Innovation-based efficient communication quantization coding method

Abstract

The invention provides an efficient communication quantization coding method based on innovation. The method comprises the following steps: establishing a state space model of a target system and a sensor node; constructing a data packet of a measurement vector and measurement information independent of a system state based on the state space model; based on a differential coding strategy, an information-based coding-decoding quantization algorithm is established, so that a coded data packet is composed of a calibration value and a value for measuring information; establishing a first optimization problem by taking data validity as a constraint condition of a reconstruction error and taking a communication bandwidth saving rate as an optimization target, and solving to obtain an optimal quantization bit number and a data item number in a data packet; and according to an innovation-based coding-decoding quantization algorithm, coding and sending of a sender and receiving and decoding of a receiver are carried out, so that the receiver obtains a reconstructed measurement vector. According to the invention, high-efficiency data transmission can be realized in a scene with limited communication bandwidth, and the validity of data transmission is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of wireless sensor networks, and in particular to a quantization coding method based on innovation, which can significantly improve communication efficiency when processing large-scale observation signals. Background Art

[0002] In recent years, with the rapid development of wireless sensor networks, the Internet of Vehicles, and the Industrial Internet, efficient data quantization and coding in networked systems have been extensively studied. In data quantization and coding, sensor nodes continuously collect observation data, encode and quantize the data to reduce data redundancy and improve transmission efficiency. Data coding and quantization can provide efficient data support for tasks such as distributed state estimation, remote monitoring, and intelligent control, improving system stability and real-time performance. Due to bandwidth limitations of communication networks, traditional data transmission methods can easily lead to excessive communication load, affecting system performance. Therefore, research on efficient data quantization and coding methods is of great significance for improving the communication efficiency of networked systems.

[0003] Existing data encoding and quantization methods often face problems such as high data redundancy and high communication overhead when processing large-scale observation signals, which can affect system performance. Therefore, we investigate an efficient communication quantization coding method based on innovations. This method can effectively reduce data transmission volume while maintaining data accuracy, thereby improving the data processing efficiency of networked systems and optimizing the communication performance and energy management capabilities of wireless sensor networks.

[0004] In view of this, it is necessary to propose a quantization coding method to ensure the quantization error when observing signals on a large scale, and design a feedback quantization mechanism to further reduce the error. Summary of the Invention

[0005] The object of the present invention is to provide an efficient communication quantization coding method based on innovation, which still has good data validity and communication efficiency when observing signals on a large scale.

[0006] In order to achieve the above object, the present invention provides an efficient communication quantization coding method based on innovation, comprising:

[0007] S1, establish the state space model of the target system and sensor nodes;

[0008] S2, constructing a data packet of measurement vectors and measurement innovations independent of the system state based on the state space model to reduce the accuracy loss caused by large-scale observation output;

[0009] S3, based on the differential coding strategy, establishes an innovation-based encoding-decoding quantization algorithm so that the encoded data packet consists of the calibration value and the measurement innovation value;

[0010] S4, establishing a first optimization problem with data validity as a constraint on the reconstruction error and a communication bandwidth saving rate η as an optimization objective; solving the first optimization problem to obtain the optimal number of quantization bits and the number of data items in the data packet required by the innovation-based encoding-decoding quantization algorithm;

[0011] S6, performing encoding and sending at the sender and receiving and decoding at the receiver according to the encoding-decoding quantization algorithm based on the innovation, so that the receiver obtains a reconstructed measurement vector.

[0012] In the step S1, the target system is a real-world physical system whose physical parameters can be measured by sensors, and the state of the target system refers to the physical parameters;

[0013] The expressions of the state space model of the system state and sensor nodes are:

[0014] x(k+1)=Ax(k)+w(k),

[0015] y(k)=Hx(k)+v(k),

[0016] Where x(k) is the true state of the system at time k, m is the dimension of the true state, x(k+1) is the true state of the system at time k+1, A is the system state transfer matrix, w(k) is the Gaussian system noise at time k, with mean zero and variance Q; y(k) is the measurement vector at time k, n is the dimension of the measurement vector, H is the measurement matrix of the sensor, v(k) is the Gaussian measurement noise at time k, with mean zero and variance R; k is the first moment.

[0017] The step S2 specifically includes:

[0018] S21, define the innovation gain matrix M = HAH -1 , the encoded measurement innovation of the measurement vector y(+τ+1) is defined as Δy(k+τ+1)=y(k+τ+1)-My(k+τ), where y(k+τ+1) and y(k+τ) are the measurement vectors at time k+τ+1 and k+τ, respectively, and τ represents an arbitrary time. The measurement vector between two communications between the sensor and the remote server is recorded as the measurement vector data packet y(k,k+T-1);

[0019] S22, based on the system state and the state space model of the sensor node, obtain the expansion of the measurement innovation;

[0020] S23, M=HAH -1 Substitute the expansion of the measurement innovation and calculate the variance matrix of the measurement innovation. According to the variance matrix of the measurement innovation Determine a normal value interval for measuring innovation. The normal value interval for measuring innovation is used to ensure that the frequency of abnormal values ​​of the measuring innovation meets the requirements.

[0021] In the innovation-based encoding-decoding quantization algorithm, the original measurement vector data packet is encoded and then sent during each communication. The encoding method includes:

[0022] S31, taking the measurement vector at the first moment in the data packet of the original measurement vector as a calibration value; encoding the measurement vectors at other moments other than the first moment as measurement innovation;

[0023] S32, for the measurement innovation Δy(k+τ+1) at the rest of the time except the first time, τ=0,1,…,T-1, T is the number of data items in the data packet, use the probabilistic quantizer to quantize the scalar of the measurement innovation Δy(k+τ+1) in each dimension, obtain the quantized innovation value of each dimension, and combine them to obtain the quantized measurement innovation Then we get the encoded and quantized data packets

[0024] S33, the scalar Δy in each dimension of the measurement innovation obtained by the probability quantizer i (k+τ+1) and the quantized innovation value of this dimension The deviation between them is defined as the quantization error n i (k+τ+1), and determine the covariance of the quantization error.

[0025] The probability quantizer used meets the following requirements: the scalar Δy in the i-th dimension according to the measurement innovation Δy(k+τ+1) i (k+τ+1) (i is the dimension number of the real state), the quantized innovation value of the i-th dimension for:

[0026]

[0027] Where i = 1, 2, ..., m, m is the dimension of the real state, i is the dimension ordinal number of the real state; Δy i (k+τ+1) is the scalar of the measurement innovation Δy(k+τ+1) in the i-th dimension; q j+1 (k) and q j (k) is the quantization level, j is the ordinal number of the quantization level; k is the first moment, Δ is the quantization step size;

[0028] The quantization step size Δ is:

[0029]

[0030] in, u, are the lower and upper limits of the signal range to be quantized, and b is the number of quantization bits;

[0031] And the relationship between the quantization level, quantization step size and quantization bit number is as follows:

[0032] Δ=q j+1 -q j ,j=0,1,…,2 b -2,

[0033] Among them, q j+1 ,q j is the quantization level, q0= u ,

[0034] In the innovation-based encoding-decoding quantization algorithm, the receiver decodes the encoded and quantized data packet to obtain the reconstructed data packet. Decoding methods include:

[0035] S31', according to the formula M=HAH -1 Calculate the innovation gain matrix M and decode it to get the reconstructed measurement vector Merge to obtain the reconstructed data packet

[0036] Reconstructed measurement vector for:

[0037]

[0038] in, is the reconstructed measurement vector at time k, which is the same as the measurement vector before encoding and quantization; is the reconstructed measurement vector at time k+τ+1, M is the innovation gain matrix, M=HAH -1 , H is the sensor measurement matrix, A is the system state transfer matrix, is the quantized measurement information at time k+τ+1;

[0039] S32', reconstruct the measurement vector at time k+τ+1 The deviation from the original measurement information y(k+τ+1) at time k+τ+1 is defined as the reconstruction error e(k+τ+1); the covariance of the reconstruction error e(k+τ+1) is calculated and the normal numerical range of the reconstruction error is determined accordingly.

