequence to generate a reverse hidden state sequence with the same vector dimension as the forward hidden state sequence. S43. Perform element-wise subtraction on the current time step hidden state in the forward hidden state sequence and the corresponding hidden state in the reverse hidden state sequence to generate a difference vector with the same dimension as the hidden state vector. Input the difference vector into the displacement mapping unit to perform matrix multiplication and vector addition to generate a time displacement vector. Perform element-wise addition on the time displacement vector and the corresponding forward hidden state to form a reverse memory traction state. S44. Input the reverse memory traction state into the Q-value estimation unit, perform matrix multiplication and nonlinear activation operations on the reverse memory traction state, and generate the Q-value vector of the current time step. S45. Within the policy output unit, select the action index with the largest value from the Q-value vector and use the action index as the policy output for the current time step.

[0011] Optionally, S5 specifically includes: S51. Input the event response feature vector and action state pair into the reverse memory driven DRQN ​​model module in chronological order. Perform forward recursive operation step by step in the forward recursive unit and reverse recursive operation in the reverse recursive unit to generate the reverse memory traction state for each time step. S52. Input the reverse memory traction state into the Q-value estimation unit, perform matrix multiplication and nonlinear activation operations on the reverse memory traction state, and generate the Q-value vector for the corresponding time step. S53. Perform a difference operation on the Q value and reward value at the corresponding position of the action state in the Q value vector to form a time-series difference error signal, and update the parameters of the reverse memory-driven DRQN ​​model module.

[0012] Optionally, S6 specifically includes: S61. During the training phase, state-action-reward-next state samples generated in chronological order are written into the priority experience replay queue. S62. Calculate the absolute value of the timing difference error signal for each sample in the priority experience playback queue, and use it as the priority of the sample; S63. Select a preset number of samples from the priority experience replay queue according to priority to form a training sample set; S64. Perform a truncation operation of a preset length on the forward hidden state sequence corresponding to each sample in the training sample set to obtain continuous hidden state segments. S65. Utilize the captured hidden state fragments to perform reverse recursion and reverse memory driving operations to obtain the reverse memory traction state; S66. Input the reverse memory traction state into the training step, generate the training error signal, and use the training error signal to update the parameters of the reverse memory-driven DRQN ​​model.

[0013] Optionally, S7 specifically includes: S71. After each round of training, calculate the average value of the absolute value of the training error signal to obtain the average error value. S72. When the average error value of 10 consecutive training rounds is not greater than 0.01, stop the parameter update operation and determine that the reverse memory-driven DRQN ​​model has converged. S73. After the convergence condition is met, output the trained model parameters and generate a multi-channel vibration recognition strategy model.

[0014] The beneficial effects of this invention are: (1) By applying the mutual inhibition reservoir echo state network to the dynamic feature extraction of multi-channel vibration signals, the reservoir state can be kept stable under high noise, weak events and multi-channel mutual interference conditions, thereby improving the distinguishability of vibration events and providing a reliable timing basis for precise triggering.

[0015] (2) By setting an event-sensitive trigger gate mechanism, key change segments in continuous time sequence are filtered, which realizes the accurate capture of weak events and mutation points, reduces the probability of false triggering and missed triggering, makes the triggering time closer to the starting point of the real event, and improves the triggering accuracy.

[0016] (3) A decision model is constructed by a deep recurrent Q-network driven by reverse memory, and the selection method of priority experience playback and reverse recursion sample segments is combined to enable the triggering strategy to utilize the forward and reverse time correlation information at the same time, thereby improving the training convergence speed and strategy stability, and realizing high-precision synchronous triggering and high-reliability decision output of multi-channel vibration signals. Attached Figure Description

[0017] 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: Figure 1 This is a flowchart of a multi-channel vibration data acquisition system with precise triggering and synchronization proposed in this invention; Figure 2 The flowchart below shows the event-sensitive trigger gate of a multi-channel vibration data acquisition system for precise triggering synchronization proposed in this invention. Figure 3 This is a multi-stream synthesis structure diagram of a multi-channel vibration data acquisition system with precise triggering and synchronization proposed in this invention. Detailed Implementation

[0018] 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.

