A method for predicting phase duration of an induction signal based on a TSMixer
By using a TSMixer-based method for predicting the phase duration of inductive signals, and employing a multi-branch model to process the phase state and conflict relationships at intersections in groups, and dynamically masking the output of the sub-models, the method addresses the limitations in the versatility and accuracy of inductive signal phase duration prediction, and achieves efficient prediction at different intersections.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TONGJI UNIV
- Filing Date
- 2025-06-30
- Publication Date
- 2026-07-21
AI Technical Summary
Existing technologies struggle to accurately characterize the impact of traffic flow state changes on the duration of sensing phases and lack universality across different intersections, resulting in low accuracy in predicting the duration of sensing signal phases.
A TSMixer-based method for predicting the phase duration of inductive signals is adopted. By collecting historical data from intersections, a time-series dataset is constructed. The phase state and conflict relationship are processed in groups using a multi-branch TSMixer model. The sub-model outputs are dynamically masked and fused to finally generate the prediction of the remaining phase duration.
It improves the versatility and accuracy of phase duration prediction of induction signals at different intersections, has the ability to make multiple predictions, and solves the problems of low accuracy and insufficient applicability of traditional methods.
Smart Images
Figure CN120726830B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of transportation technology, and in particular to a method for predicting the duration of sensing phases based on TSMixer. Background Technology
[0002] Currently, traffic lights are an important infrastructure for traffic management departments to regulate vehicle traffic order and optimize road traffic flow. Their phase timing information has a significant impact on driver decision-making, intersection operating efficiency, and traffic safety.
[0003] Traditional phase timing data typically relies on fixed timing schemes provided by traffic management departments, but this approach has limited coverage and struggles to meet the dynamic demands of complex traffic environments. Although some technical methods exist for mining fixed phase durations, these methods mostly employ statistical models or time series analysis, failing to capture the complex spatial characteristics and temporal correlations between variable phase durations and traffic flow data, and thus unable to meet the prediction requirements for sensor signal phase durations.
[0004] Furthermore, some existing technologies employ deep learning methods to predict the phase duration of sensed signals, but due to the lack of universality in their model structures, they are difficult to adapt to different intersection geometries and phase designs. Therefore, there is an urgent need to propose a universal real-time sensed phase duration prediction method to improve its applicability and prediction accuracy at different intersections. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of the prior art, which is difficult to accurately characterize the impact of traffic flow state changes on the duration of sensing phase and is difficult to apply to different inductive signal controlled intersections, and to provide an inductive signal phase duration prediction method based on TSMixer.
[0006] The objective of this invention can be achieved through the following technical solutions:
[0007] A method for predicting the phase duration of an inductive signal based on TSMixer includes the following steps:
[0008] Collect historical phase status and induction coil detection data of the inductively controlled intersection, and generate a phase switching timetable based on the historical phase switching times;
[0009] Using the green light phase switching moment as the feature sampling moment, and the current phase start moment to the feature sampling moment as the feature calculation window, the phase and induction coil data of consecutive time steps within the feature calculation step window are extracted to construct a time series dataset;
[0010] Based on the phase state and phase conflict relationship at the sampling time, the intersection phases are dynamically divided into conflict relationship groups and phase state groups, and dynamic mask representations are generated for each phase group.
[0011] A phase duration prediction model based on multi-branch TSMixer is established, in which conflict relationship group and phase state group correspond to independent sub-models. Each sub-model processes time series datasets in parallel and outputs prediction results. The outputs of each sub-model are fused through a dynamic masking mechanism to generate the final phase remaining duration prediction.
[0012] As a preferred technical solution, the status data of the sensor-controlled intersection is collected as follows:
[0013] Determine the phase of the induction signal intersection, the number of induction coils, and the correspondence between each induction coil and the phase;
[0014] The historical phase status of the signalized intersection is represented by the numbers 0 and 1, which indicate the red and green light phases respectively. The yellow light phase is considered as part of the red light phase.
[0015] The historical state of the induction coil at the intersection is represented by the numbers 0 and 1, indicating whether the induction coil is not activated or has been activated.
[0016] As a preferred technical solution, the timetable for switching the induction phase state is obtained and the calculation range of all induction phase features is defined, as follows:
[0017] Using the obtained historical sensing signal phase state data, a timetable for all phase state switching is generated based on the changes in digital 0 and 1;
[0018] Based on the phase switching timetable, determine the phase switching time when all phase states change from number 1 to number 0, and use this as the green light phase switching time.