[0040] In step S4, the first optimization problem P1 is established as:

[0041]

[0042] Among them, b i is the number of quantized bits, i is the dimension number of the real state, m is the dimension of the real state, T is the number of data items in the data packet, is the upper bound of the normal numerical interval of the reconstruction error, ∈ i For data validity, Z represents a set of integers.

[0043] For the receiver, after decoding, the reconstructed measurement vector Before, it also includes: the total number of bits B of the combination of the number of data items T in the data packet and the calibration value of the received encoded and quantized data packet and the scalar of the i-th dimension of the measurement innovation i , calculate the number of quantized bits b of the scalar of the i-th dimension of each measurement innovation i , to divide the starting and ending positions of different measurement innovations.

[0044] The efficient communication quantization coding method based on novel information also includes step S5, further introducing a feedback quantization algorithm in the encoding and decoding process; if the actual standard deviation of the measured novel information is less than the theoretical standard deviation, the feedback quantization algorithm solves the feedback-corrected quantization bit number, and the feedback-corrected quantization bit number ensures that the reconstruction error can be limited to the range of the quantization step and maximizes the communication bandwidth saving rate; then the novel information-based encoding-decoding quantization algorithm is executed to obtain the receiver reconstruction error simulated by the sender as the feedback input of the probability quantizer to perform secondary quantization on the measured novel information to ensure that the receiver reconstruction error simulated by the sender is within the range of the quantization step, and then send the data packet; otherwise, the sender directly sends the quantized data packet using the novel information-based encoding-decoding quantization algorithm in step S3 and the quantization bit number in step S4.

[0045] The step S5 specifically includes:

[0046] S51: If the actual standard deviation of the measured innovation is less than the theoretical standard deviation, the second optimization problem P2 is solved to determine the number of quantization bits for feedback correction that ensures the reconstruction error is limited to the quantization step size and maximizes the communication bandwidth saving rate. Otherwise, the quantized data packet obtained by using the encoding-decoding quantization algorithm based on the innovation in step S3 and the quantization bit number in step S4 is directly sent, and then step S6 is executed.

[0047] The second optimization problem P2 is:

[0048]

[0049] in,

[0050]

[0051] Among them, Γ is the sufficient condition for the number of quantization bits; Γ x.y , Both represent subsets of Γ; i k ,j k ∈{1,2,…,2m} is a qualified subscript index, and j, x=1, 2, ..., m, y=1, 2, ..., 2m is also a subscript index; M z,i Represents the zth row and ith column of the innovation gain matrix.

[0052] The innovation-based quantization coding method proposed in this paper can achieve excellent communication resource savings while ensuring data validity when processing large-scale observation signals. Furthermore, the feedback quantization method provided by this invention can significantly reduce data errors, improve data availability, and effectively reduce communication resource consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 The flowchart of an innovation-based efficient communication quantization coding method according to one embodiment of the present invention is shown.

[0054] Figure 2 This is an exemplary schematic diagram of quantization encoding and decoding of the present invention.

[0055] Figure 3A-Figure 3E An exemplary diagram of algorithm error and MSE provided for some embodiments of the present invention. DETAILED DESCRIPTION

[0056] The preferred embodiments of the present invention are given below in conjunction with the accompanying drawings and described in detail.

[0057] The present invention provides an efficient communication quantization coding method based on new information, which still has good data validity and communication efficiency when observing signals on a large scale. The principle of the efficient communication quantization coding method based on new information of the present invention is to ensure the quantization error when observing signals on a large scale by constructing measurement new information that is independent of the system state, and to design a feedback quantization mechanism to further reduce the error. In which, the measurement vectors at time k+τ+1 and k+τ are y(k+τ+1) and y(k+τ) respectively, the system state transfer matrix is ​​A, the sensor measurement matrix is ​​H, and the new information gain matrix M=HAH is defined. -1 , then the measurement information is defined as: Δy(k+τ+1)=y(k+τ+1)-My(k+τ).

[0058] The present invention generally provides an efficient communication quantization coding method based on innovation, which includes: establishing a state space model of sensor nodes; constructing measurement innovations independent of the system state based on the state space model; proposing an encoding-decoding quantization algorithm based on innovation based on a differential coding strategy; solving the minimum number of quantization bits under a given reconstruction error constraint to balance data validity and transmission resource consumption; and designing a feedback quantization algorithm to constrain the error within the quantization step size.

[0059] The present invention provides an efficient communication quantization coding estimation method based on innovation, which specifically includes:

[0060] Step S1, establishing a state space model of the target system and sensor nodes;

[0061] In the present invention, a target system is a real-world physical system whose physical parameters can be measured by sensors. The state of the target system refers to a physical parameter, which includes at least one of position, velocity, acceleration, angle, voltage, current, resistance, power, temperature, humidity, pressure, surface tension, and pressure. The present invention models the state of the target system, so the type of physical parameters included in the system state varies depending on the real-world physical system being modeled. For example, if the target system is a target vehicle moving in a straight line, then the system state is the vehicle's position and velocity. Sensors need to measure the vehicle's position and velocity to establish a state-space model of the vehicle's position and velocity.

[0062] The sensor node may be a plurality of sensors or a single sensor, and is used to measure the system state to obtain a measurement vector. In this embodiment, when the target system is a vehicle, the measurement vector may be position and velocity, or a linear combination of position and velocity. Generally speaking, the sensor node may be a wireless sensor network comprising multiple sensors. In this embodiment, only the case of a single sensor is described, but the method of the present invention can be extended to a multi-sensor network without additional configuration. The parameters of the measurement vector depend on the performance of the sensor. In the present invention, the sensor is required to be able to measure all states of the system.

[0063] In step S1, the expressions of the system state and the state space model of the sensor node are:

[0064] x(k+1)=Ax(k)+w(k),

[0065] y(k)=Hx(k)+v(k)

[0066] Among them, x(k) is the true state of the system at time k, m is the dimension of the true state, the target state and the true state are the same concept, x(k+1) is the true state of the system at time k+1, A is the system state transfer matrix, w(k) is the Gaussian system noise at time k, with a mean of zero and a variance of Q; y(k) is the measurement vector at time k, which comes from the sensor's measurement of the system state, n is the dimension of the measurement vector, and in the present invention, m=n is required, H is the measurement matrix of the sensor, v(k) is the Gaussian measurement noise at time k, with a mean of zero and a variance of R; k is the first moment.

[0067] Step S2: constructing a data packet of measurement vectors and measurement innovations independent of the system state based on the state space model to reduce the accuracy loss caused by large-scale observation output;

[0068] The step S2 specifically includes:

[0069] Step S21, define the innovation gain matrix M = HAH -1 , the encoded measurement innovation of the measurement vector y(+τ+1) is defined as Δy(k+τ+1)=y(k+τ+1)-My(k+τ), where y(k+τ+1) and y(k+τ) are the measurement vectors at time k+τ+1 and k+τ respectively; and the measurement vector between two communications between the sensor and the remote server is recorded as the measurement vector data packet y(k,k+T-1);

[0070] That is, the measurement innovation is the measurement value at the next moment minus M times the measurement value at the previous moment.