[0019] refer to Figure 1-3 A multi-channel vibration data acquisition system with precise triggering synchronization includes: The data acquisition and preprocessing module is used to read the timing signals of the multi-channel vibration sensor and the corresponding control state information, and to perform preprocessing on the timing signals to generate vibration state samples and action state pairs. The mutual inhibition reservoir echo state network module is used to perform input mapping, mutual inhibition reservoir state update and dynamic feature extraction based on the vibration state samples generated by the data acquisition and preprocessing module, and output compressed dynamic feature representation. The event-sensitive trigger gate module is used to generate a gated activation vector based on the control state information, and to perform time-series filtering on the intermediate states output by the mutual inhibition reservoir echo state network module, and output an event response feature vector. The reverse memory-driven DRQN ​​model module is used to receive the event response feature vector output by the event-sensitive trigger gate module, perform bidirectional time-aware Q-value estimation, and output policy actions. The error training module is used to calculate the Q-value error based on the policy action output by the reverse memory-driven DRQN ​​model module and generate training signals. The experience playback module is used to construct a priority experience playback queue based on the Q-value error, select training segments based on the reverse memory driving rule, update the parameters of the DRQN ​​model, and form a multi-channel vibration recognition strategy model.

[0020] In this embodiment, the modules are interconnected using the following method: S1. Acquire timing signals from multi-channel vibration sensors and corresponding control status information, perform preprocessing, and construct vibration status samples and action status pairs. S2. Construct an echo state network with a mutually inhibited reserve pool structure, initialize the input weight matrix, the mutually inhibited connection matrix, the output weight matrix and the state update parameters, perform input mapping and reserve state evolution on the vibration state samples, and extract compressed dynamic feature representations. S3. Construct an event-sensitive trigger gate mechanism, generate a gated activation vector based on the control state information, perform temporal filtering on the intermediate states of the echo state network, and extract event response feature vectors. S4. Input the event response feature vector into the DRQN ​​model with reverse memory driving mechanism, construct a bidirectional time-aware Q-value estimation structure, and initialize the Q-value function and policy output layer. S5. Input training is performed on the action-state pairs, forward inference and backpropagation of error are executed, and the Q-value error of the policy output is calculated. S6. Construct a priority experience replay queue based on Q-value error, select training segments using reverse memory-driven rules, and update the parameters of the DRQN ​​model; S7. Repeat steps S2 to S6 until the Q-value function converges, and output the multi-channel vibration recognition strategy model.

[0021] In this embodiment, S1 specifically includes: S11. Read the timing signals and corresponding control status information of the multi-channel vibration sensor through the data acquisition device, convert the analog vibration signals of each channel into digital sequences according to the set sampling frequency, and generate multi-channel digital vibration sequences according to the channel number. S12. Normalize the multi-channel digital vibration sequence, linearly map the vibration amplitude of each channel according to the numerical range of 0 to 1, and divide the normalized vibration sequence into windows according to the set sliding window length and sliding step size to generate a vibration segment sequence with continuous time order. S13. Perform a label alignment operation based on the timestamps of the control state information and the vibration segment sequence, pair the control state information according to the same time order as the vibration segment, and generate vibration state samples and action state pairs containing the vibration segment and the matching control state.