[0019] Arrange the green and red light switching times of all phases in chronological order, remove duplicate values, and obtain the data feature sampling time sequence;
[0020] Based on the state switching timetable and sampling time sequence, the calculation interval for the characteristic data of each phase is obtained by taking the state start time of each phase at the characteristic sampling time as the starting point of the calculation interval and the characteristic sampling time as the ending point of the calculation interval.
[0021] As a preferred technical solution, the characteristic data of the sensing signal phase specifically includes: the duration of the phase that has passed within the data feature acquisition range; the traffic flow during the current phase state; the time elapsed since the last activation of the induction coil during the current phase state; whether queuing or congestion occurred during the current phase state; the phase state at the time of data acquisition; and whether the phase state has switched.
[0022] As a preferred technical solution, the time series dataset concatenates the multidimensional feature vectors of all phases in a preset phase number order to generate the joint feature vector of the current moment;
[0023] The joint feature vectors of the current time and the preceding N-1 consecutive time steps are stacked in time sequence, where N is not less than the number of phases of the sensing intersection.
[0024] As a preferred technical solution, the phase duration prediction predicts the remaining phase duration of the phase state group through a phase state module:
[0025] The phase state module includes two identical TSMixer regression networks, namely the red light phase network and the green light phase network, which are used to predict the duration of each phase state.
[0026] Based on the phase state characteristics, a dynamic mask for the predicted value is established. The output results of the red light phase network and the green light phase network of the phase state mask vector pair are used to mask the predicted values of the phase in the red light phase network and the phase in the green light phase network, respectively.
[0027] The outputs of the masked red and green light phase networks are added together to obtain the first predicted output vector.
[0028] As a preferred technical solution, the phase duration prediction predicts the remaining phase duration of conflict relationship groups through a conflict relationship module:
[0029] The conflict relationship module divides non-conflicting phases and conflicting phases according to the preset phase group of the intersection and the group to which the current green light switching phase belongs, and constructs two TSMixer regression networks with the same structure, as non-conflicting category network and conflicting category network respectively, to predict the duration of each phase state.
[0030] A group mask is established based on the group information of each phase. The group mask is used to mask the prediction results of the non-conflict class network and the conflict class network, so that the prediction values in the output of the non-conflict class network and the conflict class network only retain the phases belonging to that class.
[0031] The outputs of the masked non-conflict class network and the conflict class network are added together to obtain the second predicted output vector.
[0032] As a preferred technical solution, the structure of the TSMixer regression network is as follows:
[0033] The input is an N×6A-dimensional temporal feature tensor, which is passed sequentially through N mixing layers, a temporal projection layer, and a feature projection layer corresponding to the time window, and outputs the prediction result of the remaining phase duration; wherein, the mixing layer includes a temporal mixing MLP and a feature mixing MLP connected in sequence.
[0034] The temporal hybrid MLP layer normalizes the input temporal features in two dimensions and then transposes them. It then inputs the transpose into a single-layer temporal MLP to model the temporal pattern of the time series. The output of the temporal MLP is transposed again and then residually linked with the two-dimensional normalized input data to obtain the output of the temporal hybrid MLP layer.
[0035] The feature mixing MLP layer is immediately followed by the time mixing MLP layer. After the output of the time mixing MLP layer is normalized in two dimensions, a feature MLP is input. The feature MLP uses two MLP layers to analyze the correlation between different features. The output of the feature MLP is residually connected with the output of the time mixing MLP layer to obtain the output of the feature mixing MLP layer.
[0036] The time projection layer is a fully connected layer applied in the time domain. It transposes the moments output by the hybrid layer and performs prediction through a fully connected layer, mapping the time series from the original input length to the target prediction length T. The prediction output of the fully connected layer is then recovered through transposition.
[0037] The feature projection layer is a fully connected layer applied on the feature domain. The matrix output from the time projection is transposed again and predicted through a fully connected layer, mapping the feature sequence from the original input length to the target prediction length, and outputting the prediction results of the remaining duration of all phases.
[0038] As a preferred technical solution, the phase duration prediction model concatenates the prediction output vectors obtained from the phase state module and the conflict relationship module and inputs them into a multilayer perceptron regression network to obtain the final prediction result of the phase state duration of each signal.