[0071] Between two communications between the sensor and the remote server, the sensor performs a total of T measurements of the system state, which are recorded as the original measurement vector data packet y(k,k+T-1)=[y(k),y(k+1),…,y(k+T-1)],

[0072] Where y(k) is the measurement vector at time k, k is the first time, and T is the number of data items in the data packet. In other words, it means that there are a total of T measurement vectors in the data packet.

[0073] Step S22, obtaining an expansion of the measurement innovation according to the system state and the state space model of the sensor node;

[0074] In step S22, the expanded form of the encoded measurement innovation Δy(k+τ+1) of the measurement vector y(+τ+1) is:

[0075] Δy(k+τ+1)=y(k+τ+1)-My(k+τ)=HAx(k+τ)-MHx(k+τ)+Hw(k)+v(vk+1)-M(k)=(HA-MH)x(k)+Hw(k)+v(k+1)-Mv(k),

[0076] Where M is the innovation gain matrix, M∈R m×m , m is the dimension of the real state, M = HAH -1 , A is the system state transfer matrix, H is the sensor measurement matrix, y(k+τ+1) and y(k+τ) are the measurement vectors at time k+τ+1 and k+τ, and x(k+τ) is the true state of the system at time k+τ; w(k) is the Gaussian system noise at time k, v(k) is the Gaussian measurement noise at time k; v(k+1) is the Gaussian measurement noise at time k+1, k is the first time, and τ represents any time.

[0077] Step S23, set M=HAH -1 Substitute the expansion of the measurement innovation and calculate the variance matrix of the measurement innovation. According to the variance matrix of the measurement innovation Determine a normal value interval for measuring innovation. The normal value interval for measuring innovation is used to ensure that the frequency of abnormal values ​​of the measuring innovation meets the requirements.

[0078] In order to make the measurement innovation independent of the system state x(k), it is required that M = HAH -1 , at this time, the measurement innovation is only related to the system state transfer matrix A, the measurement matrix H, the Gaussian system noise w(k) and the Gaussian measurement noise v(k). Since w(k) is a Gaussian system noise with zero mean and variance Q, and v(k) is a Gaussian measurement noise with zero mean and variance R, the two are independent of each other, so the measurement innovation is a zero mean, with a measurement innovation variance matrix Gaussian variables, and the variance matrix of the measurement innovation Expressed as a covariance matrix, the calculation formula is:

[0079]

[0080] Where Q is the variance of the Gaussian system noise, R is the variance of the Gaussian measurement noise, H represents the measurement matrix, H′ represents the transpose of the measurement matrix, M is the innovation gain matrix, and M′ is the transpose of the innovation gain matrix M.

[0081] For a Gaussian variable with a standard deviation of σ, 99.74% of its values ​​are distributed within the range of [-3σ, +3σ]. Therefore, using σ i The variance matrix representing the measurement innovation The square root of the i-th diagonal element of , then the scalar Δy that measures the i-th dimension of the innovationi 99.74% of the values ​​of (k+τ+1) are distributed in the normal value range of the measurement innovation [-3σ i ,+3σ i ], where i = 1, 2, ..., m, τ = 0, 1, ..., T-2, τ represents any time, T is the number of data items in the data packet, m is the dimension of the real state, i is the dimension ordinal number of the real state, σ i The variance matrix representing the measurement innovation The square root of the i-th diagonal element of . Consider the scalar Δy measuring the i-th dimension of the innovation i If (k+τ+1) exceeds the normal value range of the measurement innovation [-3σ i ,+3σ i In some embodiments, when the number of data items in a data packet is T=200, there is only one outlier. The outlier has a low frequency of occurrence and only affects the data packet where the outlier is located, with limited impact on downstream tasks.

[0082] Therefore, by constructing the measurement innovation value according to the method of the present invention, even large-scale observation output caused by system state divergence will not affect the numerical range of the measurement innovation Δy(k+τ+1), thereby avoiding quantizer saturation.

[0083] Step S3: Based on the differential coding strategy, establish an innovation-based encoding-decoding quantization algorithm so that the encoded data packet consists of the calibration value and the measurement innovation value;

[0084] In the innovation-based encoding-decoding quantization algorithm, the original measurement vector data packet is encoded and then sent during each communication. The encoding method includes:

[0085] In step S31, the measurement vector at the first moment in the data packet of the original measurement vector is used as a calibration value, which remains unchanged without any change or generation; and the measurement vectors at other moments other than the first moment are encoded as measurement innovations.

[0086] Therefore, after this encoding method, the original measurement vector data packet y(k, k+T-1) is the encoded data packet for:

[0087]

[0088] Where y(k) is the measurement vector at the first moment, which serves as the calibration value; k is the first moment; Δy(k+1),…,Δy(k+T-1) are the measurement information obtained by encoding the measurement vectors y(k+1),…,y(k+T-1) at the remaining moments other than the first moment, and T is the number of data items in the data packet.

[0089] For a system with high real-time requirements, the number of data items in a data packet, T, depends on the sensor sampling frequency and communication frequency. If the sensor sampling frequency is 60 Hz and the communication frequency is 10 Hz, then T is 60 / 10 = 6. In this case, T ranges from 1 to 600. For systems with lower real-time requirements, such as data compression, a higher T value results in a higher data compression rate. In this case, T has no specific range.

[0090] Step S32: For the measurement innovation Δy(k+τ+1) at the remaining moments other than the first moment, τ=0, 1, ..., T-1, T is the number of data items in the data packet, use the probabilistic quantizer to quantize the scalar of the measurement innovation Δy(k+τ+1) in each dimension, obtain the quantized innovation value of each dimension, and combine them to obtain the quantized measurement innovation. Then we get the encoded and quantized data packets

[0091] In this embodiment, the probabilistic quantizer used meets the following requirements:

[0092] According to the measurement innovation Δy(k+τ+1) in the i-th dimension scalar Δy i (k+τ+1) (i is the dimension number of the real state), the quantized innovation value of the i-th dimension for:

[0093]

[0094] Where i = 1, 2, ..., m, m is the dimension of the real state, i is the dimension ordinal number of the real state; Δy i (k+τ+1) is the scalar of the measurement innovation Δy(k+τ+1) in the i-th dimension; q j+1 (k) and q j (k) is the quantization level, j is the ordinal number of the quantization level, and the quantization level is used to divide the data range into discrete values; k is the first moment, and Δ is the quantization step size;

[0095] Among them, the quantization step size Δ is:

[0096]

[0097] in, u , is the lower limit and upper limit of the signal range to be quantized (that is, the range of the signal to be quantized is ), b is the number of quantization bits, and the number of quantization bits is calculated based on the given error requirement.

[0098] And the quantization level qj , the relationship between the quantization step size Δ and the number of quantization bits b satisfies the following:

[0099] Δ=q j+1 -q j ,j=0,1,…,2 b -2,

[0100] Among them, q j+1 ,q j is the quantization level, q0= u , That is, the first quantization level is the lower boundary of the input data range, and the last quantization level is the upper boundary of the input data range;

[0101] That is, for different quantization levels, the value of the quantization step Δ between two adjacent quantization levels is fixed.

[0102] Step S33: convert the measurement innovation obtained by the probability quantizer into a scalar Δy in each dimension. i (k+τ+1) and the quantized innovation value of this dimension The deviation between them is defined as the quantization error n i (k+τ+1), and determine the covariance of the quantization error.

[0103] The scalar Δy in each dimension of the measurement innovation obtained by the probability quantizer is i (k+τ+1) and the quantized innovation value of this dimension The deviation between them is defined as the quantization error n i (k+τ+1), in some embodiments, when the quantization step size Δ is much larger than the standard deviation of the input signal of the probabilistic quantizer, the quantization error has a mean of 0 and an upper bound of 0.25Δ 2 Gaussian distribution.