[0022] In this embodiment, S2 specifically includes: S21. Establish a mutually inhibited reservoir echo state network, set the input weight matrix, the mutual inhibition connection matrix, the output weight matrix and the state update parameters, and send the vibration state samples into the mutually inhibited reservoir echo state network in chronological order. In this embodiment, the initialization process in S21 specifically includes: When establishing the echo state network of the mutual inhibition reservoir, the dimension of the input weight matrix is ​​determined, and the feature dimension of the vibration state sample and the number of reservoir units are mapped to the number of rows and columns of the matrix. A pseudo-random number generation operation is performed to generate values ​​for all elements of the input weight matrix, and the generated matrix is ​​written into the input mapping unit. The mutual inhibition connection matrix is ​​defined by determining its dimension and setting the number of reservoir units to the number of rows and columns of the matrix. Pseudo-random number generation is performed on the off-diagonal elements of the mutual inhibition connection matrix to generate connection coefficients between reservoir units. The diagonal elements of the matrix are set to zero to form a mutual inhibition structure. The generated mutual inhibition connection matrix is ​​then written into the mutual inhibition processing unit. Output a weight matrix, determine the dimension of the output weight matrix, set the number of reservoir units to the number of rows in the matrix, and set the output dimension of the compressed dynamic feature representation to the number of columns in the matrix; perform a pseudo-random number generation operation to generate values ​​for all elements of the output weight matrix, and write the generated matrix into the output projection unit; The state update parameters include the leakage rate, bias vector, and activation function parameters. The leakage rate is written to the state update unit, the bias vector is written to the bias register unit through pseudo-random number generation, and the activation function parameters are written to the nonlinear activation unit according to the corresponding activation function format. The leakage rate is 0.2, and the activation function is set to the hyperbolic tangent function.

[0023] S22. Perform matrix multiplication on the vibration state sample at each moment, generate an input mapping vector using the input weight matrix, and send the input mapping vector into the reserve pool state update unit. In this embodiment, the reserve pool status update unit in S22 specifically includes: The reserve pool state update unit receives the input mapping vector and the previous reserve pool activation state at each time step, and performs matrix operations to generate the unsuppressed state update quantity. The reserve pool state update unit consists of an input mapping receiving subunit, a state reading subunit, and a linear combination calculation subunit. The input mapping receiving subunit receives the input mapping vector generated by the input weight matrix. The state reading subunit reads the previous reserve pool activation state from the reserve state storage unit. The linear combination calculation subunit combines the input mapping vector and the previous reserve pool activation state by performing matrix multiplication and vector addition operations to generate the unsuppressed state update quantity for mutual suppression processing, and outputs the unsuppressed state update quantity to the mutual suppression processing unit.

[0024] S23. Read the previous state from the reserve state storage unit, perform matrix multiplication and vector addition operations, and combine the previous state with the input mapping vector to form the unsuppressed state update quantity. S24. Perform a matrix multiplication operation on the unsuppressed state update quantity using the mutual inhibition connection matrix to generate the mutual inhibition adjustment quantity, and perform a vector subtraction operation on the mutual inhibition adjustment quantity and the unsuppressed state update quantity to form the mutual inhibition processing intermediate state. In this embodiment, the principle of mutual inhibition in S24 specifically includes: The mutual inhibition structure sets a fixed inhibition coefficient between reservoir units, ensuring that each reservoir unit receives inhibition inputs from other reservoir units simultaneously when generating state updates. Each off-diagonal element in the mutual inhibition connection matrix characterizes the mutual inhibition strength between two reservoir units; when two reservoir units have a mutual inhibition relationship, the relationship is reflected by the value of the off-diagonal element. When a reservoir unit generates its current state input, it reduces its update amount according to the inhibition coefficient at the corresponding position in the mutual inhibition connection matrix, making the reservoir unit directly constrained by other reservoir units during state updates. In this way, the state update of any reservoir unit is carried out within the range of its inhibition relationship with other reservoir units, thus forming a state constraint mechanism based on a fixed inhibition coefficient between units within the reservoir.

[0025] S25. Input the intermediate state of mutual inhibition processing into the nonlinear activation unit, perform nonlinear activation operation unit by unit to generate the current reserve pool activation state, and write the current reserve pool activation state into the reserve state storage unit; in this embodiment, the nonlinear activation unit adopts the hyperbolic tangent function. S26. Perform matrix multiplication on the current reservoir activation state and the output weight matrix to generate the compressed dynamic feature representation for the current moment.