[0039] As a preferred technical solution, the multilayer perceptron regression network is as follows:
[0040] Input layer: The dimension of the input layer is determined by the number of features in the input data, with each feature corresponding to one input unit;
[0041] Hidden layers: The network contains several hidden layers, each of which includes a fully connected layer, a batch normalization layer, and a non-linear activation layer cascaded in sequence;
[0042] The first hidden layer receives the output of the input layer and maps it to a representation of 128 neurons;
[0043] The second hidden layer processes the output of the first hidden layer again and maps it into a structure of 256 neurons.
[0044] Output layer: Predicts the duration of each phase state.
[0045] Compared with the prior art, the present invention has the following beneficial effects:
[0046] 1) This invention collects historical phase state data and coil sensor state data of inductively controlled intersections and constructs a signal phase state switching timetable; it predicts the remaining duration of all phases at any time when the green light phase switches to the red light phase, and has the ability to make multiple predictions within a cycle.
[0047] 2) This invention proposes a novel modular learning architecture to construct a phase duration prediction model. It achieves specialized learning for sub-networks through phase conflict decoupling and dynamic grouping, and employs a cross-sub-network prediction collaborative correction mechanism. Using a feature time-series dataset containing the phase states of all signals at the intersection and traffic flow states as input, it predicts the remaining phase duration. This addresses the shortcomings of traditional inductive signal phase duration prediction techniques, such as low accuracy and lack of versatility. Attached Figure Description
[0048] Figure 1 This is a flowchart of the TSMixer-based induction signal phase duration prediction method of the present invention;
[0049] Figure 2 This is a schematic diagram of the layout of the inductive signal intersection in Embodiment 1 of the present invention;
[0050] Figure 3 This is a schematic diagram of the design of the calculation interval for the phase characteristic data of the inductive signal intersection in Embodiment 1 of the present invention;
[0051] Figure 4 This is a schematic diagram of the phase grouping of the inductive signal intersection in Embodiment 1 of the present invention;
[0052] Figure 5 This is a schematic diagram of the modular learning phase duration prediction model of the present invention;
[0053] Figure 6 This is a schematic diagram of the TSMixer regression network structure of the present invention;
[0054] Figure 7 This is a schematic diagram of the hybrid layer structure in the TSMixer regression network of this invention;
[0055] Figure 8 This is a schematic diagram of the temporal projection and feature projection layers in the TSMixer regression network of this invention. Detailed Implementation
[0056] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.
[0057] Example 1
[0058] This invention proposes a method for predicting the phase duration of induced signals based on TSMixer, such as... Figure 1 As shown, it mainly includes the following steps:
[0059] S1. Collect historical signal phase status data and coil sensor status data of the inductively controlled intersection, and integrate them to construct an initial dataset;
[0060] S11. Through on-site investigation or review of design documents, determine the number of induction coils and signal phases at the intersection, and clarify the correspondence between each induction coil and its phase; such as Figure 2 The diagram shown is a schematic of the layout of the induction signal intersection in this embodiment; all phases of the intersection are numbered using i, where i∈S,S={1,2,3,…,A}; all induction coils are numbered using j, where j∈D,D={1,2,3,…,B};
[0061] S12. Obtain the phase status data per second for all phases of the historical intersection, using the numbers 0 and 1 to represent the red and green light phases respectively. The yellow light phase is treated as part of the red light phase status. The status can then be represented as:
[0062]
[0063] Among them, s i (k,.) represents the phase state of phase i at time k.
[0064] S13. Obtain the status data of all induction coils at the historical intersection. Use the numbers 0 and 1 to represent the induction coil's unactivated and activated states, respectively. The status can then be represented as:
[0065]
[0066] Where, d j (k,.) represents the excitation state of induction coil j at time k.
[0067] S2. Based on the collected dataset, obtain a timetable of all phase state switching, and define the calculation range of all phase features according to the timetable;
[0068] S21. Using the historical phase state data obtained in step S1, generate a timetable for all phase state switching based on the changes in digits 0 and 1;
[0069] S22. Based on the phase state switching timetable, determine the phase switching time when all phases change from number 1 to number 0, and define it as the green light phase switching time;
[0070] S23. Arrange the green light phase switching times of all phases in chronological order, remove duplicate values, and obtain the data feature sampling time sequence;
[0071] S24. Based on the state switching timetable and the data feature sampling time sequence, using the start time of the phase state of each phase at the feature sampling time as the starting point of the calculation interval and the data feature sampling time as the ending point of the calculation interval, obtain the feature data calculation interval for each phase, such as... Figure 3 As shown.