[0104] Since the measurement innovation is the scalar Δy in each dimension i (k+τ+1) is generally in the normal value range of measuring innovation [-3σ i ,+3σ i ], i = 1, 2, ..., m, therefore, the present invention will measure the lower bound of the normal value interval of the new information -3σ i As the lower limit of the signal range to be quantized u , that is -3σ i = u , the upper bound of the normal numerical interval of the measurement innovation is +3σ i The upper limit of the signal range to be quantified Right now Data beyond the upper boundary are quantified as +3σ i, the value less than the lower bound is quantized to -3σ i However, the encoding method proposed in this patent can effectively ensure that almost no data exceeds the lower and upper limits of the signal range to be quantized. u ,

[0105] Since the measurement innovation is a vector, the scalar of each dimension needs to be quantized. The scalar Δy of the measurement innovation in the i-th dimension i (k+τ+1) uses the given number of quantization bits b of the i-th dimension i To quantize, we can get the quantization step size Δ under the i-th dimension i for:

[0106]

[0107] Among them, σ i The variance matrix representing the measurement innovation The square root of the i-th diagonal element of i is the number of quantization bits for the i-th dimension.

[0108] At this time, the quantization error n i Covariance of (k+τ+1) for:

[0109]

[0110] Among them, Δ i is the quantization step size in the i-th dimension, σ i The variance matrix representing the measurement innovation The square root of the i-th diagonal element of i is the number of quantization bits for the i-th dimension.

[0111] In the innovation-based encoding-decoding quantization algorithm, the receiver decodes the encoded and quantized data packet to obtain the reconstructed data packet. Decoding methods include:

[0112] Step S31', according to the formula M=HAH -1 Calculate the innovation gain matrix M and decode it to get the reconstructed measurement vector Merge to obtain the reconstructed data packet

[0113] Reconstructed measurement vector for:

[0114]

[0115] in, is the reconstructed measurement vector at time k, which is the same as the measurement vector before encoding and quantization; is the reconstructed measurement vector at time k+τ+1, M is the innovation gain matrix, M=HAH -1 , H is the sensor measurement matrix, A is the system state transfer matrix, is the quantized measurement information at time k+τ+1.

[0116] Step S32', reconstruct the measurement vector at time k+τ+1 The deviation from the original measurement information y(k+τ+1) at time k+τ+1 is defined as the reconstruction error e(k+τ+1); the covariance of the reconstruction error e(k+τ+1) is calculated and the normal numerical range of the reconstruction error is determined accordingly.

[0117] It should be noted that the receiver cannot obtain the true measurement value (i.e., the original measurement innovation y(k+τ+1)), so it is impossible to calculate the reconstruction error e(k+τ+1). However, the receiver can calculate the diagonal elements σ of the variance matrix of the measurement innovation based on the system parameters. i .

[0118] That is, the reconstruction error e(k+τ+1) is:

[0119]

[0120] According to the above recursive relationship and e(k+0)=0, the reconstruction error e(k+τ+1) is:

[0121] e(k+τ+1)=M τ n(k+1)+M τ-1 n(k+2)+…+Mn(k+τ)+ n(k+τ+1)

[0122] Among them, e(k+τ+1) is the reconstruction error at time k+τ+1, M is the innovation gain matrix, M τ represents M to the power of τ, n(k+1) is the quantization error of the measurement vector y(k+1) at time k+1, n(k+τ+1) is the quantization error of the measurement vector y(k+τ+1) at time k+τ+1, and so on.

[0123] Therefore, the covariance of the reconstruction error e(k+τ+1) for:

[0124]

[0125]

[0126] Among them, Δi is the quantization step size in the i-th dimension, σ i The variance matrix representing the measurement innovation The square root of the i-th diagonal element of i is the number of quantization bits of the i-th dimension, is the covariance of the reconstruction error, is the covariance of the quantization error, M is the innovation gain matrix, (M τ )′ represents the transpose of M raised to the power of τ, and T is the number of data items in the data packet.

[0127] Since the quantization errors are independent of each other, we can get Similar to the measurement innovation range, the i-th scalar e of the reconstruction error at time k+τ+1 is i Most of the values ​​of (k+τ+1) are distributed in the normal range of reconstruction error The values ​​outside this range are defined as abnormal values, where Represents the covariance of the reconstruction error The element in the i-th row and i-th column of .

[0128] Step S4: Establish a first optimization problem with data validity as the constraint of reconstruction error and communication bandwidth saving rate η as the optimization target; solve the first optimization problem to obtain the optimal number of quantization bits b required for the encoding-decoding quantization algorithm based on the innovation. i and the number of data items T in the data packet.

[0129] Thus, the first optimization problem is established to balance data availability and bandwidth consumed by transmission resources.

[0130] Encoded data packet is the encoded data containing the number of data items in T data packets. In some embodiments, the error when quantizing data using 32 bits can be ignored and the data is considered accurate; therefore, when no quantization strategy is used, the scalar Δy of the i-th dimension of the new information is measured. i (k+τ+1) are all transmitted using 32 bits; when using the quantization strategy, in order to make the calibration value (i.e., the measurement vector y(k) at time k) error-free, 32 bits are used to transmit the scalar y of the i-th dimension of the calibration value i (k), in order to save bandwidth, measure the scalar Δy of the i-th dimension of the innovation i (k+τ+1) uses the number of quantization bits b i Transmitted after quantization, where i = 1, 2, …, m, τ = 0, 1, …, T-1, i is the dimensionality ordinal of the real state, m is the dimension of the real state, τ represents any time, and T is the number of data items in the data packet.

[0131] Therefore, the communication bandwidth saving rate η is:

[0132]

[0133] Among them, b i is the number of quantized bits, i is the dimension number of the real state, m is the dimension of the real state, and T is the number of data items in the data packet.

[0134] In order to ensure the validity of the reconstructed data, data validity ∈ i As an upper bound imposed on the normal value interval of the reconstruction error superior.

[0135] Therefore, in step S4, the first optimization problem P1 is established as:

[0136]

[0137] Among them, b i is the number of quantized bits, i is the dimension number of the real state, m is the dimension of the real state, T is the number of data items in the data packet, is the upper bound of the normal numerical interval of the reconstruction error, ∈ i For data validity, Z represents a set of integers.

[0138] Since the number of data items in the data packet is T and the number of quantization bits is b i are all finite integers, so the first optimization problem P1 can be solved by exhaustive method, and the objective function is monotonic with respect to the optimization parameters, so the time complexity of the bisection optimization algorithm can be used from O(m 4 ×32 m ) to O(m 5 ).

[0139] Therefore, solving the first optimization problem P1 can obtain the given system parameters (A, H, Q, R) and data validity ∈ i Under the constraint of , we can get the number of data items T in the data packet that maximizes the communication bandwidth saving rate and the number of quantized bits b in each dimension of the measurement innovation. i .

[0140] Here, data validity ∈ i The value of is artificially given. The values ​​of the system state transfer matrix A and the measurement matrix H are also known and are used when calculating the innovation gain matrix M. The upper bound of the normal numerical range of the reconstruction error is The covariance of the reconstruction error is calculated by combining the innovation gain matrix M with the variance Q of the Gaussian system noise and the variance R of the Gaussian measurement noise.

[0141] Here, the covariance of the reconstruction error is for:

[0142]

[0143]

[0144] Among them, Δ i is the quantization step size in the i-th dimension, σ i The variance matrix representing the measurement innovation The square root of the i-th diagonal element of i is the number of quantization bits of the i-th dimension, is the covariance of the reconstruction error, is the covariance of the quantization error, M is the innovation gain matrix, (M τ )′ represents the transpose of M raised to the power of τ, and T is the number of data items in the data packet.