[0026] In this embodiment, S3 specifically includes: S31. A gating activation vector generation unit is set up in the event-sensitive trigger gate module. The control status information is input to the gating activation vector generation unit in chronological order. The control status information consists of control quantities recorded during the operation of the acquisition device, including acquisition command identifier, sampling time identifier, and external control input quantity. The acquisition command identifier is used to characterize the status identifier of the current acquisition process in the start-up phase, stable phase, or stop phase. The sampling time identifier is generated by the acquisition device and is used to characterize the sampling time corresponding to the vibration state sample. The external control input quantity is provided by the vibration acquisition environment and is used to record the operating speed, loading status, or change in operating conditions of the device during the acquisition process. The control status information is input to the event-sensitive trigger gate module in the form of a vector arranged in chronological order to perform numerical change rate calculation and gating activation vector generation operations. S32. Perform numerical rate of change calculation on the control state information, generate a rate of change vector, and input the rate of change vector into the gated function unit; the numerical rate of change calculation is completed by performing a difference operation on the control state information between adjacent time steps; S33. In the gated function unit, a gated activation vector is generated based on the rate of change vector. A threshold comparison operation is performed on each value in the rate of change vector. When the value is greater than the preset threshold, a first activation value is written into the gated activation vector. When the value is not greater than the preset threshold, a second activation value is written into the gated activation vector. The first activation value is greater than the second activation value. The preset threshold is 0.1, the first activation value is 1.0, and the second activation value is 0.1. S34. Perform element-wise multiplication on the gating activation vector and the reservoir activation state generated by the mutual inhibition reservoir echo state network module to generate the gating filtering state. S35. Perform a timing combination operation on the gated filtering states to combine the gated filtering states of consecutive time steps into an event response feature vector. The timing combination operation is completed by writing the gated filtering states sequentially within consecutive time steps; the gated filtering states generated in the current time step are written to the timing buffer unit in chronological order, and the gated filtering states of multiple adjacent time steps are sequentially concatenated in the vector dimension to form a fixed-length event response feature vector.

[0027] In this embodiment, S4 specifically includes: S41. In the reverse memory-driven DRQN ​​model module, a forward recursive unit is set up. The event response feature vector is input into the forward recursive unit in chronological order. The forward recursive operation is performed on the event response feature vector step by step to generate a forward hidden state sequence with fixed dimensions. The forward recursive unit adopts a recurrent neural network structure. The input is the event response feature vector sequence and the output is the hidden state sequence corresponding to the input dimension. S42. In the reverse memory-driven DRQN ​​model module, a reverse recursion unit is set up to extract a set length of continuous hidden states from the forward hidden state sequence. The extracted hidden state sequence is input into the reverse recursion unit in reverse time order, and the reverse recursion operation is performed on the hidden state sequence to generate a reverse hidden state sequence with the same vector dimension as the forward hidden state sequence. The reverse recursion unit also adopts a recurrent neural network structure. The input is the hidden state sequence arranged in reverse time order, and the output is the reverse hidden state sequence with the same vector dimension as the input sequence. The preset length is written into the reverse recursion unit by the external configuration unit and takes the value of a positive integer. The preset length is selected from multiple candidate values ​​through parameter search before the model training starts. The selected value is recorded in the parameter configuration file. After receiving the forward hidden state sequence, the preset length of hidden states is extracted from the end of the sequence and arranged in reverse time order. During the training phase, all hidden states within the preset length participate in gradient backpropagation, and hidden states outside the preset length continue to participate in the time-step recursive calculation and gradient backpropagation of the forward recursion unit. S43. Perform element-wise subtraction on the current time step hidden state in the forward hidden state sequence and the corresponding hidden state in the reverse hidden state sequence to generate a difference vector with the same dimension as the hidden state vector. Input the difference vector into the displacement mapping unit to perform matrix multiplication and vector addition to generate a time displacement vector. Perform element-wise addition on the time displacement vector and the corresponding forward hidden state to form a reverse memory traction state. S44. Input the reverse memory traction state into the Q-value estimation unit, perform matrix multiplication and nonlinear activation operations on the reverse memory traction state to generate the Q-value vector of the current time step; the Q-value estimation unit is used to generate the Q-value vector of the current time step based on the reverse memory traction state; the Q-value estimation unit sets a Q-value weight matrix and a Q-value bias vector, the number of rows of the Q-value weight matrix is ​​equal to the number of actions, the number of columns is equal to the vector dimension of the reverse memory traction state, and the length of the Q-value bias vector is equal to the number of actions; the nonlinear activation operation performs a maximum value selection operation on each element in the intermediate vector, compares each element with the zero value, and takes the larger value as the output value of the current element to form the vector after nonlinear activation; during the training phase, the state-action-reward-next state sample sequence is constructed using the event response feature vector, action index, and reward value generated based on the trigger time error, synchronization deviation, and trigger judgment, and the weights of the Q-value estimation unit are iteratively updated using a temporal difference update method; In this embodiment, the reward function in S44 specifically includes: The absolute value of the time difference between the action output at the current time step and the actual trigger moment is calculated. The result is used as the trigger moment error signal, and a linear mapping operation is performed on the trigger moment error signal according to a preset threshold range to generate a trigger reward component. The absolute value of the synchronization time deviation of multi-channel data after action execution is calculated, and the obtained deviation value is multiplied by a preset synchronization deviation penalty coefficient to generate a synchronization penalty component. A trigger determination signal is generated based on whether the action at the current time step is consistent with the actual trigger area. A penalty value is generated when the action is triggered in a non-trigger behavior space, and a penalty value is generated when the action is not triggered in the actual trigger behavior space. The trigger reward component, the synchronization penalty component, and the trigger determination signal are then combined. The signal undergoes addition to form a scalar reward value, which is recorded in the state-action-reward-next state sample sequence. The trigger time error threshold is set to 3 sampling points. When the absolute value of the trigger time error is not greater than 3, the error is used as the input value and multiplied by the linear mapping coefficient 0.5 to form the trigger reward component. When the absolute value of the trigger time error is greater than 3, the linear mapping output value is fixed at −1. The synchronization deviation penalty coefficient is set to 0.2. After calculating the absolute value of the multi-channel synchronization deviation, the absolute value is multiplied by the synchronization deviation penalty coefficient to form the synchronization penalty component. The determination of false triggering and missed triggering is based on the real event interval marking: false triggering -2, missed triggering -3.