[0072] S3. Using the initial dataset and the phase state switching timetable, construct a dataset containing all phase state features and traffic flow state features of the intersection, and construct a time series dataset from the feature data of continuous time steps.
[0073] S31. For A phases, calculate 6 feature data X1 to X6 for each phase, specifically including:
[0074] X1 represents the elapsed time of the current phase state within the data feature acquisition range;
[0075] X2 represents the traffic flow during the period of the current phase state;
[0076] X3 represents the time elapsed since the last activation of the coil sensor, during which the current phase state has been reached;
[0077] X4 indicates whether queuing or congestion has occurred during the current phase state. If the coil sensor triggering time exceeds the threshold, it is counted as 1, indicating that the above phenomenon has occurred.
[0078] X5 indicates the phase state at the time of data acquisition, with the numbers 1 and 0 representing the green and red light phase states, respectively.
[0079] X6 indicates whether a green light phase switch has occurred, with the numbers 1 and 0 representing whether a green light phase switch has occurred or not, respectively.
[0080] S32. Generate 6-dimensional feature data for each phase, stack the feature data of all phases and convert them into tensor form to form a T×6-dimensional feature data tensor.
[0081] S33. At the predicted time T n The 6-dimensional feature vectors of all A phases of the intersection are concatenated in a preset phase number order to generate a 6A-dimensional joint feature vector at the current moment.
[0082] S34. Set the current time T n The 6A-dimensional joint feature vectors of the preceding N-1 consecutive time steps are stacked in time sequence to construct a time sequence feature tensor of dimension N×6A, where N is the preset time window length and N≥A.
[0083] S35. Use the remaining duration of the current phase state of A phases as the predicted value Y.
[0084] S4. At the predicted time, divide the conflict relationship groups according to the preset conflict relationship and the actual green light switching phase, divide them according to the current phase state, and generate dynamic conflict relationship mask and phase state mask.
[0085] S41. Through on-site investigation or review of design data, determine the traffic flow conflict relationships of each phase control, and divide the phases into several groups: traffic flows controlled by phases within a group are non-conflicting phases and can be released simultaneously; traffic flows controlled by phases between groups constitute conflicting phases and simultaneous release is prohibited. For example... Figure 4 As shown, phases within each group are non-conflicting phases, while phases between groups are conflicting phases. Conflicting and non-conflicting categories are dynamically assigned based on preset phase grouping rules. At any given prediction time, phases in the same group as the green light switching phase are classified as non-conflicting, while the remaining phases are classified as conflicting. For example, at sampling time T, phases 1 and 4 switch from the green light phase to the red light phase. Phases 5, 8, and 10 in the same group are classified as non-conflicting. Phases 2, 3, 6, 7, and 9 in another group are classified as conflicting.
[0086] S42. Based on the aforementioned dynamic classification method for conflict and non-conflict categories, establish a dynamic conflict relationship mask vector m. c ∈{0,1} A ,in:
[0087]
[0088] S43. Based on the phase state feature X5 at the predicted time in S31, establish the phase state mask vector m. s ∈{0,1} A ,in:
[0089]
[0090] S5. Construct a phase duration prediction model based on TSMixer, and propose a novel architecture using a modular learning architecture, such as... Figure 5 As shown:
[0091] S51. In the phase state module, establish two TSMixer regression networks with identical structures: a red light phase network and a green light phase network, to predict the duration of A phase states. Specifically, the TSMixer regression network structure is as follows: Figure 6As shown, its input is an N×6A-dimensional temporal feature tensor, which is passed sequentially through N mixing layers, a temporal projection layer, and a feature projection layer corresponding to the time window, and outputs the predicted phase remaining duration. The mixing layers include sequentially connected temporal mixing MLPs and feature mixing MLPs.
[0092] like Figure 7 As shown, the temporal features input to the temporal hybrid MLP layer are normalized and transposed in two dimensions, and then input into a single-layer temporal MLP. By transposing the input, the temporal hybrid MLP shares features across all layers. The temporal MLP consists of a fully connected layer, an activation function, and a random deactivation layer, and is responsible for modeling the temporal patterns of the time series. The output of the temporal MLP is transposed again and then residually linked with the two-dimensional normalized input data to serve as the output of the temporal hybrid MLP layer.