[0145] When the data packet of the original measurement vector is quantized according to the coding-decoding quantization algorithm based on the innovation in step S3 and the optimal quantization bit number b in step S4 i After encoding and quantization, it is transmitted to the remote server in binary form through the wireless channel.

[0146] For the receiver, after decoding, the reconstructed measurement vector Before, it also includes: the total number of bits B of the combination of the number of data items T in the data packet and the calibration value of the received encoded and quantized data packet and the scalar of the i-th dimension of the measurement innovation i , calculate the number of quantized bits b of the scalar of the i-th dimension of each measurement innovation i , to divide the starting and ending positions of different measurement innovations.

[0147] The number of quantized bits b of the scalar of the i-th dimension of each measurement innovation i for:

[0148]

[0149] Where T is the number of data items in the data packet, which is determined by the receiver based on the local clock; B i is the number of bits of the combination of the calibration value of the data packet and the scalar of the i-th dimension of the measurement innovation.

[0150] When the remote server knows the sensor's measurement matrix H, the variance R of the Gaussian measurement noise, the system state transfer matrix A, and the variance Q of the Gaussian system noise, it can use the formula M=HAH -1 Calculate the innovation gain matrix M, and based on the variance matrix of the measured innovation The calculation formula Calculate the diagonal elements σ of the variance matrix of the measurement innovation i , and accordingly determine the range of the innovation signal [-3σ i ,+3σ i ]; At this time, for the quantization bit number b i The receiver can decode it as +3σ in decimal. i ; For the quantization bit number b i The receiver can decode it as -3σ in decimal. i ; This avoids adding the information required for decimal and binary conversion and improves communication bandwidth utilization.

[0151] Step S5, further introducing a feedback quantization algorithm during the encoding and decoding process to further reduce the error;

[0152] If the actual standard deviation of the measured innovation is less than the theoretical standard deviation, the feedback quantization algorithm solves the number of quantization bits for feedback correction, and the number of quantization bits for feedback correction ensures that the reconstruction error can be limited to the range of the quantization step and maximizes the communication bandwidth saving rate; then the encoding-decoding quantization algorithm based on the innovation is executed to obtain the receiver reconstruction error simulated by the sender as the feedback input of the probability quantizer to perform secondary quantization on the measured innovation to ensure that the receiver reconstruction error simulated by the sender is within the range of the quantization step, and then send the data packet; otherwise, the sender directly sends the quantized data packet using the encoding-decoding quantization algorithm based on the innovation in step S3 and the quantization bit number in step S4.

[0153] Since the sender already knows the encoding-decoding quantization algorithm based on the new information adopted by the receiver in step S2, in step S5, the probability quantizer first quantizes and reconstructs the new information based on the encoding-decoding quantization algorithm in step S2 to obtain the encoded and quantized data packet and the reconstructed data packet, and compares the reconstructed data packet with the data packet of the original measurement vector to obtain the receiver reconstruction error simulated by the sender, and uses it as the feedback input of the probability quantizer to perform secondary quantization (i.e., feedback quantization) on the quantized measurement new information, and then sends the data packet. Among them, when trying to perform secondary quantization on the quantized measurement new information, the values ​​of each dimension of the secondary quantized measurement new information are required to be All in [-Δ i ,+Δ i ],i=1,2,…,m, if the values ​​of each dimension of the new information are measured Already in [-Δ i ,+Δ i ], no secondary quantization is performed.

[0154] The step S5 specifically includes:

[0155] In step S51, if the actual standard deviation of the measured innovation is less than the theoretical standard deviation, the number of quantization bits for feedback correction is calculated to ensure that the reconstruction error can be limited to the range of the quantization step and that the communication bandwidth saving rate is maximized; otherwise, the quantized data packet obtained by using the encoding-decoding quantization algorithm based on the innovation in step S3 and the quantization bit number in step S4 is directly sent, and then step S6 is executed.

[0156] The step S51 specifically includes:

[0157] Step S511: The sender determines the variance matrix of the measurement innovation based on the measurement matrix H, the variance R of the Gaussian measurement noise, the system state transfer matrix A, and the variance Q of the Gaussian system noise. The square root of the i-th diagonal element σ i ,i=1,2,…,m, as the theoretical standard deviation;

[0158] As mentioned above, the variance matrix of the measurement innovation is for:

[0159]

[0160] Where Q is the variance of the Gaussian system noise, R is the variance of the Gaussian measurement noise, H represents the measurement matrix, H′ represents the transpose of the measurement matrix, M is the innovation gain matrix, and M′ is the transpose of the innovation gain matrix M.

[0161] Step S512: According to the encoded data packet (i.e., encoded but not quantized data packets), calculate the actual standard deviation σ of the measured innovation i ′,i=1,2,…,m;

[0162] Step S513: If σ i ′<σ i , then calculate the tightening coefficient a i =1-σ i ′ / σ i ,i=1,2,…,m;

[0163] If σ i ′>σ i , then it means If there are outliers, requantization cannot be performed to limit the error to the quantization step size, then S5 is skipped (i.e., no requantization is performed), and the sender directly sends the quantized data packet using the encoding-decoding quantization algorithm based on the innovation in step S3 and the quantization bit number in step S4;

[0164] If σ i ′=σ i , then it means If there is no room for tightening, requantization may not be performed. If the error is limited to the quantization step size, S5 is skipped (i.e., no requantization is performed). The sender directly sends the quantized data packet obtained by using the encoding-decoding quantization algorithm based on the innovation in step S3 and the quantization bit number in step S4.

[0165] Step S514: Establish a second optimization problem P2 for feedback correction to ensure that the reconstruction error can be limited to the range of the quantization step size and maximize the communication bandwidth saving rate, and solve the problem to obtain the number of quantization bits b for feedback correction. i .

[0166] Therefore, after quantizing the innovation information using the quantization bit number obtained by solving the first optimization problem, if the actual standard deviation is less than the theoretical standard deviation, the first optimization problem can be further combined with the second optimization problem to solve the feedback-corrected quantization bit number b. i By ensuring that the reconstruction error can be limited to the range of the quantization step size, the data validity is greatly improved.

[0167] The second optimization problem P2 is:

[0168]

[0169] in,

[0170] in

[0171]

[0172] Among them, Γ is the sufficient condition for the number of quantization bits; Γ x.y , Both represent subsets of Γ; i k ,j k ∈{1,2,…,2m} is a qualified subscript index, and j, x=1, 2, ..., m, y=1, 2, ..., 2m is also a subscript index; M z,i Represents the zth row and ith column of the innovation gain matrix.

[0173] That is to say, They refer to Γ x.y Take k in x and i in y k 、j k The value at time.

[0174] In order to limit the reconstruction error within the range of the quantization step when executing the feedback quantization algorithm, the number of quantization bits must satisfy the sufficient condition Γ for the number of quantization bits; when the number of quantization bits b i When ∈Γ, the reconstruction error can be limited to the range of the quantization step size. The sufficient condition Γ for the number of quantization bits is used in step S52 below.

[0175] The following specifically describes the process of solving the sufficient condition Γ for the number of quantization bits.