[0028] S45. Within the policy output unit, select the action index with the largest value from the Q-value vector and use the action index as the policy output for the current time step.

[0029] In this embodiment, S5 specifically includes: S51. Input the event response feature vector and action state pair into the reverse memory driven DRQN ​​model module in chronological order. Perform forward recursive operation step by step in the forward recursive unit and reverse recursive operation in the reverse recursive unit to generate the reverse memory traction state for each time step. S52. Input the reverse memory traction state into the Q-value estimation unit, perform matrix multiplication and nonlinear activation operations on the reverse memory traction state, and generate the Q-value vector for the corresponding time step. S53. Perform a difference operation on the Q value and reward value at the corresponding position of the action state in the Q value vector to form a time-series difference error signal, and update the parameters of the reverse memory-driven DRQN ​​model module.

[0030] In this embodiment, S6 specifically includes: S61. During the training phase, state-action-reward-next state samples generated in chronological order are written into the priority experience replay queue. S62. Calculate the absolute value of the timing difference error signal for each sample in the priority experience playback queue, and use it as the priority of the sample; S63. Select a preset number of samples from the priority experience replay queue according to priority to form a training sample set; in this embodiment, the preset number is 32. S64. Perform a truncation operation of a preset length on the forward hidden state sequence corresponding to each sample in the training sample set to obtain continuous hidden state segments. S65. Utilize the captured hidden state fragments to perform reverse recursion and reverse memory driving operations to obtain the reverse memory traction state; S66. Input the reverse memory traction state into the training step, generate the training error signal, and use the training error signal to update the parameters of the reverse memory-driven DRQN ​​model.

[0031] In this embodiment, the training error signal in S66 specifically includes: During the training phase, the reverse memory traction state is input into the Q-value calculation step to generate the training Q-value; for each sample in the training sample set, the reward value is read, and the reverse memory traction state of the next time step is input into the Q-value calculation step to obtain the Q-value vector of the next time step. The maximum value in the Q-value vector is selected as the maximum Q-value of the next time step; the target Q-value is generated according to the following formula: Target Q value = Reward value + 0.9 × Maximum Q value at the next time step; Where 0.9 is a fixed discount factor; after obtaining the target Q value, the difference between the training Q value and the target Q value is calculated: Training error signal = training Q-value - target Q-value; During the parameter update process, the parameters of the reverse memory-driven DRQN ​​model are updated by writing the training error signal into the parameter update.