[0093] like Figure 7 As shown, the feature mixing MLP layer is immediately followed by the time mixing MLP layer. After the output of the time mixing MLP layer is normalized in two dimensions, a feature MLP is input. The feature MLP uses two MLP layers to analyze the correlation between different features. Since it is applied to the feature domain, it is shared across all time steps. The output of the feature MLP is concatenated with the output residual of the time mixing MLP layer to obtain the output of the feature mixing MLP layer.
[0094] Furthermore, TSMixer applies 2D normalization in both the time and feature dimensions, which is a fusion of batch normalization and layer normalization, while standardizing the data in both the time axis and feature dimensions.
[0095] Nonlinear activation function: The activation function is applied to each MLP component in the mixing layer, i.e., temporal mixing MLP and feature mixing MLP. In this embodiment, the Gaussian Error Linear Unit (GELU) is preferably used as the activation function, and its definition is as follows:
[0096]
[0097] The GELU activation function ensures that the network does not suffer from the vanishing gradient problem during backpropagation while reducing the problem of neuron death, making it suitable for deep fully connected networks.
[0098] Residual connections: Residual connections are applied to each MLP component in the hybrid layer.
[0099] Temporal projection layer: The temporal projection layer is a fully connected layer applied in the time domain. The matrix output from the mixing layer is transposed again and predicted through a fully connected layer. After prediction, it is recovered by transposition. These layers not only learn the temporal pattern but also map the time series from the original input length L to the target prediction length T. Preferably, in this embodiment, the target prediction length T is 1, such as... Figure 8 As shown.
[0100] Feature projection layer: The feature projection layer is a fully connected layer applied on the feature domain. The matrix output from the time projection is transposed again here and predicted through a fully connected layer, mapping the feature sequence from the original input length 6A to the target prediction length A, and outputting the prediction results of the remaining duration of all phases.
[0101] S52. Retain the predicted phase value Y of the red light phase network when the light is red. r The predicted phase value Y of the green light phase network in the green light phase state g The phase state mask vectors are used to mask the outputs of the red and green light phase networks and then sum them to obtain an A-dimensional phase state prediction output vector, which can be specifically represented as follows:
[0102] Y phase =(Y g ⊙m s )+(Y r ⊙(1-m s ))
[0103]
[0104] Among them, Y phase represents the phase state network prediction value after mask merging, and ⊙ represents the element-wise multiplication of the prediction vector and the mask vector.
[0105] S53. In the conflict relationship module, construct two TSMixer regression networks with the same structure to predict the duration of A phase states, which serve as the non-conflict category network and the conflict category network, respectively; specifically, the TSMixer regression network in the conflict relationship module adopts the same TSMixer regression network structure as that in the phase state module.
[0106] S54. Use a dynamic conflict relationship masking vector to mask the prediction results of the non-conflict category network and the conflict category network to ensure that the phase prediction value in the output of each TSMixer regression network only retains the phase belonging to that category; add the two masked output results to obtain another A-dimensional conflict relationship prediction output vector;
[0107] Y rel =(Y comp ⊙m c)+(Y conf ⊙(1-m c ))
[0108]
[0109] Among them, Y comp Y represents the network prediction result for the non-conflict category. conf Conflict category network prediction results; m c This is a dynamic conflict relationship mask vector.
[0110] Because the predicted phase values Y differ significantly across different categories and exhibit periodicity, this invention establishes multiple sub-networks with identical structures to learn from each group, which helps the sub-networks converge. Since the prediction results of the sub-networks are masked, each sub-network can only receive the loss and gradient backpropagation for a specific type of phase. For example, the green light phase sub-network can only receive the error for predicting the green light phase duration; the non-conflict category sub-network learns from the phases within the same group as the green light switching phase.
[0111] S55. Combine the predicted values obtained in steps S53 and S54:
[0112]
[0113] Among them, Y concat Indicate Y phase and Y rel The vectors are then concatenated to obtain a prediction vector with a dimension of 2A.
[0114] S56. Input the concatenated predicted values into a multilayer perceptron (MLP) regression network to obtain the final predicted result Y for the duration of each signal phase state. pred .