[0176] When the system is a two-dimensional system, at time k+0, according to the following formula, it is obvious that the reconstruction error is within the quantization step size:

[0177]

[0178] At time k+1, according to the following formula, it is obvious that the reconstruction error is within the quantization step size:

[0179]

[0180] At time k+2, we have:

[0181]

[0182] Expand to get:

[0183] e 1 (k+2)=M 1,1 e 1 (k+1)+M 1,2 e 2 (k+1)+n 1 (k+2) and

[0184] e 2 (k+2)=M 2,1 e 1 (k+1)+M 2,2 e 2 (k+1)+n 2 (k+2)

[0185] There are four extreme cases:

[0186] Extreme Case 1: e 1 (k+1)=+Δ 1 and e 2 (k+1)=+Δ 2 ;

[0187] Extreme Case 2: e 1 (k+1)=+Δ 1 and e 2 (k+1)=-Δ 2 ;

[0188] Extreme Case 3: e 1 (k+1)=-Δ 1 and e 2 (k+1)=-Δ 2 ;

[0189] Extreme Case 4: e1 (k+1)=-Δ 1 and e 2 (k+1)=+Δ 2 ;

[0190] Among them, extreme case 3 is equivalent to extreme case 1, and extreme case 4 is equivalent to extreme case 2, so only extreme case 1 and extreme case 2 need to be analyzed.

[0191] It is required to reconstruct the first dimension value e of the error 1 (k+2) and the second dimension value e 2 (k+2) are limited to [-Δ 1 ,+Δ 1 ] and [-Δ 2 ,+Δ 2 ], for the first dimension value e of the reconstruction error 1 The requirement of (k+2) acts on the quantization error n caused by the probabilistic quantizer 1 There are the following inequalities on (k+2):

[0192] n 1 (k+2)∈[(+1-M 1,1 )Δ 1 -M 1,2 Δ 2 ,(-1-M 1,1 )Δ 1 -M 1,2 Δ 2 ].

[0193] Among them, M 1,1 and M 1,2 The first row and first column scalar and the first row and second column scalar of the innovation gain matrix M; combined with n 1 The value range of (k+2) is n 1 (k+2)∈[Δy 1 (k+2)-3σ 1 ,Δy 1 (k+2)+3σ 1 ], you can get

[0194] Δy 1 (k+τ+1)-3σ 1 ≤(+1-M 1,1 )Δ 1 -M 1,2 Δ 2 ≤Δy 1 (k+τ+ 1)+3σ 1 and Δy 1 (k+τ+1)-3σ 1 ≤(-1-M 1,1)Δ 1 -M 1,2 Δ 2 ≤Δy 1 (k+τ+1)+3σ 1 .

[0195] Solve the above inequalities and get the solution sets:

[0196] and

[0197] Among them, Γ 1.1 , Γ 1.2 is parameter b 1 ,b 2 The solution set of b 1 ,b 2 Represents the quantization error of the first and second dimensions, a 1 It represents the difference between the actual standard deviation and the expected standard deviation, and the calculation formula is a i =1-σ i′ / σ i ,i=1,2,…,m.

[0198] So for any e 1 (k+1)∈[-Δ 1 ,+Δ 1 ] and the first dimension scalar Δy for any encoded measurement innovation Δy(k+τ+1) 1 (k+2)∈[-(3-a 1 )σ 1 ,+(3-a 1 )σ 1 ], when (b 1 ,b 2 )∈(Γ 1.1 ∪Γ 1.2 ), there must be -Δ 1 ≤e 1 (k+2)≤Δ 1 , e 1 (k+2) is the first dimension value of the reconstruction error.

[0199] The same analysis process is used for e 2 (k+2), the solution set can be derived:

[0200] and

[0201] So for any e 2 (k+1)∈[-Δ 2 ,+Δ 2] and for any Δy 2 (k+2)∈[-(3-a 2 )σ 2 ,+(3-a 2 )σ 2 ], when (b 1 ,b 2 )∈(Γ 1.3 ∪Γ 1.4 ), there must be -Δ 2 ≤e 2 (k+2)≤Δ 2 .

[0202] So in extreme case 1, when (Γ 1.1 ∪Γ 1.2 )∩(Γ 1.3 ∪Γ 1.4 ), for any e i (k+1)∈[-Δ i ,+Δ i ] and for any Δy i (k+2)∈[-(3-a i )σ i ,+(3-a i )σ i ], there must be -Δ i ≤e i (k+2)≤Δ i ,i=1,2.

[0203] Similarly, in extreme case 2, when (Γ 2.1 ∪Γ 2.2 )∩(Γ 2.3 ∪Γ 2.4 ), for any e i (k+1)∈[-Δ i ,+Δ i ] and for any Δy i (k+2)∈[-(3-a i )σ i ,+(3-a i )σ i ], there must be -Δ i ≤e i (k+2)≤Δ i ,i=1,2, where

[0204]

[0205]

[0206]

[0207]

[0208] definition When (b 1 ,b 2 )∈Γ, the secondary quantization can definitely limit the error within the quantization step size.

[0209] Extending Γ to the m-dimensional system, the sufficient condition Γ for the number of quantization bits obtained by solving the second optimization problem P2 is:

[0210]

[0211] in

[0212]

[0213] Where, Γ is a sufficient condition for the number of quantization bits; in order to limit the reconstruction error within the range of the quantization step when executing the feedback quantization algorithm, the number of quantization bits must meet the sufficient condition for the number of quantization bits; Γ x.y , Both represent subsets of Γ; i k ,j k ∈{1,2,…,2m} is a qualified subscript index, and j, x=1, 2, ..., m, y=1, 2, ..., 2m is also a subscript index; M z,i Represents the zth row and ith column of the innovation gain matrix.

[0214] Step S52: The sender uses the quantization bit number b obtained in step S51 to i The encoding-decoding quantization algorithm based on the new information of step S3 is used to encode the data packet. Quantization, to obtain encoded and quantized data packets

[0215] Among them, the encoding-decoding quantization algorithm based on the innovation can be based on the quantization bit number b i Solve the quantization step and quantization level of each dimension, and then convert the encoded data packet Quantification.

[0216] Step S53: The sender reconstructs the encoded and quantized data packet. Get the reconstructed data packet The reconstructed data packet Subtract the original measurement vector data packet y(k,k+T-1) to obtain the data packet e(k,k+T-1) of the receiver reconstruction error e(k+τ+1) simulated by the sender;

[0217] It should be noted that since the sender already knows the receiver's new information-based encoding-decoding quantization algorithm, the receiver reconstruction error e(k+τ+1) simulated by the sender here is completely consistent with the reconstruction error e(k+τ+1) obtained by the receiver in terms of both numerical value and calculation method.

[0218] Step S54: traverse the receiver reconstruction error e(k+τ+1) simulated by the sender from τ=0 to τ=T-2; during the traversal, if the value of any dimension in the receiver reconstruction error e(k+τ+1) simulated by the sender |e i (k+τ+1)|>quantization step size Δ i , the receiver reconstruction error e(k+τ+1) simulated by the sender is used as the feedback input of the probability quantizer for secondary quantization. The secondary quantization includes traversing the measurement of the new information Δy(k+τ+1)=[Δy 1 (k+τ+1),Δy 2 (k+τ+1),…,Δy m (k+τ+1)] corresponding to the possible value combination of the quantization level [q 1 ,q 2 ,…,q m ] and obtain the corresponding sender-simulated receiver reconstruction error e(k+τ+1), until the values ​​of all dimensions in the sender-simulated receiver reconstruction error |e are found. i (k+τ+1)| all satisfy|e i (k+τ+1)|≤Δ i The value combination of the quantization level is used as the new quantized measurement information and replace the encoded and quantized data packet of step S52 Otherwise, directly use the encoded and quantized data packet of step S52 The original quantified measurement information.

[0219] Where, for each measurement innovation Δy(k+τ+1), the scalar Δy of the i-th dimension is i (k+τ+1), the possible value q of the corresponding quantization level i satisfy are the quantization bits b in step S514 i The quantization level obtained by the solution.

[0220] Step S55: Send the final encoded and quantized data packet

[0221] Step S6: The transmitter encodes and transmits the data and the receiver receives and decodes the data according to the encoding-decoding quantization algorithm and the feedback quantization algorithm based on the new information, so that the receiver obtains the reconstructed measurement vector.