[0032] In this embodiment, S7 specifically includes: S71. After each round of training, calculate the average value of the absolute value of the training error signal to obtain the average error value. S72. When the average error value of 10 consecutive training rounds is not greater than 0.01, stop the parameter update operation and determine that the reverse memory-driven DRQN ​​model has converged. S73. After the convergence condition is met, output the trained model parameters and generate a multi-channel vibration recognition strategy model.

[0033] Example 1: To verify the feasibility of this invention in practice, it was applied to a structural vibration monitoring system with multiple vibration acquisition channels. Under these conditions, the vibration signals from different channels exhibit amplitude differences, propagation delays, operating noise interference, and intermittent impacts. The original vibration time-series signals from each channel were read and preprocessed. Continuous vibration state samples were generated through normalization, segmentation, and noise filtering. Control state information was then read to construct action state pairs. A cross-inhibition reservoir echo state network was used to extract features from the processed vibration samples. The cross-inhibition structure limits the cross-channel amplification effect of the reservoir's internal states, making the dynamic feature representation more stable and preventing over-response even under high noise levels. After dynamic feature generation, the control state was used as the event-driven basis. Event-sensitive trigger gates were used to filter out time segments with changing trends and remove meaningless stationary segments, improving the accuracy of event extraction.

[0034] After obtaining the event response feature vector, it is input into the reverse memory-driven DRQN ​​model for decision-making reasoning. The model retains the forward recursive structure while incorporating backward recursive fragments, allowing the hidden state to simultaneously consider local information from the past and future, reducing decision lag and short-sighted behavior. During the training phase, the state-action-reward-next state data is recorded through an experience replay mechanism, and a priority ranking method based on temporal difference error is adopted to make training more focused on critical moments. Hidden state fragments for backward recursion are constructed through a preset length truncation operation, ultimately forming a reverse memory-driven state used to update the Q-value estimate, allowing training to make fuller use of temporal correlations. The training loop uses a condition that the average error of ten consecutive rounds is less than 0.01 as a convergence criterion to ensure the stability of the model's output actions.

[0035] Table 1: Comparison of Multi-channel Vibration Acquisition Data

[0036] In actual data acquisition, the model of this invention was used to determine the trigger moment in real time and maintain synchronization across multiple channels. To compare with traditional methods, multiple vibration sequences were continuously recorded for the same mechanical structure under identical operating conditions, and the triggering errors and missed triggers of the traditional method and this invention were quantitatively tested. As shown in Table 1, this invention has advantages in scenarios with strong noise and weak events. In "Sequence A," the average triggering error of the traditional method was 4.8 milliseconds, while the error of this invention was only 1.2 milliseconds. Similarly, in "Sequences B" to "Sequence D," the traditional error generally exceeded 5 milliseconds, while this invention remained stably between 1.0 and 1.6 milliseconds. Regarding the number of missed triggers, traditional acquisition methods generally resulted in 3 to 6 missed triggers, while this invention had no missed triggers in most cases, even in sequences with complex vibration patterns, where only one missed trigger occurred. Overall, this invention reduces the triggering error by more than 70% and the missed trigger rate by more than 80% compared to traditional methods, improving the ability to extract and synchronously acquire vibration events.

[0037] The mutual inhibition reservoir structure of this invention ensures stable dynamic characteristics even under high-frequency noise interference. The event-sensitive trigger gate improves the ability to capture weak events, while the reverse memory-driven DRQN ​​model enhances policy learning efficiency and decision reliability. This enables the entire triggering system to more accurately predict event occurrence times and ensure multi-channel synchronization. This embodiment fully demonstrates that this invention not only improves triggering accuracy but also enhances the system's robustness in complex vibration environments, possessing strong engineering applicability and broad promotional value.