[0115] Specifically, a multilayer perceptron regression network consists of an input layer, multiple hidden layers, and an output layer. The input layer receives the concatenated predicted values, the hidden layers extract high-level features from the input features, and the output layer generates the prediction result.
[0116] Input layer: The dimension of the input layer is 2A;
[0117] Hidden layers: The network contains several hidden layers. Each hidden layer includes a fully connected layer, a batch normalization layer, and a GELU activation layer cascaded in sequence. The first hidden layer receives the output of the input layer and maps it to a representation of 128 neurons. The second hidden layer processes the output of the first hidden layer again and maps it to a structure of 256 neurons.
[0118] Output layer: Based on the output of the last hidden layer, generate A phase state duration predictions.
[0119] S6. Using the feature dataset as input and the phase duration as output, train the prediction model and evaluate its accuracy, and optimize the model based on the evaluation results.
[0120] S61. Divide the feature dataset into a training set (60%), a validation set (20%), and a test set (30%);
[0121] S62. The optimizer is selected as Adaptive Moment Estimator (Adam), the batch size is set to 12, the learning rate is fixed at 0.001, and the loss function is Mean Squared Error (MAE):
[0122]
[0123] Where f represents the loss function, MAE represents the MAE loss function used for the surrogate model, and Y... i This represents the actual phase state duration data at time i. This represents the predicted phase state duration data at time i, and N represents the number of samples used for prediction.
[0124] S63. The model accuracy is evaluated using mean absolute error (MAE), root mean square error (RMSE), perfect hit rate (EH), and approximate hit rate (NM):
[0125]
[0126] Where RMSE represents the root mean square error; EH represents the perfect hit rate; NM represents the approximate hit rate; Y i This represents the actual phase state duration data at time i. This represents the predicted phase state duration data at time i, and N represents the number of samples used for prediction.
[0127] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.
Claims
1. A method for predicting the phase duration of an inductive signal based on TSMixer, characterized in that, Includes the following steps: Collect historical phase status and induction coil detection data of the inductively controlled intersection, and generate a phase switching timetable based on the historical phase switching times; Using the green light phase switching moment as the feature sampling moment, and the current phase start moment to the feature sampling moment as the feature calculation window, the phase and induction coil data of consecutive time steps within the feature calculation step window are extracted to construct a time series dataset; Based on the phase state and phase conflict relationship at the sampling time, the intersection phases are dynamically divided into conflict relationship groups and phase state groups, and dynamic mask representations are generated for each phase group. A phase duration prediction model based on multi-branch TSMixer is established, in which conflict relationship group and phase state group correspond to independent sub-models. Each sub-model processes time series datasets in parallel and outputs prediction results. The outputs of each sub-model are fused through a dynamic masking mechanism to generate the final phase remaining duration prediction.
2. The method for predicting the phase duration of an inductive signal based on TSMixer according to claim 1, characterized in that, The status data of the sensor-controlled intersection is collected as follows: Determine the phase of the induction signal intersection, the number of induction coils, and the correspondence between each induction coil and the phase; The historical phase status of the signalized intersection is represented by the numbers 0 and 1, which indicate the red and green light phases respectively. The yellow light phase is considered as part of the red light phase. The historical state of the induction coil at the intersection is represented by the numbers 0 and 1, indicating whether the induction coil is not activated or has been activated.
3. The method for predicting the phase duration of an inductive signal based on TSMixer according to claim 1, characterized in that, Obtain the induction phase state switching timetable and define the calculation range for all induction phase features, as follows: Using the obtained historical sensing signal phase state data, a timetable for all phase state switching is generated based on the changes in digital 0 and 1; Based on the phase switching timetable, determine the phase switching time when all phase states change from number 1 to number 0, and use this as the green light phase switching time. Arrange the green and red light switching times of all phases in chronological order, remove duplicate values, and obtain the data feature sampling time sequence; Based on the state switching timetable and sampling time sequence, the calculation interval for the characteristic data of each phase is obtained by taking the state start time of each phase at the characteristic sampling time as the starting point of the calculation interval and the characteristic sampling time as the ending point of the calculation interval.
4. The method for predicting the phase duration of an inductive signal based on TSMixer according to claim 1, characterized in that, The characteristic data of the sensing signal phase specifically includes: the duration of the phase that has passed within the data feature acquisition range; the traffic flow during the current phase state; the time elapsed since the last activation of the induction coil during the current phase state; whether queuing or congestion occurred during the current phase state; the phase state at the time of data acquisition; and whether the phase state has switched.