[0222] The step S6 further includes: performing state estimation of the target system according to the reconstructed measurement vector obtained by the receiver.

[0223] In this embodiment, the target system is a target vehicle in the Internet of Vehicles; performing state estimation of the target system refers to tracking and state estimation of the target vehicle, and the state estimation value includes the position and speed of the target vehicle.

[0224] Therefore, the present invention encodes and quantizes the measurement vector at the transmitter and decodes it at the receiver to obtain a lossy measurement vector, which can be used for tracking and state estimation of target vehicles in the Internet of Vehicles.

[0225] In the connected vehicle (IoV) environment, communication frequencies are often lower than measurement frequencies, requiring efficient data transmission within limited communication resources. Therefore, IoV is naturally suited to packet transmission for information exchange within bandwidth-constrained conditions. Furthermore, due to limited communication bandwidth, quantization algorithms must be used to compress measurement data to reduce data transmission overhead and improve communication efficiency. In this context, connected vehicles must quantize and encode measurement data and send the quantized data packets to a server. Upon receiving the data, the server performs a series of downstream tasks, including state estimation, decision analysis, and storage. State estimates in IoV may include the position and velocity of target vehicles. The velocity of a vehicle in a certain direction may serve as a basis for path planning for other vehicles. However, quantization encoding not only affects the accuracy of state estimation but can also have ripple effects on the decision-making process. For example, when vehicles transmit low-precision quantized position and velocity information, the server may introduce additional errors in state estimation, impacting critical decisions such as path planning and obstacle avoidance.

[0226] In the global coordinate system, the vehicle state may gradually diverge (for example, the vehicle state data includes position and speed. In the global coordinate system, since the vehicle may continue to move forward, its position will continue to increase, so the state may gradually diverge), which will cause the measurement data to diverge. This situation will cause the existing quantization coding method to face the problem of quantizer saturation, which will affect the accuracy of data beyond the quantization range. However, the algorithm proposed in the present invention can effectively circumvent this problem. Specifically, the measurement innovation value constructed by this method only depends on the system state transfer matrix A, the measurement matrix H, the variance Q of the system Gaussian noise w(k), and the variance R of the measurement Gaussian noise v(k). As long as the defined abnormal innovation value does not appear (the probability of its occurrence is extremely low, only 0.26%), the quantizer saturation phenomenon will not occur. In addition, further analysis combined with the encoding process shows that even if an abnormal value occurs, since each data packet contains a completely accurate calibration value. Therefore, the impact of the abnormal innovation value is limited to the reconstructed data in the current data packet, so the impact of the abnormal value on the overall data reconstruction is limited.

[0227] With the above quantization coding technology, the receiver only needs to know the encoding and decoding rules, as well as the system state transition matrix A, the measurement matrix H, the variance Q of the system Gaussian noise w(k), and the variance R of the measurement Gaussian noise v(k) in advance to decode the data. No additional redundant data required for binary and decimal conversion is required each time a data packet is decoded, thus achieving self-explanatory coding.

[0228] The innovation-based quantization coding method proposed in this paper can achieve excellent communication resource conservation while ensuring data validity when processing large-scale observation signals. Furthermore, the feedback quantization method provided by this invention can significantly reduce data errors, improve data availability, and effectively reduce communication resource consumption.

[0229] Simulation results:

[0230] Figure 2 This is an exemplary schematic diagram of quantization encoding and decoding of the present invention. Figure 3A-Figure 3E An exemplary diagram of quantization error, reconstruction error, and MSE of an estimation task is provided for some embodiments of the present invention. Figure 3A Contains experimental results on quantization error using the innovation-based encoding quantization method (IEQ). Figure 3B Contains experimental results on reconstruction error of the innovation-based quantization coding method; Figure 3C and Figure 3D It is the experimental result of the mean square error (MSE) when using the decoded data to perform the state estimation task; Figure 3EContains experimental results on the quantization error of the reconstruction error-based feedback quantization coding method (denoted as REFQ).

[0231] Figure 3A In the figure, the curves marked with dim1 and dim2 respectively represent the experimental results of the quantization error of the two-dimensional measurement values, and the orange and blue dotted lines correspond to the experimental results of their limited range. Figure 3B In the figure, the curves marked with dim1 and dim2 represent the experimental results of the reconstruction error of the two-dimensional measurement values, and the red dotted line represents the experimental results of the range that should be limited. Figure 3C In the figure, the curve labeled "Baseline" represents the MSE of the first dimension of the system state when using lossless data for estimation. The curves labeled "IEQ(∈=0.3)" and "IEQ(∈=1.0)" correspond to the experimental results of the innovation-based quantization coding algorithm of the present invention when the reconstruction error is set to 0.3 and 1.0, respectively. Figure 3D In the figure, the Baseline curve represents the MSE result of estimating the second dimension of the system state using lossless data. The IEQ(∈=0.3) and IEQ(∈=1.0) curves correspond to the experimental results of the quantization coding algorithm based on innovation when the reconstruction error is set to 0.3 and 1.0, respectively. Figure 3E In the figure, the curves marked with dim1 and dim2 represent the experimental results of the reconstruction error of the two-dimensional measurement values ​​under the feedback quantization coding method based on innovation of the present invention, and the orange and blue dotted lines correspond to the experimental results of the range that should be limited.

[0232] like Figure 2 As shown in FIG3 , the present invention uses a vehicle network driving state tracking model to verify the effectiveness of the proposed algorithm. The state space equation of this tracking model is:

[0233]

[0234]

[0235] in Indicates the state of the tracking target. and is the position and velocity of the target in the x-axis direction; dt = 0.1 sec is the time interval, and the covariance matrix of the noise is:

[0236]

[0237] Among them, x(0), P i (0) and P i,j The initial value of (0) is:

[0238]

[0239] The performance of these algorithms is compared using quantization error, reconstruction error, and mean square error (MSE). The quantization error is defined as:

[0240]

[0241] The reconstruction error is defined as:

[0242]

[0243] The root mean square error is defined as:

[0244]

[0245] where Δy i (k+τ+1) is the original innovation value, is the quantized new value, y(k+τ+1) is the original measurement value, is the reconstructed measurement value, is the total number of Monte Carlo experiments, in some instances r represents the rth Monte Carlo experiment, x r (k) is the kth moment in the rth Monte Carlo experiment, is the estimated state value of the i-th sensor at time k in the r-th Monte Carlo experiment.

[0246] The state estimation performance of a quantization-based coding algorithm was compared with that of an unprocessed data case. Simulation results show that, while the estimation performance of the innovation-based quantization coding method provided by the present invention is slightly lower than that of the unprocessed data case, this method can save up to 74.01% of communication bandwidth. Furthermore, the innovation-based feedback quantization coding method outperforms the innovation-based quantization coding method in terms of reconstruction error.

[0247] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of the present invention. Various modifications and variations are possible. In other words, any simple, equivalent changes and modifications made in accordance with the claims and description of the present invention are within the scope of protection of the patent claims. Anything not fully described in this invention constitutes conventional technology.

Claims

1. An efficient communication quantization coding method based on innovation, characterized in that: include: Step S1, establishing a state space model of the target system and sensor nodes; Step S2: constructing a data packet of measurement vectors and measurement innovations independent of the system state based on the state space model to reduce the accuracy loss caused by large-scale observation output; Step S3: Based on the differential coding strategy, establish an innovation-based encoding-decoding quantization algorithm so that the encoded data packet consists of the calibration value and the measurement innovation value; Step S4, establishing a first optimization problem with data validity as a constraint on the reconstruction error and communication bandwidth saving rate η as an optimization target; Obtaining the optimal number of quantization bits and the number of data items in the data packet required by the innovation-based encoding-decoding quantization algorithm based on the first optimization problem; Step S6: encoding and sending by the sender and receiving and decoding by the receiver are performed according to the coding-decoding quantization algorithm based on the innovation, so that the receiver obtains the reconstructed measurement vector.