[0038] 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 multi-channel vibration data acquisition system with precise triggering synchronization, characterized in that, include: The data acquisition and preprocessing module is used to read the timing signals of the multi-channel vibration sensor and the corresponding control state information, and to perform preprocessing on the timing signals to generate vibration state samples and action state pairs. The mutual inhibition reservoir echo state network module is used to perform input mapping, mutual inhibition reservoir state update and dynamic feature extraction based on the vibration state samples generated by the data acquisition and preprocessing module, and output compressed dynamic feature representation. The event-sensitive trigger gate module is used to generate a gated activation vector based on the control state information, and to perform time-series filtering on the intermediate states output by the mutual inhibition reservoir echo state network module, and output an event response feature vector. The reverse memory-driven DRQN ​​model module is used to receive the event response feature vector output by the event-sensitive trigger gate module, perform bidirectional time-aware Q-value estimation, and output policy actions. The error training module is used to calculate the Q-value error based on the policy action output by the reverse memory-driven DRQN ​​model module and generate training signals. The experience playback module is used to construct a priority experience playback queue based on the Q-value error, select training segments based on the reverse memory driving rule, update the parameters of the DRQN ​​model, and form a multi-channel vibration recognition strategy model.

2. The multi-channel vibration data acquisition system with precise triggering synchronization according to claim 1, characterized in that, The modules are connected in the following way: S1. Acquire timing signals from multi-channel vibration sensors and corresponding control status information, perform preprocessing, and construct vibration status samples and action status pairs. S2. Construct an echo state network with a mutually inhibited reserve pool structure, initialize the input weight matrix, the mutually inhibited connection matrix, the output weight matrix and the state update parameters, perform input mapping and reserve state evolution on the vibration state samples, and extract compressed dynamic feature representations. S3. Construct an event-sensitive trigger gate mechanism, generate a gated activation vector based on the control state information, perform temporal filtering on the intermediate states of the echo state network, and extract event response feature vectors. S4. Input the event response feature vector into the DRQN ​​model with reverse memory driving mechanism, construct a bidirectional time-aware Q-value estimation structure, and initialize the Q-value function and policy output layer. S5. Input training is performed on the action-state pairs, forward inference and backpropagation of error are executed, and the Q-value error of the policy output is calculated. S6. Construct a priority experience replay queue based on Q-value error, select training segments using reverse memory-driven rules, and update the parameters of the DRQN ​​model; S7. Repeat steps S2 to S6 until the Q-value function converges, and output the multi-channel vibration recognition strategy model.

3. The multi-channel vibration data acquisition system with precise triggering synchronization according to claim 2, characterized in that, S2 specifically includes: S21. Establish a mutually inhibited reservoir echo state network, set the input weight matrix, the mutual inhibition connection matrix, the output weight matrix and the state update parameters, and send the vibration state samples into the mutually inhibited reservoir echo state network in chronological order. S22. Perform matrix multiplication on the vibration state sample at each moment, generate an input mapping vector using the input weight matrix, and send the input mapping vector into the reserve pool state update unit. S23. Read the previous state from the reserve state storage unit, perform matrix multiplication and vector addition operations, and combine the previous state with the input mapping vector to form the unsuppressed state update quantity. S24. Perform a matrix multiplication operation on the unsuppressed state update quantity using the mutual inhibition connection matrix to generate the mutual inhibition adjustment quantity, and perform a vector subtraction operation on the mutual inhibition adjustment quantity and the unsuppressed state update quantity to form the mutual inhibition processing intermediate state. S25. Input the intermediate state of mutual inhibition processing into the nonlinear activation unit, perform nonlinear activation operation unit by unit, generate the current reserve pool activation state, and write the current reserve pool activation state into the reserve state storage unit. S26. Perform matrix multiplication on the current reservoir activation state and the output weight matrix to generate the compressed dynamic feature representation for the current moment.