5. The method for predicting the phase duration of an inductive signal based on TSMixer according to claim 1, characterized in that, The time series dataset concatenates the multidimensional feature vectors of all phases in a preset phase number order to generate the joint feature vector at the current moment; The joint feature vectors of the current time and the preceding N-1 consecutive time steps are stacked in time sequence, where N is not less than the number of phases of the sensing intersection.
6. The method for predicting the phase duration of an inductive signal based on TSMixer according to claim 1, characterized in that, The phase duration prediction uses the phase state module to predict the remaining phase duration of the phase state group: The phase state module includes two identical TSMixer regression networks, namely the red light phase network and the green light phase network, which are used to predict the duration of each phase state. Based on the phase state characteristics, a dynamic mask for the predicted value is established. The output results of the red light phase network and the green light phase network of the phase state mask vector pair are used to mask the predicted values of the phase in the red light phase network and the phase in the green light phase network, respectively. The outputs of the masked red and green light phase networks are added together to obtain the first predicted output vector.
7. The method for predicting the phase duration of an inductive signal based on TSMixer according to claim 6, characterized in that, The phase duration prediction uses the conflict relationship module to predict the remaining phase duration of conflict relationship groups: The conflict relationship module divides non-conflicting phases and conflicting phases according to the preset phase group of the intersection and the group to which the current green light switching phase belongs, and constructs two TSMixer regression networks with the same structure, as non-conflicting category network and conflicting category network respectively, to predict the duration of each phase state. A group mask is established based on the group information of each phase. The group mask is used to mask the prediction results of the non-conflict class network and the conflict class network, so that the prediction values in the output of the non-conflict class network and the conflict class network only retain the phases belonging to that class. The outputs of the masked non-conflict class network and the conflict class network are added together to obtain the second predicted output vector.
8. The method for predicting the phase duration of an inductive signal based on TSMixer according to claim 7, characterized in that, The specific structure of the TSMixer regression network is as follows: The input is a δ×6A-dimensional temporal feature tensor, which is passed sequentially through N mixing layers, a temporal projection layer, and a feature projection layer corresponding to the time window, and the output is the prediction result of the remaining phase duration; wherein, the mixing layer includes a temporal mixing MLP and a feature mixing MLP connected in sequence. The temporal hybrid MLP layer normalizes the input temporal features in two dimensions and then transposes them. It then inputs the transpose into a single-layer temporal MLP to model the temporal pattern of the time series. The output of the temporal MLP is transposed again and then residually linked with the two-dimensional normalized input data to obtain the output of the temporal hybrid MLP layer. The feature mixing MLP layer is immediately followed by the time mixing MLP layer. After the output of the time mixing MLP layer is normalized in two dimensions, a feature MLP is input. The feature MLP uses two MLP layers to analyze the correlation between different features. The output of the feature MLP is residually connected with the output of the time mixing MLP layer to obtain the output of the feature mixing MLP layer. The time projection layer is a fully connected layer applied in the time domain. It transposes the moments output by the hybrid layer and performs prediction through a fully connected layer, mapping the time series from the original input length to the target prediction length T. The prediction output of the fully connected layer is then recovered through transposition. The feature projection layer is a fully connected layer applied on the feature domain. The matrix output from the time projection is transposed again and predicted through a fully connected layer, mapping the feature sequence from the original input length to the target prediction length, and outputting the prediction results of the remaining duration of all phases.
9. The method for predicting the phase duration of an inductive signal based on TSMixer according to claim 7, characterized in that, The phase duration prediction model concatenates the prediction output vectors obtained from the phase state module and the conflict relationship module and inputs them into a multilayer perceptron regression network to obtain the final prediction result of the phase state duration of each signal.
10. The method for predicting the phase duration of an inductive signal based on TSMixer according to claim 9, characterized in that, The multilayer perceptron regression network is as follows: Input layer: The dimension of the input layer is determined by the number of features in the input data, with each feature corresponding to one input unit; Hidden layers: The network contains several hidden layers, each of which includes a fully connected layer, a batch normalization layer, and a non-linear activation layer cascaded in sequence; The first hidden layer receives the output of the input layer and maps it to a representation of 128 neurons; The second hidden layer processes the output of the first hidden layer again and maps it into a structure of 256 neurons. Output layer: Predicts the duration of each phase state.