2. The innovation-based efficient communication quantization coding method according to claim 1, characterized in that: In step S1, the target system is a real-world physical system whose physical parameters can be measured by sensors, and the state of the target system refers to the physical parameters; The expressions of the state space model of the system state and sensor nodes are: x(k+1)=Ax(k)+w(k), y(k)=Hx(k)+v(k), Where x(k) is the true state of the system at time k, m is the dimension of the true state, x(k+1) is the true state of the system at time k+1, A is the system state transfer matrix, w(k) is the Gaussian system noise at time k, with mean zero and variance Q; y(k) is the measurement vector at time k, n is the dimension of the measurement vector, H is the measurement matrix of the sensor, v(k) is the Gaussian measurement noise at time k, with mean zero and variance R; k is the first moment.

3. The innovation-based efficient communication quantization coding method according to claim 2, wherein: The step S2 specifically includes: Step S21, define the innovation gain matrix M = HAH -1 , the encoded measurement innovation of the measurement vector y(+τ+1) is defined as Δy(k+τ+1)=y(k+τ+1)-My(k+τ), where y(k+τ+1) and y(k+τ) are the measurement vectors at time k+τ+1 and k+τ, respectively, and τ represents an arbitrary time. The measurement vector between two communications between the sensor and the remote server is recorded as the measurement vector data packet y(k,k+T-1); Step S22, obtaining an expansion of the measurement innovation according to the system state and the state space model of the sensor node; Step S23, set M=HAH -1 Substitute the expansion of the measurement innovation and calculate the variance matrix of the measurement innovation. According to the variance matrix of the measurement innovation Determine a normal value interval for measuring innovation. The normal value interval for measuring innovation is used to ensure that the frequency of abnormal values ​​of the measuring innovation meets the requirements.

4. The innovation-based efficient communication quantization coding method according to claim 1, wherein: In the innovation-based encoding-decoding quantization algorithm, the original measurement vector data packet is encoded and then sent during each communication. The encoding method includes: Step S31: The measurement vector at the first moment in the data packet of the original measurement vector is used as a calibration value; and the measurement vectors at other moments other than the first moment are encoded as measurement innovations; Step S32: For the measurement innovation Δy(k+τ+1) at the remaining moments other than the first moment, τ=0, 1, ..., T-1, T is the number of data items in the data packet, use the probabilistic quantizer to quantize the scalar of the measurement innovation Δy(k+τ+1) in each dimension, obtain the quantized innovation value of each dimension, and combine them to obtain the quantized measurement innovation. Then we get the encoded and quantized data packets Step S33: convert the measurement innovation obtained by the probability quantizer into a scalar Δy in each dimension. i (k+τ+1) and the quantized innovation value of this dimension The deviation between them is defined as the quantization error n i (k+τ+1), and determine the covariance of the quantization error.

5. The innovation-based efficient communication quantization coding method according to claim 4, characterized in that: The probability quantizer used meets the following requirements: the scalar Δy in the i-th dimension according to the measurement innovation Δy(k+τ+1) i (k+τ+1) (i is the dimension number of the real state), the quantized innovation value of the i-th dimension for: Where i = 1, 2, ..., m, m is the dimension of the real state, i is the dimension ordinal number of the real state; Δy i (k+τ+1) is the scalar of the measurement innovation Δy(k+τ+1) in the i-th dimension; q j+1 (k) and q j (k) is the quantization level, j is the ordinal number of the quantization level; k is the first moment, Δ is the quantization step size; The quantization step size Δ is: in, are the lower and upper limits of the signal range to be quantized, and b is the number of quantization bits; And the relationship between the quantization level, quantization step size and quantization bit number is as follows: Δ=q j+1 -q j ,j=0,1,…,2 b -2, Among them, q j+1 ,q j is the quantization level, 6. The innovation-based efficient communication quantization coding method according to claim 4, characterized in that: In the innovation-based encoding-decoding quantization algorithm, the receiver decodes the encoded and quantized data packet to obtain the reconstructed data packet. Decoding methods include: Step S31', according to the formula M=HAH -1 Calculate the innovation gain matrix M and decode it to get the reconstructed measurement vector Merge to obtain reconstructed data packets Reconstructed measurement vector for: in, is the reconstructed measurement vector at time k, which is the same as the measurement vector before encoding and quantization; is the reconstructed measurement vector at time k+τ+1, M is the innovation gain matrix, M=HAH -1 , H is the sensor measurement matrix, A is the system state transfer matrix, is the quantized measurement information at time k+τ+1; Step S32', reconstruct the measurement vector at time k+τ+1 The deviation from the original measurement information y(k+τ+1) at time k+τ+1 is defined as the reconstruction error e(k+τ+1); the covariance of the reconstruction error e(k+τ+1) is calculated and the normal numerical range of the reconstruction error is determined accordingly.

7. The innovation-based efficient communication quantization coding method according to claim 6, characterized in that: In step S4, the first optimization problem P1 is established as: Among them, b i is the number of quantized bits, i is the dimension number of the real state, m is the dimension of the real state, T is the number of data items in the data packet, is the upper bound of the normal numerical interval of the reconstruction error, ∈ i For data validity, Z represents a set of integers.

8. The innovation-based efficient communication quantization coding method according to claim 6, wherein: For the receiver, after decoding, the reconstructed measurement vector Before, it also includes: the total number of bits B of the combination of the number of data items T in the data packet and the calibration value of the received encoded and quantized data packet and the scalar of the i-th dimension of the measurement innovation i , calculate the number of quantized bits b of the scalar of the i-th dimension of each measurement innovation i , to divide the starting and ending positions of different measurement innovations.

9. The innovation-based efficient communication quantization coding method according to claim 6, wherein: The step S5 is also included, wherein a feedback quantization algorithm is further introduced into the encoding and decoding process; If the actual standard deviation of the measured innovation is less than the theoretical standard deviation, the feedback quantization algorithm calculates the number of quantization bits for feedback correction, which ensures that the reconstruction error can be limited to the range of the quantization step size and maximizes the communication bandwidth saving rate; Then, the innovation-based encoding-decoding quantization algorithm is executed to obtain the receiver reconstruction error simulated by the sender. This is used as the feedback input of the probabilistic quantizer to re-quantize the measured innovation to ensure that the receiver reconstruction error simulated by the sender is within the range of the quantization step size. The data packet is then sent. Otherwise, the sender directly sends the data packet quantized by the encoding-decoding quantization algorithm based on the innovation in step S3 and the quantization bit number in step S4.

10. The innovation-based efficient communication quantization coding method according to claim 9, characterized in that: The step S5 specifically includes: Step S51: If the actual standard deviation of the measured innovation is less than the theoretical standard deviation, the second optimization problem P2 is solved to determine the number of quantization bits for feedback correction that ensures the reconstruction error is limited to the quantization step size and maximizes the communication bandwidth saving rate. Otherwise, the quantized data packet obtained by using the innovation-based encoding-decoding quantization algorithm in step S3 and the quantization bit number in step S4 is directly sent, and then step S6 is executed. The second optimization problem P2 is: in, Among them, Γ is the sufficient condition for the number of quantization bits; Γ x.y , Both represent subsets of Γ; i k ,j k ∈{1,2,…,2m} is a qualified subscript index, and j, x=1, 2, ..., m, y=1, 2, ..., 2m is also a subscript index; M z,i Represents the zth row and ith column of the innovation gain matrix.

Images