4. The multi-channel vibration data acquisition system with precise triggering synchronization according to claim 3, characterized in that, S3 specifically includes: S31. Set up a gate activation vector generation unit in the event-sensitive trigger gate module, and input the control state information to the gate activation vector generation unit in chronological order. S32. Perform numerical rate of change calculation on the control state information, generate a rate of change vector, and input the rate of change vector into the gated function unit; S33. In the gated function unit, a gated activation vector is generated based on the rate of change vector. A threshold comparison operation is performed on each value in the rate of change vector. When the value is greater than the preset threshold, a first activation value is written into the gated activation vector. When the value is not greater than the preset threshold, a second activation value is written into the gated activation vector. The first activation value is greater than the second activation value. S34. Perform element-wise multiplication on the gating activation vector and the reservoir activation state generated by the mutual inhibition reservoir echo state network module to generate the gating filtering state. S35. Perform a timing combination operation on the gating and filtering states to combine the gating and filtering states of consecutive time steps into an event response feature vector.

5. A multi-channel vibration data acquisition system for precise triggering synchronization according to claim 4, characterized in that, S4 specifically includes: S41. In the reverse memory-driven DRQN ​​model module, set up a forward recursive unit, input the event response feature vector into the forward recursive unit in chronological order, perform forward recursive operation on the event response feature vector step by step, and generate a forward hidden state sequence with fixed dimension. S42. In the reverse memory-driven DRQN ​​model module, set up a reverse recursion unit to extract a continuous hidden state of a preset length from the forward hidden state sequence. Input the extracted hidden state sequence into the reverse recursion unit in reverse time order and perform reverse recursion operation on the hidden state sequence to generate a reverse hidden state sequence with the same vector dimension as the forward hidden state sequence. S43. Perform element-wise subtraction on the current time step hidden state in the forward hidden state sequence and the corresponding hidden state in the reverse hidden state sequence to generate a difference vector with the same dimension as the hidden state vector. Input the difference vector into the displacement mapping unit to perform matrix multiplication and vector addition to generate a time displacement vector. Perform element-wise addition on the time displacement vector and the corresponding forward hidden state to form a reverse memory traction state. S44. Input the reverse memory traction state into the Q-value estimation unit, perform matrix multiplication and nonlinear activation operations on the reverse memory traction state, and generate the Q-value vector of the current time step. S45. Within the policy output unit, select the action index with the largest value from the Q-value vector and use the action index as the policy output for the current time step.

6. The multi-channel vibration data acquisition system with precise triggering synchronization according to claim 5, characterized in that, S5 specifically includes: S51. Input the event response feature vector and action state pair into the reverse memory driven DRQN ​​model module in chronological order. Perform forward recursive operation step by step in the forward recursive unit and reverse recursive operation in the reverse recursive unit to generate the reverse memory traction state for each time step. S52. Input the reverse memory traction state into the Q-value estimation unit, perform matrix multiplication and nonlinear activation operations on the reverse memory traction state, and generate the Q-value vector for the corresponding time step. S53. Perform a difference operation on the Q value and reward value at the corresponding position of the action state in the Q value vector to form a time-series difference error signal, and update the parameters of the reverse memory-driven DRQN ​​model module.

7. A multi-channel vibration data acquisition system for precise triggering synchronization according to claim 6, characterized in that, S6 specifically includes: S61. During the training phase, state-action-reward-next state samples generated in chronological order are written into the priority experience replay queue. S62. Calculate the absolute value of the timing difference error signal for each sample in the priority experience playback queue, and use it as the priority of the sample; S63. Select a preset number of samples from the priority experience replay queue according to priority to form a training sample set; S64. Perform a truncation operation of a preset length on the forward hidden state sequence corresponding to each sample in the training sample set to obtain continuous hidden state segments. S65. Utilize the captured hidden state fragments to perform reverse recursion and reverse memory driving operations to obtain the reverse memory traction state; S66. Input the reverse memory traction state into the training step, generate the training error signal, and use the training error signal to update the parameters of the reverse memory-driven DRQN ​​model.

8. A multi-channel vibration data acquisition system with precise triggering synchronization according to claim 7, characterized in that, Specifically, S7 includes: S71. After each round of training, calculate the average value of the absolute value of the training error signal to obtain the average error value. S72. When the average error value of 10 consecutive training rounds is not greater than 0.01, stop the parameter update operation and determine that the reverse memory-driven DRQN ​​model has converged. S73. After the convergence condition is met, output the trained model parameters and generate a multi-channel vibration recognition strategy model.

Images